question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
Training with synthetic data makes evaluation decontamination harder.The use of synthetic data in model training poses significant challenges to fair evaluation. Evaluation benchmarks are often created by referring to public text sources, such as coursework websites or forums. Consequently, it is arguable that all publicly available benchmark test cases might occasionally be included in the pre-training data of LLMs (Gao et al., 2021; Hoffmann et al., 2022). The use of synthetic data exacerbates this issue rather than mitigating it. Although the community has proposed several techniques to detect such evaluation contamination, such as _min-k9 prob_(Shi et al., 2023), which checks the probabilities of \(k\) long-tail tokens, these token-level decontamination methods are inadequate whenthe model is trained with synthetic data. Synthetic data might include rephrased versions of the benchmark data (Mattern et al., 2023; Oren et al., 2023), rendering token-level decontamination ineffective. In addition to developing more advanced evaluation contamination detection techniques, we recommend that model developers invest in creating and maintaining in-house and protected evaluation benchmarks. These proprietary benchmarks should be carefully safeguarded to prevent leakage and ensure the integrity of the evaluation process.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ๊ฐ€ ๋” ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ•™์Šต์— synthetic data๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. Evaluation benchmarks๋Š” ์ฝ”์Šค ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋งŒ๋“ค์–ด์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์‚ฌ๋ก€๊ฐ€ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2021; Hoffmann et al., 2022). Synthetic data๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ์™„ํ™”๋˜๊ธฐ๋ณด๋‹ค๋Š” ์˜คํžˆ๋ ค ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์˜ค์—ผ์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•ด \(k\) ๋กฑํ…Œ์ผ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_(Shi et al., 2023) ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ์ด synthetic data๋กœ ํ•™์Šต๋œ ๊ฒฝ์šฐ์—๋Š” ์ ์ ˆํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Synthetic data์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ๊ณ ์•ˆ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ(Mattern et al., 2023; Oren et al., 2023), ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๊ฐ€ ํšจ๊ณผ์ ์ด์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ๊ธ‰ ํ‰๊ฐ€ ์˜ค์—ผ ๊ฐ์ง€ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ evaluation benchmarks๋ฅผ ๋งŒ๋“ค๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ๊ฐ€ ๋” ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ•™์Šต์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ฝ”์Šค ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋งŒ๋“ค์–ด์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์‚ฌ๋ก€๊ฐ€ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2021; Hoffmann et al., 2022). ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ์™„ํ™”๋˜๊ธฐ๋ณด๋‹ค๋Š” ์˜คํžˆ๋ ค ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์˜ค์—ผ์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•ด \(k\) ๋กฑํ…Œ์ผ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_(Shi et al., 2023) ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๊ฒฝ์šฐ์—๋Š” ์ ์ ˆํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ๊ณ ์•ˆ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ(Mattern et al., 2023; Oren et al., 2023), ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๊ฐ€ ํšจ๊ณผ์ ์ด์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ๊ธ‰ ํ‰๊ฐ€ ์˜ค์—ผ ๊ฐ์ง€ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
The issue of hallucination in LLMs has been explored in multiple contexts and models (Ji et al., 2023; Kaddour et al., 2023). As a response, RAG systems have been shown to reduce hallucination (Shuster et al., 2021; Kang et al., 2023). Previous works have explored automated RAG evaluation frameworks in various settings (Es et al., 2023; Hoshi et al., 2023; Saad-Falcon et al., 2023; Zhang et al., 2024). For example, some studies use LLMs to evaluate the faithfulness, answer relevance, and context relevance of RAG systems by using GPT-3.5 as an evaluator (Es et al., 2023; Saad-Falcon et al., 2023). In another study, the authors propose metrics such as noise robustness, negative rejection, information integration, and counterfactual robustness (Chen et al., 2024). Multiple studies have shown that RAG can mislead LLMs in the presence of complex or misleading search results and that such models can still make mistakes even when given the correct response (Foulds et al., 2024; Shuster et al., 2021). In relation to understanding model priors, other works have used log probabilities to assess the LLM's confidence in responses (Mitchell et al., 2023; Zhao et al., 2024). However, so far there has not been a systematic exploration of a model's confidence (via logprobs) and the model's preference for RAG-provided information.
LLM์˜ hallucination ๋ฌธ์ œ๋Š” ์—ฌ๋Ÿฌ ๋งฅ๋ฝ๊ณผ ๋ชจ๋ธ(Ji et al., 2023; Kaddour et al., 2023)์—์„œ ํƒ๊ตฌ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ๋Œ€์‘์œผ๋กœ RAG ์‹œ์Šคํ…œ์€ hallucination์„ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Shuster et al., 2021; Kang et al., 2023). ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ ์ž๋™ํ™”๋œ RAG ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค(Es et al., 2023; Hoshi et al., 2023; Saad-Falcon et al., 2023; Zhang et al., 2024). ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” GPT-3.5๋ฅผ ํ‰๊ฐ€์ž๋กœ ์‚ฌ์šฉํ•˜์—ฌ RAG ์‹œ์Šคํ…œ์˜ ์ถฉ์‹ค๋„, ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ ๋ฐ ์ปจํ…์ŠคํŠธ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค(Es et al., 2023; Saad-Falcon et al., 2023). ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋…ธ์ด์ฆˆ robustness, ๋ถ€์ • ๊ฑฐ๋ถ€, ์ •๋ณด ํ†ตํ•ฉ ๋ฐ ๋ฐ˜์‚ฌ์‹ค์  robustness์™€ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค(Chen et al., 2024). ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋ณต์žกํ•˜๊ฑฐ๋‚˜ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ RAG๊ฐ€ LLM์„ ์˜ค๋„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ •๋‹ต์ด ์ฃผ์–ด์ง„ ๊ฒฝ์šฐ์—๋„ ์—ฌ์ „ํžˆ ์‹ค์ˆ˜๋ฅผ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Foulds et al., 2024; Shuster et al., 2021). ๋ชจ๋ธ์˜ ์„ ํ—˜์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จํ•˜์—ฌ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” log probabilities๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์‹ ๋ขฐ๋„๋ฅผ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค(Mitchell et al., 2023; Zhao et al., 2024). ๊ทธ๋Ÿฌ๋‚˜ ์ง€๊ธˆ๊นŒ์ง€ ๋ชจ๋ธ์˜ ์‹ ๋ขฐ๋„(logprobs๋ฅผ ํ†ตํ•ด)์™€ RAG๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ํƒ๊ตฌ๋Š” ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
LLM์˜ ํ™˜๊ฐ ๋ฌธ์ œ๋Š” ์—ฌ๋Ÿฌ ๋งฅ๋ฝ๊ณผ ๋ชจ๋ธ(Ji et al., 2023; Kaddour et al., 2023)์—์„œ ํƒ๊ตฌ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ๋Œ€์‘์œผ๋กœ RAG ์‹œ์Šคํ…œ์€ ํ™˜๊ฐ์„ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Shuster et al., 2021; Kang et al., 2023). ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ ์ž๋™ํ™”๋œ RAG ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค(Es et al., 2023; Hoshi et al., 2023; Saad-Falcon et al., 2023; Zhang et al., 2024). ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” GPT-3.5๋ฅผ ํ‰๊ฐ€์ž๋กœ ์‚ฌ์šฉํ•˜์—ฌ RAG ์‹œ์Šคํ…œ์˜ ์ถฉ์‹ค๋„, ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ ๋ฐ ์ปจํ…์ŠคํŠธ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค(Es et al., 2023; Saad-Falcon et al., 2023). ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋…ธ์ด์ฆˆ ๊ฒฌ๊ณ ์„ฑ, ๋ถ€์ • ๊ฑฐ๋ถ€, ์ •๋ณด ํ†ตํ•ฉ ๋ฐ ๋ฐ˜์‚ฌ์‹ค์  ๊ฒฌ๊ณ ์„ฑ๊ณผ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค(Chen et al., 2024). ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋ณต์žกํ•˜๊ฑฐ๋‚˜ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ RAG๊ฐ€ LLM์„ ์˜ค๋„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ •๋‹ต์ด ์ฃผ์–ด์ง„ ๊ฒฝ์šฐ์—๋„ ์—ฌ์ „ํžˆ ์‹ค์ˆ˜๋ฅผ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Foulds et al., 2024; Shuster et al., 2021). ๋ชจ๋ธ์˜ ์„ ํ—˜์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จํ•˜์—ฌ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋กœ๊ทธ ํ™•๋ฅ ์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์‹ ๋ขฐ๋„๋ฅผ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค(Mitchell et al., 2023; Zhao et al., 2024). ๊ทธ๋Ÿฌ๋‚˜ ์ง€๊ธˆ๊นŒ์ง€ ๋ชจ๋ธ์˜ ์‹ ๋ขฐ๋„(๋กœ๊ทธ ํ™•๋ฅ ์„ ํ†ตํ•ด)์™€ RAG๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ํƒ๊ตฌ๋Š” ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
We study whether Large Language Models (LLMs) latently perform multi-hop reasoning with complex prompts such as "The mother of the singer of 'Superstition' is". We look for evidence of a latent reasoning pathway where an LLM (1) latently identifies "the singer of 'Superstition"' as Stevie Wonder, the _bridge entity_, and (2) uses its knowledge of Stevie Wonder's mother to complete the prompt. We analyze these two hops individually and consider their co-occurrence as indicative of latent multi-hop reasoning. For the first hop, we test if changing the prompt to indirectly mention the bridge entity instead of any other entity increases the LLM's internal recall of the bridge entity. For the second hop, we test if increasing this recall causes the LLM to better utilize what it knows about the bridge entity. We find strong evidence of latent multi-hop reasoning for the prompts of certain relation types, with the reasoning pathway used in more than 80% of the prompts. However, the utilization is highly contextual, varying across different types of prompts. Also, on average, the evidence for the second hop and the full multi-hop traversal is rather moderate and only substantial for the first hop. Moreover, we find a clear scaling trend with increasing model size for the first hop of reasoning but not for the second hop.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด "The mother of the singer of 'Superstition' is"์™€ ๊ฐ™์€ ๋ณต์žกํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ์ž ์žฌ์ ์œผ๋กœ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. LLM์ด (1) ์ž ์žฌ์ ์œผ๋กœ '_bridge entity_'์ธ ์Šคํ‹ฐ๋น„ ์›๋”์˜ '์ˆ˜ํผ์Šคํ‹ฐ์…˜'์˜ ๊ฐ€์ˆ˜๋กœ ์ž ์žฌ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ , (2) ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ์— ๋Œ€ํ•œ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๋Š” ์ž ์žฌ์  ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ฆ๊ฑฐ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ํ™‰์„ ๊ฐœ๋ณ„์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ณ  ์ด๋“ค์˜ ๊ณต๋™ ๋ฐœ์ƒ์„ ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์˜ ์ง€ํ‘œ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ํ™‰์˜ ๊ฒฝ์šฐ, ๋‹ค๋ฅธ ์—”ํ‹ฐํ‹ฐ ๋Œ€์‹  bridge entity๋ฅผ ๊ฐ„์ ‘์ ์œผ๋กœ ์–ธ๊ธ‰ํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณ€๊ฒฝํ•˜๋ฉด bridge entity์— ๋Œ€ํ•œ LLM์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ™‰์˜ ๊ฒฝ์šฐ, ์ด ๋ฆฌ์ฝœ์„ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด LLM์ด bridge entity์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ๋” ์ž˜ ํ™œ์šฉํ•˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ํŠน์ • ๊ด€๊ณ„ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์˜ ๊ฐ•๋ ฅํ•œ ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ด ์ถ”๋ก  ๊ฒฝ๋กœ๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ 80% ์ด์ƒ์—์„œ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ™œ์šฉ๋„๋Š” ํ”„๋กฌํ”„ํŠธ ์œ ํ˜•์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š” ๋“ฑ ์ƒํ™ฉ์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ‰๊ท ์ ์œผ๋กœ ๋‘ ๋ฒˆ์งธ ํ™‰๊ณผ ์ „์ฒด ๋‹ค์ค‘ ํ™‰ ํšก๋‹จ์— ๋Œ€ํ•œ ์ฆ๊ฑฐ๋Š” ๋‹ค์†Œ ๋ณดํ†ต์ด๋ฉฐ ์ฒซ ๋ฒˆ์งธ ํ™‰์—๋งŒ ์ƒ๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ ํ™‰์—์„œ๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ํ™•์‹คํ•œ ํ™•์žฅ ์ถ”์„ธ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ ๋‘ ๋ฒˆ์งธ ํ™‰์—์„œ๋Š” ๊ทธ๋ ‡์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด "The mother of the singer of 'Superstition' is"์™€ ๊ฐ™์€ ๋ณต์žกํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ์ž ์žฌ์ ์œผ๋กœ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. LLM์ด (1) ์ž ์žฌ์ ์œผ๋กœ '_๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ'์ธ ์Šคํ‹ฐ๋น„ ์›๋”์˜ '์ˆ˜ํผ์Šคํ‹ฐ์…˜'์˜ ๊ฐ€์ˆ˜๋กœ ์ž ์žฌ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ , (2) ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ์— ๋Œ€ํ•œ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๋Š” ์ž ์žฌ์  ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ฆ๊ฑฐ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ํ™‰์„ ๊ฐœ๋ณ„์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ณ  ์ด๋“ค์˜ ๊ณต๋™ ๋ฐœ์ƒ์„ ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์˜ ์ง€ํ‘œ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ํ™‰์˜ ๊ฒฝ์šฐ, ๋‹ค๋ฅธ ์—”ํ‹ฐํ‹ฐ ๋Œ€์‹  ๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ฐ„์ ‘์ ์œผ๋กœ ์–ธ๊ธ‰ํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณ€๊ฒฝํ•˜๋ฉด ๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ LLM์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ™‰์˜ ๊ฒฝ์šฐ, ์ด ๋ฆฌ์ฝœ์„ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด LLM์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ๋” ์ž˜ ํ™œ์šฉํ•˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ํŠน์ • ๊ด€๊ณ„ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์˜ ๊ฐ•๋ ฅํ•œ ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ด ์ถ”๋ก  ๊ฒฝ๋กœ๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ 80% ์ด์ƒ์—์„œ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ™œ์šฉ๋„๋Š” ํ”„๋กฌํ”„ํŠธ ์œ ํ˜•์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š” ๋“ฑ ์ƒํ™ฉ์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ‰๊ท ์ ์œผ๋กœ ๋‘ ๋ฒˆ์งธ ํ™‰๊ณผ ์ „์ฒด ๋‹ค์ค‘ ํ™‰ ํšก๋‹จ์— ๋Œ€ํ•œ ์ฆ๊ฑฐ๋Š” ๋‹ค์†Œ ๋ณดํ†ต์ด๋ฉฐ ์ฒซ ๋ฒˆ์งธ ํ™‰์—๋งŒ ์ƒ๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ ํ™‰์—์„œ๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ํ™•์‹คํ•œ ํ™•์žฅ ์ถ”์„ธ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ ๋‘ ๋ฒˆ์งธ ํ™‰์—์„œ๋Š” ๊ทธ๋ ‡์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Formatting design._ As an important factor, the design of natural language format also highly impacts the generalization performance of LLMs [88]. Typically, we can add task descriptions and optional demonstrations to the input-output pairs of existing datasets, where the task description is the most key part for LLMs to understand the task [88]. Further, it can lead to substantial improvements by using an appropriate number of exemplars as demonstrations [69], which also alleviates the model sensitivity to instruction engineering [67, 69]. However, incorporating other components (_e.g.,_ things to avoid, reasons, and suggestions) into instructions may have a negligible or even adverse effect on the performance of LLMs [88, 166]. Recently, to elicit the step-by-step reasoning ability of LLMs, some work [69] proposes to include chain-of-thought (CoT) examples for some reasoning datasets, such as arithmetic reasoning. It has been shown that fine-tuning LLMs with both CoT and non-CoT examples can lead to a good performance across various reasoning tasks, including those that require multi-hop reasoning ability (_e.g.,_ commonsense question answering and arithmetic reasoning) as well as those without the need for such a reasoning way (_e.g.,_ sentiment analysis and extractive question answering) [69, 95].
\(\bullet\)_ํฌ๋งท ๋””์ž์ธ._ ์ž์—ฐ์–ด ํ˜•์‹์˜ ๋””์ž์ธ๋„ LLM์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์ค‘์š”ํ•œ ์š”์†Œ์ž…๋‹ˆ๋‹ค[88]. ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์— ์ž‘์—… ์„ค๋ช…๊ณผ ์„ ํƒ์  ๋ฐ๋ชจ๋ฅผ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ž‘์—… ์„ค๋ช…์€ LLM์ด ์ž‘์—…์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๋ถ€๋ถ„์ž…๋‹ˆ๋‹ค[88]. ๋˜ํ•œ ์ ์ ˆํ•œ ์ˆ˜์˜ ์˜ˆ์‹œ๋ฅผ ๋ฐ๋ชจ๋กœ ์‚ฌ์šฉํ•˜๋ฉด ์ƒ๋‹นํ•œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ[69], ์ด๋Š” ๋ชจ๋ธ์ด Instruction ์—”์ง€๋‹ˆ์–ด๋ง์— ๋Œ€ํ•œ ๋ฏผ๊ฐ๋„๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ์—๋„ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[67, 69]. ๊ทธ๋Ÿฌ๋‚˜ Instruction์— ๋‹ค๋ฅธ ๊ตฌ์„ฑ ์š”์†Œ(_์˜ˆ:_ ํ”ผํ•ด์•ผ ํ•  ์‚ฌํ•ญ, ์ด์œ  ๋ฐ ์ œ์•ˆ)๋ฅผ ํฌํ•จํ•˜๋ฉด LLM์˜ ์„ฑ๋Šฅ์— ๋ฏธ๋ฏธํ•˜๊ฑฐ๋‚˜ ์‹ฌ์ง€์–ด ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[88, 166]. ์ตœ๊ทผ์—๋Š” LLM์˜ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ์‚ฐ์ˆ  ์ถ”๋ก ๊ณผ ๊ฐ™์€ ์ผ๋ถ€ ์ถ”๋ก  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Chain-of-Thought (CoT) ์˜ˆ์ œ๋ฅผ ํฌํ•จํ•˜๋„๋ก ์ œ์•ˆํ•˜๋Š” ์ž‘์—…[69]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CoT์™€ ๋น„ CoT ์˜ˆ์ œ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ Fine-tuningํ•˜๋ฉด ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ์ž‘์—…(์˜ˆ: ์ƒ์‹ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก )๊ณผ ๊ทธ๋Ÿฌํ•œ ์ถ”๋ก  ๋ฐฉ์‹์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ์ž‘์—…(์˜ˆ: ๊ฐ์ • ๋ถ„์„ ๋ฐ ์ถ”์ถœํ˜• ์งˆ๋ฌธ ๋‹ต๋ณ€)์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ์ž‘์—…์—์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[69, 95].
\(\bullet\)_ํฌ๋งท ๋””์ž์ธ._ ์ž์—ฐ์–ด ํ˜•์‹์˜ ๋””์ž์ธ๋„ LLM์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์ค‘์š”ํ•œ ์š”์†Œ์ž…๋‹ˆ๋‹ค[88]. ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์— ์ž‘์—… ์„ค๋ช…๊ณผ ์„ ํƒ์  ๋ฐ๋ชจ๋ฅผ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ž‘์—… ์„ค๋ช…์€ LLM์ด ์ž‘์—…์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๋ถ€๋ถ„์ž…๋‹ˆ๋‹ค[88]. ๋˜ํ•œ ์ ์ ˆํ•œ ์ˆ˜์˜ ์˜ˆ์‹œ๋ฅผ ๋ฐ๋ชจ๋กœ ์‚ฌ์šฉํ•˜๋ฉด ์ƒ๋‹นํ•œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ[69], ์ด๋Š” ๋ชจ๋ธ์ด ์ง€์‹œ ์—”์ง€๋‹ˆ์–ด๋ง์— ๋Œ€ํ•œ ๋ฏผ๊ฐ๋„๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ์—๋„ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[67, 69]. ๊ทธ๋Ÿฌ๋‚˜ ์ง€์นจ์— ๋‹ค๋ฅธ ๊ตฌ์„ฑ ์š”์†Œ(_์˜ˆ:_ ํ”ผํ•ด์•ผ ํ•  ์‚ฌํ•ญ, ์ด์œ  ๋ฐ ์ œ์•ˆ)๋ฅผ ํฌํ•จํ•˜๋ฉด LLM์˜ ์„ฑ๋Šฅ์— ๋ฏธ๋ฏธํ•˜๊ฑฐ๋‚˜ ์‹ฌ์ง€์–ด ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[88, 166]. ์ตœ๊ทผ์—๋Š” LLM์˜ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ์‚ฐ์ˆ  ์ถ”๋ก ๊ณผ ๊ฐ™์€ ์ผ๋ถ€ ์ถ”๋ก  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT) ์˜ˆ์ œ๋ฅผ ํฌํ•จํ•˜๋„๋ก ์ œ์•ˆํ•˜๋Š” ์ž‘์—…[69]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CoT์™€ ๋น„ CoT ์˜ˆ์ œ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋ฉด ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ์ž‘์—…(์˜ˆ: ์ƒ์‹ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก )๊ณผ ๊ทธ๋Ÿฌํ•œ ์ถ”๋ก  ๋ฐฉ์‹์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ์ž‘์—…(์˜ˆ: ๊ฐ์ • ๋ถ„์„ ๋ฐ ์ถ”์ถœํ˜• ์งˆ๋ฌธ ๋‹ต๋ณ€)์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ์ž‘์—…์—์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[69, 95].
Learning to plan in synthetic environments.An important feature of the agent in Autonomous Machine Intelligence (LeCun, 2022) is planning--an ability of decomposing complex tasks into subtasks and finishing the subtasks in a reward-optimal way (Kambhampati et al., 2024). Synthetic data can be a valuable tool here as it can serve as the feedback signal collected from a simulator (Park et al., 2023), and learning on it can make the agent aware of affordances (Ahn et al., 2022; Liang et al., 2022). For example, Inner Monologue (Huang et al., 2022) leverages natural language form feedback generated by the simulated environment to teach LLM-based robots planning. They find that such feedback significantly improves high-level instruction completion on both simulated and real-world domains. To compose a large number of realistic planning tasks (e.g., _"Rearrange objects on a table to match a given scene."_), VIMA (Jiang et al., 2022) creates a multi-modality simulated environment called VIMA-Bench, which supports extensible collections of objects and textures. In the Minecraft game, Voyager (Wang et al., 2023) deploys a number of GPT-4 based agents to interact with the synthetic environment and finds that the agents can unlock new skills faster and complete planning more efficiently with the help of synthetic feedback.
ํ•ฉ์„ฑ ํ™˜๊ฒฝ์—์„œ ๊ณ„ํšํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์šฐ๋Š” ๊ฒƒ. Autonomous Machine Intelligence(LeCun, 2022)์—์„œ ์—์ด์ „ํŠธ์˜ ์ค‘์š”ํ•œ ํŠน์ง• ์ค‘ ํ•˜๋‚˜๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•˜์œ„ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•˜๊ณ  ๋ณด์ƒ ์ตœ์ ์˜ ๋ฐฉ์‹์œผ๋กœ ํ•˜์œ„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๊ณ„ํš ๋Šฅ๋ ฅ(Kambhampati et al., 2024)์ž…๋‹ˆ๋‹ค. Synthetic data๋Š” ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ์—์„œ ์ˆ˜์ง‘ํ•œ ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—ฌ๊ธฐ์„œ ์œ ์šฉํ•œ ๋„๊ตฌ๊ฐ€ ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ(Park et al., 2023), ์ด๋ฅผ ํ•™์Šตํ•˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ affordances๋ฅผ ์ธ์‹ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Ahn et al., 2022; Liang et al., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, Inner Monologue(Huang et al., 2022)๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ํ™˜๊ฒฝ์—์„œ ์ƒ์„ฑ๋œ ์ž์—ฐ์–ด ํ˜•์‹์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜์—ฌ LLM ๊ธฐ๋ฐ˜ ๋กœ๋ด‡์˜ ๊ณ„ํš์„ ๊ฐ€๋ฅด์นฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ ์‹ค์ œ ๋„๋ฉ”์ธ ๋ชจ๋‘์—์„œ ์ƒ์œ„ ์ˆ˜์ค€์˜ instruction ์™„๋ฃŒ๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์‚ฌ๋ก€(์˜ˆ: _"์ฃผ์–ด์ง„ ์žฅ๋ฉด๊ณผ ์ผ์น˜ํ•˜๋„๋ก ํ…Œ์ด๋ธ” ์œ„์˜ ๋ฌผ์ฒด๋ฅผ ์žฌ๋ฐฐ์น˜ํ•˜์„ธ์š”."_)์˜ ์‚ฌ์‹ค์ ์ธ ๊ณ„ํš ์ž‘์—…์„ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด VIMA(Jiang et al., 2022)๋Š” ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฌผ์ฒด ๋ฐ ํ…์Šค์ฒ˜ ์ปฌ๋ ‰์…˜์„ ์ง€์›ํ•˜๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์ธ VIMA-Bench๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ ๊ฒŒ์ž„์—์„œ Voyager(Wang et al., 2023)๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ GPT-4 ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ๋ฐฐํฌํ•˜์—ฌ ํ•ฉ์„ฑ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๊ณ , ์—์ด์ „ํŠธ๊ฐ€ synthetic feedback์˜ ๋„์›€์„ ๋ฐ›์•„ ์ƒˆ๋กœ์šด ์Šคํ‚ฌ์„ ๋” ๋นจ๋ฆฌ ์ž ๊ธˆ ํ•ด์ œํ•˜๊ณ  ๊ณ„ํš์„ ๋” ํšจ์œจ์ ์œผ๋กœ ์™„๋ฃŒํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ํ™˜๊ฒฝ์—์„œ ๊ณ„ํšํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์šฐ๋Š” ๊ฒƒ.์ž์œจ ๋จธ์‹  ์ธํ…”๋ฆฌ์ „์Šค(LeCun, 2022)์—์„œ ์—์ด์ „ํŠธ์˜ ์ค‘์š”ํ•œ ํŠน์ง• ์ค‘ ํ•˜๋‚˜๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•˜์œ„ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•˜๊ณ  ๋ณด์ƒ ์ตœ์ ์˜ ๋ฐฉ์‹์œผ๋กœ ํ•˜์œ„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๊ณ„ํš ๋Šฅ๋ ฅ(Kambhampati et al., 2024)์ž…๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ์—์„œ ์ˆ˜์ง‘ํ•œ ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—ฌ๊ธฐ์„œ ์œ ์šฉํ•œ ๋„๊ตฌ๊ฐ€ ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ(Park et al., 2023), ์ด๋ฅผ ํ•™์Šตํ•˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ์–ดํฌ๋˜์Šค๋ฅผ ์ธ์‹ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Ahn et al., 2022; Liang et al., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด๋„ˆ ๋ชจ๋†€๋กœ๊ทธ(Huang et al., 2022)๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ํ™˜๊ฒฝ์—์„œ ์ƒ์„ฑ๋œ ์ž์—ฐ์–ด ํ˜•์‹์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜์—ฌ LLM ๊ธฐ๋ฐ˜ ๋กœ๋ด‡์˜ ๊ณ„ํš์„ ๊ฐ€๋ฅด์นฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ ์‹ค์ œ ๋„๋ฉ”์ธ ๋ชจ๋‘์—์„œ ์ƒ์œ„ ์ˆ˜์ค€์˜ ์ง€์นจ ์™„๋ฃŒ๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์‚ฌ๋ก€(์˜ˆ: _"์ฃผ์–ด์ง„ ์žฅ๋ฉด๊ณผ ์ผ์น˜ํ•˜๋„๋ก ํ…Œ์ด๋ธ” ์œ„์˜ ๋ฌผ์ฒด๋ฅผ ์žฌ๋ฐฐ์น˜ํ•˜์„ธ์š”."_)์˜ ์‚ฌ์‹ค์ ์ธ ๊ณ„ํš ์ž‘์—…์„ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด VIMA(Jiang et al., 2022)๋Š” ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฌผ์ฒด ๋ฐ ํ…์Šค์ฒ˜ ์ปฌ๋ ‰์…˜์„ ์ง€์›ํ•˜๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์ธ VIMA-Bench๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ ๊ฒŒ์ž„์—์„œ Voyager(Wang et al., 2023)๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ GPT-4 ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ๋ฐฐํฌํ•˜์—ฌ ํ•ฉ์„ฑ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๊ณ , ์—์ด์ „ํŠธ๊ฐ€ ํ•ฉ์„ฑ ํ”ผ๋“œ๋ฐฑ์˜ ๋„์›€์„ ๋ฐ›์•„ ์ƒˆ๋กœ์šด ์Šคํ‚ฌ์„ ๋” ๋นจ๋ฆฌ ์ž ๊ธˆ ํ•ด์ œํ•˜๊ณ  ๊ณ„ํš์„ ๋” ํšจ์œจ์ ์œผ๋กœ ์™„๋ฃŒํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
We conduct a systematic study on the premise order effect using a variety of SoTA LLMs, including GPT-4-turbo, GPT-3.5-turbo (OpenAI, 2023), PaLM 2-L (Google, 2023), and Gemini Pro (Gemini, 2023). Our primary focus is deductive reasoning, and we benchmark all LLMs on problems that only involve _modus ponens_ (if P then Q; P; therefore Q), where all LLMs in our evaluation at least achieve decent performance with a small number of premises. We show that the accuracy decrease caused by different ordering can be more than 30%. The ordering effect is further amplified when irrelevant premises (i.e., premises that are not needed to derive a conclusion) are presented in the prompt. Figure 1 illustrates a failure case, where all LLMs fail to generate the proof after changing the order of relevant rules. Interestingly, while all LLMs perform best when the premise order follows the ground truth proof, they reveal different preferences on other alternative orderings. Specifically, compared to randomly ordering the premises, GPT-4-turbo and GPT-3.5-turbo generally achieve betterperformance when the premise order is exactly the reverse of the ground truth proof, which enables LLMs to perform derivation via backward chaining. On the other hand, PaLM 2-L generally achieves the **worst performance** with such a reversed order.
GPT-4-turbo, GPT-3.5-turbo (OpenAI, 2023), PaLM 2-L (Google, 2023), Gemini Pro (Gemini, 2023) ๋“ฑ ๋‹ค์–‘ํ•œ SoTA LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ œ ์ˆœ์„œ ํšจ๊ณผ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ฃผ์š” ์ดˆ์ ์€ ์—ฐ์—ญ์  ์ถ”๋ก ์ด๋ฉฐ, ๋ชจ๋“  LLM์ด ์ ์–ด๋„ ์†Œ์ˆ˜์˜ ์ „์ œ๋งŒ์œผ๋กœ๋„ ์ ์ ˆํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” _modus ponens_ (if P then Q; P; therefore Q)๋งŒ ํฌํ•จํ•˜๋Š” ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋ชจ๋“  LLM์„ ๋ฒค์น˜๋งˆํ‚นํ•ฉ๋‹ˆ๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ์ˆœ์„œ๋กœ ์ธํ•œ ์ •ํ™•๋„ ๊ฐ์†Œ๋Š” 30% ์ด์ƒ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ด€๋ จ ์—†๋Š” ์ „์ œ(์ฆ‰, ๊ฒฐ๋ก ์„ ๋„์ถœํ•˜๋Š” ๋ฐ ํ•„์š”ํ•˜์ง€ ์•Š์€ ์ „์ œ)๊ฐ€ prompt์— ์ œ์‹œ๋˜๋ฉด ์ˆœ์„œ ํšจ๊ณผ๋Š” ๋”์šฑ ์ฆํญ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์€ ๋ชจ๋“  LLM์ด ๊ด€๋ จ ๊ทœ์น™์˜ ์ˆœ์„œ๋ฅผ ๋ณ€๊ฒฝํ•œ ํ›„ ์ฆ๋ช…์„ ์ƒ์„ฑํ•˜์ง€ ๋ชปํ•˜๋Š” ์‹คํŒจ ์‚ฌ๋ก€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋ชจ๋“  LLM์ด ์ „์ œ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์„ ๋”ฐ๋ฅผ ๋•Œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ, ๋‹ค๋ฅธ ๋Œ€์ฒด ์ˆœ์„œ์— ๋Œ€ํ•ด์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜ธ๋„๋ฅผ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. ํŠนํžˆ, ์ „์ œ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ •๋ ฌํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด GPT-4-turbo ๋ฐ GPT-3.5-turbo๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์˜ ์ •ํ™•ํ•œ ๋ฐ˜๋Œ€์ผ ๋•Œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์—ญ๋ฐฉํ–ฅ ์ฒด์ธ์œผ๋กœ ๋„์ถœ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— PaLM 2-L์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์—ญ์ˆœ์œผ๋กœ **์ตœ์•…์˜ ์„ฑ๋Šฅ**์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
GPT-4-turbo, GPT-3.5-turbo(OpenAI, 2023), PaLM 2-L(Google, 2023), Gemini Pro(Gemini, 2023) ๋“ฑ ๋‹ค์–‘ํ•œ SoTA LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ œ ์ˆœ์„œ ํšจ๊ณผ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ฃผ์š” ์ดˆ์ ์€ ์—ฐ์—ญ์  ์ถ”๋ก ์ด๋ฉฐ, ๋ชจ๋“  LLM์ด ์ ์–ด๋„ ์†Œ์ˆ˜์˜ ์ „์ œ๋งŒ์œผ๋กœ๋„ ์ ์ ˆํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_(P์ด๋ฉด Q; P; ๋”ฐ๋ผ์„œ Q)๋งŒ ํฌํ•จํ•˜๋Š” ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋ชจ๋“  LLM์„ ๋ฒค์น˜๋งˆํ‚นํ•ฉ๋‹ˆ๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ์ˆœ์„œ๋กœ ์ธํ•œ ์ •ํ™•๋„ ๊ฐ์†Œ๋Š” 30% ์ด์ƒ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ด€๋ จ ์—†๋Š” ์ „์ œ(์ฆ‰, ๊ฒฐ๋ก ์„ ๋„์ถœํ•˜๋Š” ๋ฐ ํ•„์š”ํ•˜์ง€ ์•Š์€ ์ „์ œ)๊ฐ€ ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋˜๋ฉด ์ˆœ์„œ ํšจ๊ณผ๋Š” ๋”์šฑ ์ฆํญ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์€ ๋ชจ๋“  LLM์ด ๊ด€๋ จ ๊ทœ์น™์˜ ์ˆœ์„œ๋ฅผ ๋ณ€๊ฒฝํ•œ ํ›„ ์ฆ๋ช…์„ ์ƒ์„ฑํ•˜์ง€ ๋ชปํ•˜๋Š” ์‹คํŒจ ์‚ฌ๋ก€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋ชจ๋“  LLM์ด ์ „์ œ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์„ ๋”ฐ๋ฅผ ๋•Œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ, ๋‹ค๋ฅธ ๋Œ€์ฒด ์ˆœ์„œ์— ๋Œ€ํ•ด์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜ธ๋„๋ฅผ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. ํŠนํžˆ, ์ „์ œ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ •๋ ฌํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด GPT-4-turbo ๋ฐ GPT-3.5-turbo๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์˜ ์ •ํ™•ํ•œ ๋ฐ˜๋Œ€์ผ ๋•Œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์—ญ๋ฐฉํ–ฅ ์ฒด์ธ์œผ๋กœ ๋„์ถœ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— PaLM 2-L์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์—ญ์ˆœ์œผ๋กœ **์ตœ์•…์˜ ์„ฑ๋Šฅ**์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
The State Space Model originates from Kalman filtering [35] which mainly introduces a linear filtering and prediction method. Kalman filtering can be divided into two steps, i.e., the prediction and correction step. The prediction is to estimate the current state based on the state of the previous time, and the correction is to estimate the optimal state by integrating the estimated state and the observed state of the current time. The State Space Model is a mathematical model that describes the behavior of a dynamic system using a set of first-order differential equations (continuous-time systems) or difference equations (discrete-time systems) to represent the evolution of the internal state of the system, and another set of equations to describe the relationship between the state and the output of the system. These equations can be expressed in matrix and vector form to deal with multivariable systems. Subsequently, Gu et al. [34] introduces a Linear State Space Layer (LSSL) that combines the advantages of recurrent neural networks (RNNs), temporal convolutional networks, and neural differential equations (NDEs) while addressing their shortcomings in model power and computational efficiency. This new sequence model is inspired by control systems and implemented through the linear state space layer (LSSL).
์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์€ ์ฃผ๋กœ ์„ ํ˜• ํ•„ํ„ฐ๋ง ๋ฐ ์˜ˆ์ธก ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๋Š” Kalman filtering[35]์—์„œ ์œ ๋ž˜ํ–ˆ์Šต๋‹ˆ๋‹ค. Kalman filtering์€ ์˜ˆ์ธก ๋‹จ๊ณ„์™€ ๋ณด์ • ๋‹จ๊ณ„์˜ ๋‘ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์ธก์€ ์ด์ „ ์‹œ๊ฐ„์˜ ์ƒํƒœ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ๋ณด์ •์€ ์ถ”์ •๋œ ์ƒํƒœ์™€ ํ˜„์žฌ ์‹œ๊ฐ„์˜ ๊ด€์ธก ์ƒํƒœ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์ตœ์ ์˜ ์ƒํƒœ๋ฅผ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์€ ๋™์  ์‹œ์Šคํ…œ์˜ ๋™์ž‘์„ ์„ค๋ช…ํ•˜๋Š” ์ˆ˜ํ•™์  ๋ชจ๋ธ๋กœ, ์‹œ์Šคํ…œ์˜ ๋‚ด๋ถ€ ์ƒํƒœ์˜ ์ง„ํ™”๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ผ๋ จ์˜ 1์ฐจ ๋ฏธ๋ถ„ ๋ฐฉ์ •์‹(์—ฐ์† ์‹œ๊ฐ„ ์‹œ์Šคํ…œ) ๋˜๋Š” ์ฐจ์ด ๋ฐฉ์ •์‹(๋ถˆ์—ฐ์† ์‹œ๊ฐ„ ์‹œ์Šคํ…œ)๊ณผ ์ƒํƒœ์™€ ์‹œ์Šคํ…œ์˜ ์ถœ๋ ฅ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ๋ฐฉ์ •์‹์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์ •์‹์€ ํ–‰๋ ฌ๊ณผ ๋ฒกํ„ฐ ํ˜•์‹์œผ๋กœ ํ‘œํ˜„ํ•˜์—ฌ ๋‹ค๋ณ€์ˆ˜ ์‹œ์Šคํ…œ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์–ด์„œ Gu ๋“ฑ[34]์€ ์ˆœํ™˜ ์‹ ๊ฒฝ๋ง(RNN), temporal convolutional networks, neural differential equations (NDE)์˜ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•˜๋ฉด์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ ๊ณ„์‚ฐ ํšจ์œจ์„ฑ์˜ ๋‹จ์ ์„ ํ•ด๊ฒฐํ•œ Linear State Space Layer (LSSL)์„ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ์‹œํ€€์Šค ๋ชจ๋ธ์€ ์ œ์–ด ์‹œ์Šคํ…œ์—์„œ ์˜๊ฐ์„ ์–ป์–ด Linear State Space Layer (LSSL)์„ ํ†ตํ•ด ๊ตฌํ˜„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์€ ์ฃผ๋กœ ์„ ํ˜• ํ•„ํ„ฐ๋ง ๋ฐ ์˜ˆ์ธก ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๋Š” ์นผ๋งŒ ํ•„ํ„ฐ๋ง[35]์—์„œ ์œ ๋ž˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์นผ๋งŒ ํ•„ํ„ฐ๋ง์€ ์˜ˆ์ธก ๋‹จ๊ณ„์™€ ๋ณด์ • ๋‹จ๊ณ„์˜ ๋‘ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์ธก์€ ์ด์ „ ์‹œ๊ฐ„์˜ ์ƒํƒœ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ๋ณด์ •์€ ์ถ”์ •๋œ ์ƒํƒœ์™€ ํ˜„์žฌ ์‹œ๊ฐ„์˜ ๊ด€์ธก ์ƒํƒœ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์ตœ์ ์˜ ์ƒํƒœ๋ฅผ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์€ ๋™์  ์‹œ์Šคํ…œ์˜ ๋™์ž‘์„ ์„ค๋ช…ํ•˜๋Š” ์ˆ˜ํ•™์  ๋ชจ๋ธ๋กœ, ์‹œ์Šคํ…œ์˜ ๋‚ด๋ถ€ ์ƒํƒœ์˜ ์ง„ํ™”๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ผ๋ จ์˜ 1์ฐจ ๋ฏธ๋ถ„ ๋ฐฉ์ •์‹(์—ฐ์† ์‹œ๊ฐ„ ์‹œ์Šคํ…œ) ๋˜๋Š” ์ฐจ์ด ๋ฐฉ์ •์‹(๋ถˆ์—ฐ์† ์‹œ๊ฐ„ ์‹œ์Šคํ…œ)๊ณผ ์ƒํƒœ์™€ ์‹œ์Šคํ…œ์˜ ์ถœ๋ ฅ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ๋ฐฉ์ •์‹์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์ •์‹์€ ํ–‰๋ ฌ๊ณผ ๋ฒกํ„ฐ ํ˜•์‹์œผ๋กœ ํ‘œํ˜„ํ•˜์—ฌ ๋‹ค๋ณ€์ˆ˜ ์‹œ์Šคํ…œ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์–ด์„œ Gu ๋“ฑ[34]์€ ์ˆœํ™˜ ์‹ ๊ฒฝ๋ง(RNN), ์‹œ๊ฐ„์  ์ปจ๋ณผ๋ฃจ์…˜ ๋„คํŠธ์›Œํฌ, ์‹ ๊ฒฝ ๋ฏธ๋ถ„ ๋ฐฉ์ •์‹(NDE)์˜ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•˜๋ฉด์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ ๊ณ„์‚ฐ ํšจ์œจ์„ฑ์˜ ๋‹จ์ ์„ ํ•ด๊ฒฐํ•œ ์„ ํ˜• ์ƒํƒœ ๊ณต๊ฐ„ ๊ณ„์ธต(LSSL)์„ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ์‹œํ€€์Šค ๋ชจ๋ธ์€ ์ œ์–ด ์‹œ์Šคํ…œ์—์„œ ์˜๊ฐ์„ ์–ป์–ด ์„ ํ˜• ์ƒํƒœ ๊ณต๊ฐ„ ๊ณ„์ธต(LSSL)์„ ํ†ตํ•ด ๊ตฌํ˜„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
\begin{table} \begin{tabular}{l|l} \hline \hline \multicolumn{1}{c}{**Transformation**} & \multicolumn{1}{c}{**Training example**} \\ \hline None & Cruise was born on July 3, 1962, in Syracuse, New York, to Mary \\ & Lee Pfeiffer. \\ \hline Word reversal & Preiffer Lee Mary to, York New, Syracuse in, 1962, 3 July on born \\ & was Cruise \\ \hline Entity-preserving reversal & Mary Lee Pfeiffer to, Syracuse, New York in, 1962, 3 July on born \\ & was Cruise \\ \hline Random segment reversal & [REV] York, to Mary Lee Pfeiffer. [REV] in Syracuse, New [REV] \\ & on July 3, 1962, [REV] born [REV] Cruise was \\ \hline \hline \end{tabular} \end{table} Table 2: **Reversal transformations: examples of different reversal types on a given string. In practice, training examples can be much longer (e.g., entire documents during pre-training). The language model is still trained left-to-right on such transformations, and in the word reversal case is essentially predicting the sentence backwards (right-to-left) starting from the last word. The entities which have their word ordering preserved are highlighted by underlines. In random segment reversal, the segments are separated by โ€œ[REV]โ€. _Reverse training_ on both the standard (โ€Noneโ€ transformation) and reversed examples, hence doubling the amount of training tokens.
\begin{table} \begin{tabular}{l|l} \hline \hline \multicolumn{1}{c}{**Transformation**} & \multicolumn{1}{c}{**Training example**} \\ \hline None & ํฌ๋ฃจ์ฆˆ๋Š” 1962๋…„ 7์›” 3์ผ ๋‰ด์š•์ฃผ ์‹œ๋Ÿฌํ์Šค์—์„œ ๋ฉ”๋ฆฌ \\ & ๋ฆฌ ํŒŒ์ดํผ. \\ \hline Word reversal & Preiffer Lee Mary to, ๋‰ด์š•์ฃผ ์‹œ๋Ÿฌํ์Šค, 1962๋…„ 7์›” 3์ผ์ƒ \\ & was Cruise \\ \hline Entity-preserving reversal & Mary Lee Pfeiffer to, ๋‰ด์š•์ฃผ ์‹œ๋Ÿฌํ์Šค, 1962๋…„ 7์›” 3์ผ์ƒ \\ & was Cruise \\ \hline Random segment reversal & [REV] ๋‰ด์š•, ๋ฉ”๋ฆฌ ๋ฆฌ ํŒŒ์ดํผ์—๊ฒŒ. [REV] ์‹œ๋Ÿฌํ์Šค์—์„œ, ๋‰ด์š• [REV] \\ & 1962๋…„ 7์›” 3์ผ, [REV] ํฌ๋ฃจ์ฆˆ๊ฐ€ [REV] ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค \\ \hline \hline \end{tabular} \end{table} ํ‘œ 2: **Reversal transformations: ์ฃผ์–ด์ง„ ๋ฌธ์ž์—ด์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ์—ญ๋ณ€ํ™˜ ์œ ํ˜•์˜ ์˜ˆ์‹œ. ์‹ค์ œ๋กœ Training examples๋Š” ํ›จ์”ฌ ๋” ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘ ์ „์ฒด ๋ฌธ์„œ). The language model์€ ์ด๋Ÿฌํ•œ ๋ณ€ํ™˜์— ๋Œ€ํ•ด ์—ฌ์ „ํžˆ ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ํ›ˆ๋ จ๋˜๋ฉฐ, ๋‹จ์–ด ๋ฐ˜์ „์˜ ๊ฒฝ์šฐ ๋งˆ์ง€๋ง‰ ๋‹จ์–ด๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ๋ฌธ์žฅ์„ ๊ฑฐ๊พธ๋กœ(์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ) ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹จ์–ด ์ˆœ์„œ๊ฐ€ ์œ ์ง€๋˜๋Š” ์—”ํ‹ฐํ‹ฐ๋Š” ๋ฐ‘์ค„๋กœ ๊ฐ•์กฐ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„ ์„ธ๊ทธ๋จผํŠธ ๋ฐ˜์ „์˜ ๊ฒฝ์šฐ ์„ธ๊ทธ๋จผํŠธ๋Š” "[REV]"๋กœ ๊ตฌ๋ถ„๋ฉ๋‹ˆ๋‹ค. ํ‘œ์ค€("None" ๋ณ€ํ™˜)๊ณผ ๋ฐ˜์ „๋œ ์˜ˆ์ œ ๋ชจ๋‘์— ๋Œ€ํ•ด _Reverse training_์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ํ›ˆ๋ จ ํ† ํฐ์˜ ์–‘์„ ๋‘ ๋ฐฐ๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค.
\begin{table} \begin{tabular}{l|l} \hline \hline \multicolumn{1}{c}{**๋ณ€ํ™˜**} & \multicolumn{1}{c}{**ํ›ˆ๋ จ ์˜ˆ์‹œ**} \\ \hline ์—†์Œ & ํฌ๋ฃจ์ฆˆ๋Š” 1962๋…„ 7์›” 3์ผ ๋‰ด์š•์ฃผ ์‹œ๋Ÿฌํ์Šค์—์„œ ๋ฉ”๋ฆฌ \\ & ๋ฆฌ ํŒŒ์ดํผ. \\ \hline ๋‹จ์–ด ๋ฐ˜์ „ & Preiffer Lee Mary to, ๋‰ด์š•์ฃผ ์‹œ๋Ÿฌํ์Šค, 1962๋…„ 7์›” 3์ผ์ƒ \\ & was Cruise \\ \hline ์—”ํ‹ฐํ‹ฐ ๋ณด์กด ๋ฐ˜์ „ & Mary Lee Pfeiffer to, ๋‰ด์š•์ฃผ ์‹œ๋Ÿฌํ์Šค, 1962๋…„ 7์›” 3์ผ์ƒ \\ & was Cruise \\ \hline ๋ฌด์ž‘์œ„ ์„ธ๊ทธ๋จผํŠธ ๋ฐ˜์ „ & [REV] ๋‰ด์š•, ๋ฉ”๋ฆฌ ๋ฆฌ ํŒŒ์ดํผ์—๊ฒŒ. [REV] ์‹œ๋Ÿฌํ์Šค์—์„œ, ๋‰ด์š• [REV] \\ & 1962๋…„ 7์›” 3์ผ, [REV] ํฌ๋ฃจ์ฆˆ๊ฐ€ [REV] ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค \\ \hline \hline \end{tabular} \end{table} ํ‘œ 2: **์—ญ๋ณ€ํ™˜: ์ฃผ์–ด์ง„ ๋ฌธ์ž์—ด์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ์—ญ๋ณ€ํ™˜ ์œ ํ˜•์˜ ์˜ˆ์‹œ. ์‹ค์ œ๋กœ ํ›ˆ๋ จ ์˜ˆ๋Š” ํ›จ์”ฌ ๋” ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘ ์ „์ฒด ๋ฌธ์„œ). ์–ธ์–ด ๋ชจ๋ธ์€ ์ด๋Ÿฌํ•œ ๋ณ€ํ™˜์— ๋Œ€ํ•ด ์—ฌ์ „ํžˆ ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ํ›ˆ๋ จ๋˜๋ฉฐ, ๋‹จ์–ด ๋ฐ˜์ „์˜ ๊ฒฝ์šฐ ๋งˆ์ง€๋ง‰ ๋‹จ์–ด๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ๋ฌธ์žฅ์„ ๊ฑฐ๊พธ๋กœ(์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ) ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹จ์–ด ์ˆœ์„œ๊ฐ€ ์œ ์ง€๋˜๋Š” ์—”ํ‹ฐํ‹ฐ๋Š” ๋ฐ‘์ค„๋กœ ๊ฐ•์กฐ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„ ์„ธ๊ทธ๋จผํŠธ ๋ฐ˜์ „์˜ ๊ฒฝ์šฐ ์„ธ๊ทธ๋จผํŠธ๋Š” "[REV]"๋กœ ๊ตฌ๋ถ„๋ฉ๋‹ˆ๋‹ค. ํ‘œ์ค€("์—†์Œ" ๋ณ€ํ™˜)๊ณผ ๋ฐ˜์ „๋œ ์˜ˆ์ œ ๋ชจ๋‘์— ๋Œ€ํ•ด _์—ญ๋ฐฉํ–ฅ ํ›ˆ๋ จ_์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ํ›ˆ๋ จ ํ† ํฐ์˜ ์–‘์„ ๋‘ ๋ฐฐ๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค.
History-taking and the clinical interview are widely taught in both medical schools' and postgraduate curricula [37, 38, 39, 40, 41, 42]. Consensus on physician-patient communication has evolved to embrace patient-centred communication practices, with recommendations that communication in clinical encounters should address six core functions: fostering the relationship, gathering information, providing information, making decisions, responding to emotions and enabling disease- and treatment-related behavior [43, 44, 20]. Specific skills and behaviours for meeting these goals have also been described, taught and assessed [45, 20] with validated tools [45]. Medical conventions consistently cite that certain categories of information should be gathered during a clinical interview, comprising topics such as the presenting complaint, past medical history and medication history, social and family history, and systems review [46, 47]. Clinicians' ability to meet these goals is commonly assessed using the framework of an objective structured clinical examination (OSCE) [31, 32, 33]. Such assessments vary in their reproducibility or implementation and have even been adapted for remote practice as virtual OSCEs (vOSCEs) with telemedical scenarios, an issue of particular relevance during the COVID-19 pandemic [48].
๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ž„์ƒ ๋ฉด๋‹ด์€ ์˜๊ณผ๋Œ€ํ•™ ๋ฐ ๋Œ€ํ•™์› ๊ณผ์ •์—์„œ ๋„๋ฆฌ ๊ฐ€๋ฅด์น˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[37, 38, 39, 40, 41, 42]. ์˜์‚ฌ-ํ™˜์ž ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์— ๋Œ€ํ•œ ํ•ฉ์˜๋Š” ํ™˜์ž ์ค‘์‹ฌ์˜ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ด€ํ–‰์„ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ง„ํ™”ํ•˜์—ฌ ์ž„์ƒ ๋งŒ๋‚จ์—์„œ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์ด 6๊ฐ€์ง€ ํ•ต์‹ฌ ๊ธฐ๋Šฅ์„ ๋‹ค๋ฃจ์–ด์•ผ ํ•œ๋‹ค๋Š” ๊ถŒ์žฅ ์‚ฌํ•ญ์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค: ๊ด€๊ณ„ ์กฐ์„ฑ, ์ •๋ณด ์ˆ˜์ง‘, ์ •๋ณด ์ œ๊ณต, ์˜์‚ฌ ๊ฒฐ์ •, ๊ฐ์ •์— ๋Œ€ํ•œ ๋Œ€์‘, ์งˆ๋ณ‘ ๋ฐ ์น˜๋ฃŒ ๊ด€๋ จ ํ–‰๋™ ๊ฐ€๋Šฅ[43, 44, 20]. ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ๊ตฌ์ฒด์ ์ธ ๊ธฐ์ˆ ๊ณผ ํ–‰๋™๋„ ๊ฒ€์ฆ๋œ ๋„๊ตฌ[45]๋ฅผ ํ†ตํ•ด ์„ค๋ช…๋˜๊ณ , ๊ฐ€๋ฅด์น˜๊ณ , ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค[45, 20]. ์˜ํ•™ ๊ทœ์น™์€ ์ž„์ƒ ๋ฉด๋‹ด ์ค‘์— ์ œ์‹œ๋˜๋Š” ๋ถˆ๋งŒ, ๊ณผ๊ฑฐ ๋ณ‘๋ ฅ ๋ฐ ์•ฝ๋ฌผ ๋ณต์šฉ๋ ฅ, ์‚ฌํšŒ ๋ฐ ๊ฐ€์กฑ๋ ฅ, ์‹œ์Šคํ…œ ๊ฒ€ํ† ์™€ ๊ฐ™์€ ์ฃผ์ œ๋กœ ๊ตฌ์„ฑ๋œ ํŠน์ • ๋ฒ”์ฃผ์˜ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ผ๊ด€๋˜๊ฒŒ ๊ทœ์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[46, 47]. ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ์ž„์ƒ์˜์˜ ๋Šฅ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ๊ด€์  ๊ตฌ์กฐํ™”๋œ ์ž„์ƒ ๊ฒ€์‚ฌ(Objective Structured Clinical Examination, OSCE)์˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค[31, 32, 33]. ์ด๋Ÿฌํ•œ ํ‰๊ฐ€๋Š” ์žฌํ˜„์„ฑ ๋˜๋Š” ๊ตฌํ˜„ ๋ฐฉ์‹์ด ๋‹ค์–‘ํ•˜๋ฉฐ, ์‹ฌ์ง€์–ด COVID-19 ํŒฌ๋ฐ๋ฏน ๊ธฐ๊ฐ„ ๋™์•ˆ ํŠนํžˆ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ์›๊ฒฉ ์ง„๋ฃŒ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํ†ตํ•ด ๊ฐ€์ƒ OSCE(virtual OSCE, vOSCE)๋กœ ๊ฐœ์กฐ๋˜๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค[48].
๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ž„์ƒ ๋ฉด๋‹ด์€ ์˜๊ณผ๋Œ€ํ•™ ๋ฐ ๋Œ€ํ•™์› ๊ณผ์ •์—์„œ ๋„๋ฆฌ ๊ฐ€๋ฅด์น˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[37, 38, 39, 40, 41, 42]. ์˜์‚ฌ-ํ™˜์ž ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์— ๋Œ€ํ•œ ํ•ฉ์˜๋Š” ํ™˜์ž ์ค‘์‹ฌ์˜ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ด€ํ–‰์„ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ง„ํ™”ํ•˜์—ฌ ์ž„์ƒ ๋งŒ๋‚จ์—์„œ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์ด 6๊ฐ€์ง€ ํ•ต์‹ฌ ๊ธฐ๋Šฅ์„ ๋‹ค๋ฃจ์–ด์•ผ ํ•œ๋‹ค๋Š” ๊ถŒ์žฅ ์‚ฌํ•ญ์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค: ๊ด€๊ณ„ ์กฐ์„ฑ, ์ •๋ณด ์ˆ˜์ง‘, ์ •๋ณด ์ œ๊ณต, ์˜์‚ฌ ๊ฒฐ์ •, ๊ฐ์ •์— ๋Œ€ํ•œ ๋Œ€์‘, ์งˆ๋ณ‘ ๋ฐ ์น˜๋ฃŒ ๊ด€๋ จ ํ–‰๋™ ๊ฐ€๋Šฅ[43, 44, 20]. ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ๊ตฌ์ฒด์ ์ธ ๊ธฐ์ˆ ๊ณผ ํ–‰๋™๋„ ๊ฒ€์ฆ๋œ ๋„๊ตฌ[45]๋ฅผ ํ†ตํ•ด ์„ค๋ช…๋˜๊ณ , ๊ฐ€๋ฅด์น˜๊ณ , ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค[45, 20]. ์˜ํ•™ ๊ทœ์น™์€ ์ž„์ƒ ๋ฉด๋‹ด ์ค‘์— ์ œ์‹œ๋˜๋Š” ๋ถˆ๋งŒ, ๊ณผ๊ฑฐ ๋ณ‘๋ ฅ ๋ฐ ์•ฝ๋ฌผ ๋ณต์šฉ๋ ฅ, ์‚ฌํšŒ ๋ฐ ๊ฐ€์กฑ๋ ฅ, ์‹œ์Šคํ…œ ๊ฒ€ํ† ์™€ ๊ฐ™์€ ์ฃผ์ œ๋กœ ๊ตฌ์„ฑ๋œ ํŠน์ • ๋ฒ”์ฃผ์˜ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ผ๊ด€๋˜๊ฒŒ ๊ทœ์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[46, 47]. ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ์ž„์ƒ์˜์˜ ๋Šฅ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ๊ด€์  ๊ตฌ์กฐํ™”๋œ ์ž„์ƒ ๊ฒ€์‚ฌ(OSCE)์˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค[31, 32, 33]. ์ด๋Ÿฌํ•œ ํ‰๊ฐ€๋Š” ์žฌํ˜„์„ฑ ๋˜๋Š” ๊ตฌํ˜„ ๋ฐฉ์‹์ด ๋‹ค์–‘ํ•˜๋ฉฐ, ์‹ฌ์ง€์–ด COVID-19 ํŒฌ๋ฐ๋ฏน ๊ธฐ๊ฐ„ ๋™์•ˆ ํŠนํžˆ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ์›๊ฒฉ ์ง„๋ฃŒ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํ†ตํ•ด ๊ฐ€์ƒ OSCE(vOSCE)๋กœ ๊ฐœ์กฐ๋˜๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค[48].
To evaluate Self-Instruct empirically, we run this framework on GPT3 (Brown et al., 2020), which is a vanilla LM (SS3). The iterative Self-Instruct process on this model leads to about 52k instructions, paired with about 82K instance inputs and target outputs. We observe that the resulting data provides a diverse range of creative tasks, as is demonstrated by examples in Figure 1. These generated tasks deviate from the distribution of typical NLP tasks, and also have fairly small overlap with the seed tasks (SS3.2). On this resulting data, we build GPT3SELF-INST by finetuning GPT3 (i.e., the same model used for generating the instruction data). We evaluate GPT3SELF-INST in comparison to various other models on both typical NLP tasks included in SuperNI (Wang et al., 2022), and a set of new instructions that are created for novel usage of instruction-following models (SS4). The results indicate that GPT3SELF-INST outperforms GPT3 (the original model) by a large margin (+33.1%) and nearly matches the performance of InstructGPT001. Moreover, our human evaluation on the newly-created instruction set shows that GPT3SELF-INST demonstrates a broad range of instruction following ability, outperforming models trained on other publicly available instruction datasets and leaving only a 5% gap behind InstructGPT001.
Self-Instruct๋ฅผ ๊ฒฝํ—˜์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ๋ฐ”๋‹๋ผ LM(SS3)์ธ GPT3(Brown et al., 2020)์—์„œ ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ฐ˜๋ณต์ ์ธ Self-Instruct ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ์•ฝ 52,000๊ฐœ์˜ instructions์™€ ์•ฝ 82,000๊ฐœ์˜ ์ธ์Šคํ„ด์Šค ์ž…๋ ฅ ๋ฐ ๋Œ€์ƒ ์ถœ๋ ฅ์ด ์ง์„ ์ด๋ฃน๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋ฐ์ดํ„ฐ๋Š” ๊ทธ๋ฆผ 1์˜ ์˜ˆ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋‹ค์–‘ํ•œ ์ฐฝ์˜์ ์ธ ์ž‘์—…์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒ์„ฑ๋œ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์ธ NLP ์ž‘์—…์˜ ๋ถ„ํฌ์—์„œ ๋ฒ—์–ด๋‚˜๋ฉฐ, ์‹œ๋“œ ์ž‘์—…(SS3.2)๊ณผ๋„ ์ƒ๋‹นํžˆ ๊ฒน์น˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ GPT3(์ฆ‰, instruction ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์— ์‚ฌ์šฉ๋œ ๋™์ผํ•œ ๋ชจ๋ธ)๋ฅผ Fine-tuningํ•˜์—ฌ GPT3SELF-INST๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. SuperNI(Wang et al., 2022)์— ํฌํ•จ๋œ ์ผ๋ฐ˜์ ์ธ NLP ์ž‘์—…๊ณผ instruction ๋ชจ๋ธ์˜ ์ƒˆ๋กœ์šด ์‚ฌ์šฉ์„ ์œ„ํ•ด ์ƒ์„ฑ๋œ ์ƒˆ๋กœ์šด instruction ์„ธํŠธ(SS4)์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ GPT3SELF-INST๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, GPT3SELF-INST๊ฐ€ GPT3(์›๋ž˜ ๋ชจ๋ธ)๋ณด๋‹ค ํฐ ํญ(+33.1%)์œผ๋กœ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ InstructGPT001์˜ ์„ฑ๋Šฅ์— ๊ฑฐ์˜ ๊ทผ์ ‘ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ƒˆ๋กœ ์ƒ์„ฑ๋œ instruction ์„ธํŠธ์— ๋Œ€ํ•œ ์ธ๊ฐ„ ํ‰๊ฐ€ ๊ฒฐ๊ณผ, GPT3SELF-INST๋Š” ๋‹ค๋ฅธ ๊ณต๊ฐœ๋œ instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ๊ด‘๋ฒ”์œ„ํ•œ instruction ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์œผ๋ฉฐ, InstructGPT001์— ๋น„ํ•ด 5%์˜ ๊ฒฉ์ฐจ๋งŒ ๋‚จ๊ฒผ์Šต๋‹ˆ๋‹ค.
Self-Instruct๋ฅผ ๊ฒฝํ—˜์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ๋ฐ”๋‹๋ผ LM(SS3)์ธ GPT3(Brown et al., 2020)์—์„œ ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ฐ˜๋ณต์ ์ธ Self-Instruct ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ์•ฝ 52,000๊ฐœ์˜ ๋ช…๋ น์–ด์™€ ์•ฝ 82,000๊ฐœ์˜ ์ธ์Šคํ„ด์Šค ์ž…๋ ฅ ๋ฐ ๋Œ€์ƒ ์ถœ๋ ฅ์ด ์ง์„ ์ด๋ฃน๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋ฐ์ดํ„ฐ๋Š” ๊ทธ๋ฆผ 1์˜ ์˜ˆ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋‹ค์–‘ํ•œ ์ฐฝ์˜์ ์ธ ์ž‘์—…์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒ์„ฑ๋œ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์ธ NLP ์ž‘์—…์˜ ๋ถ„ํฌ์—์„œ ๋ฒ—์–ด๋‚˜๋ฉฐ, ์‹œ๋“œ ์ž‘์—…(SS3.2)๊ณผ๋„ ์ƒ๋‹นํžˆ ๊ฒน์น˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ GPT3(์ฆ‰, ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์— ์‚ฌ์šฉ๋œ ๋™์ผํ•œ ๋ชจ๋ธ)๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ GPT3SELF-INST๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. SuperNI(Wang et al., 2022)์— ํฌํ•จ๋œ ์ผ๋ฐ˜์ ์ธ NLP ์ž‘์—…๊ณผ ๋ช…๋ น์–ด ๋ชจ๋ธ์˜ ์ƒˆ๋กœ์šด ์‚ฌ์šฉ์„ ์œ„ํ•ด ์ƒ์„ฑ๋œ ์ƒˆ๋กœ์šด ๋ช…๋ น์–ด ์„ธํŠธ(SS4)์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ GPT3SELF-INST๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, GPT3SELF-INST๊ฐ€ GPT3(์›๋ž˜ ๋ชจ๋ธ)๋ณด๋‹ค ํฐ ํญ(+33.1%)์œผ๋กœ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ InstructGPT001์˜ ์„ฑ๋Šฅ์— ๊ฑฐ์˜ ๊ทผ์ ‘ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ๋ช…๋ น์–ด ์„ธํŠธ์— ๋Œ€ํ•œ ์ธ๊ฐ„ ํ‰๊ฐ€ ๊ฒฐ๊ณผ, GPT3SELF-INST๋Š” ๋‹ค๋ฅธ ๊ณต๊ฐœ๋œ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์œผ๋ฉฐ, InstructGPT001์— ๋น„ํ•ด 5%์˜ ๊ฒฉ์ฐจ๋งŒ ๋‚จ๊ฒผ์Šต๋‹ˆ๋‹ค.
**Action Goal**: The agent can perform actions with various objectives. Here, we present several representative examples: (1) _Task Completion_. In this scenario, the agent's actions are aimed at accomplishing specific tasks, such as crafting an iron pick-axe in Minecraft [38] or completing a function in software development [18]. These actions usually have well-defined objectives, and each action contributes to the completion of the final task. Actions aimed at this type of goal are very common in existing literature. (2) _Communication_. In this case, the actions are taken to communicate with the other agents or real humans for sharing information or collaboration. For example, the agents in ChatDev [18] may communicate with each other to collectively accomplish software development tasks. In Inner Monologue [61], the agent actively engages in communication with humans and adjusts its action strategies based on human feedback. (3) _Environment Exploration_. In this example, the agent aims to explore unfamiliar environments to expand its perception and strike a balance between exploring and exploiting. For instance, the agent in Voyager [38] may explore unknown skills in their task completion process, and continually refine the skill execution code based on environment feedback through trial and error.
**Action Goal**: ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค: (1) _์ž‘์—… ์™„๋ฃŒ_. ์ด ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์€ ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ [38]์—์„œ ์ฒ  ๊ณก๊ดญ์ด๋ฅผ ์ œ์ž‘ํ•˜๊ฑฐ๋‚˜ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ [18]์—์„œ ํ•จ์ˆ˜๋ฅผ ์™„๋ฃŒํ•˜๋Š” ๋“ฑ ํŠน์ • ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ์ž˜ ์ •์˜๋œ ๋ชฉํ‘œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ๊ฐ ์ž‘์—…์€ ์ตœ์ข… ์ž‘์—… ์™„๋ฃŒ์— ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ชฉํ‘œ๋ฅผ ์œ„ํ•œ ํ–‰๋™์€ ๊ธฐ์กด ๋ฌธํ—Œ์—์„œ ๋งค์šฐ ํ”ํ•˜๊ฒŒ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (2) _Communication_. ์ด ๊ฒฝ์šฐ, ๋‹ค๋ฅธ ์—์ด์ „ํŠธ ๋˜๋Š” ์‹ค์ œ ์‚ฌ๋žŒ๊ณผ ์ •๋ณด๋ฅผ ๊ณต์œ ํ•˜๊ฑฐ๋‚˜ ํ˜‘์—…ํ•˜๊ธฐ ์œ„ํ•ด ํ–‰๋™์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatDev [18]์˜ ์—์ด์ „ํŠธ๋Š” ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ์ž‘์—…์„ ๊ณต๋™์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์„œ๋กœ ์†Œํ†ตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Inner Monologue [61]์—์„œ ์—์ด์ „ํŠธ๋Š” ์ธ๊ฐ„๊ณผ ์ ๊ทน์ ์œผ๋กœ ์†Œํ†ตํ•˜๊ณ  ์ธ๊ฐ„์˜ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ํ–‰๋™ ์ „๋žต์„ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. (3) _Environment Exploration_. ์ด ์˜ˆ์—์„œ ์—์ด์ „ํŠธ๋Š” ์ต์ˆ™ํ•˜์ง€ ์•Š์€ ํ™˜๊ฒฝ์„ ํƒ์ƒ‰ํ•˜์—ฌ ์ธ์‹์„ ํ™•์žฅํ•˜๊ณ  ํƒ์ƒ‰๊ณผ ์ฐฉ์ทจ ์‚ฌ์ด์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Voyager [38]์˜ ์—์ด์ „ํŠธ๋Š” ์ž‘์—… ์™„๋ฃŒ ๊ณผ์ •์—์„œ ์•Œ๋ ค์ง€์ง€ ์•Š์€ ์Šคํ‚ฌ์„ ํƒ์ƒ‰ํ•˜๊ณ  ์‹œํ–‰์ฐฉ์˜ค๋ฅผ ํ†ตํ•ด ํ™˜๊ฒฝ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ์Šคํ‚ฌ ์‹คํ–‰ ์ฝ”๋“œ๋ฅผ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์•ก์…˜ ๋ชฉํ‘œ**: ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค: (1) _์ž‘์—… ์™„๋ฃŒ_. ์ด ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์€ ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ [38]์—์„œ ์ฒ  ๊ณก๊ดญ์ด๋ฅผ ์ œ์ž‘ํ•˜๊ฑฐ๋‚˜ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ [18]์—์„œ ํ•จ์ˆ˜๋ฅผ ์™„๋ฃŒํ•˜๋Š” ๋“ฑ ํŠน์ • ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ์ž˜ ์ •์˜๋œ ๋ชฉํ‘œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ๊ฐ ์ž‘์—…์€ ์ตœ์ข… ์ž‘์—… ์™„๋ฃŒ์— ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ชฉํ‘œ๋ฅผ ์œ„ํ•œ ํ–‰๋™์€ ๊ธฐ์กด ๋ฌธํ—Œ์—์„œ ๋งค์šฐ ํ”ํ•˜๊ฒŒ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (2) _์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜_. ์ด ๊ฒฝ์šฐ, ๋‹ค๋ฅธ ์—์ด์ „ํŠธ ๋˜๋Š” ์‹ค์ œ ์‚ฌ๋žŒ๊ณผ ์ •๋ณด๋ฅผ ๊ณต์œ ํ•˜๊ฑฐ๋‚˜ ํ˜‘์—…ํ•˜๊ธฐ ์œ„ํ•ด ํ–‰๋™์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatDev [18]์˜ ์—์ด์ „ํŠธ๋Š” ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ์ž‘์—…์„ ๊ณต๋™์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์„œ๋กœ ์†Œํ†ตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‚ด๋Ÿฌ ๋ชจ๋†€๋กœ๊ทธ [61]์—์„œ ์—์ด์ „ํŠธ๋Š” ์ธ๊ฐ„๊ณผ ์ ๊ทน์ ์œผ๋กœ ์†Œํ†ตํ•˜๊ณ  ์ธ๊ฐ„์˜ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ํ–‰๋™ ์ „๋žต์„ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. (3) _ํ™˜๊ฒฝ ํƒ์ƒ‰_. ์ด ์˜ˆ์—์„œ ์—์ด์ „ํŠธ๋Š” ์ต์ˆ™ํ•˜์ง€ ์•Š์€ ํ™˜๊ฒฝ์„ ํƒ์ƒ‰ํ•˜์—ฌ ์ธ์‹์„ ํ™•์žฅํ•˜๊ณ  ํƒ์ƒ‰๊ณผ ์ฐฉ์ทจ ์‚ฌ์ด์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Voyager [38]์˜ ์—์ด์ „ํŠธ๋Š” ์ž‘์—… ์™„๋ฃŒ ๊ณผ์ •์—์„œ ์•Œ๋ ค์ง€์ง€ ์•Š์€ ์Šคํ‚ฌ์„ ํƒ์ƒ‰ํ•˜๊ณ  ์‹œํ–‰์ฐฉ์˜ค๋ฅผ ํ†ตํ•ด ํ™˜๊ฒฝ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ์Šคํ‚ฌ ์‹คํ–‰ ์ฝ”๋“œ๋ฅผ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Numerous studies have extended the framework of chain-of-thought to various formats to enhance the capacity for planning further. Tree-of-Thought (Yao et al., 2023) enables LLMs to consider multiple reasoning paths in a tree and self-evaluate to determine the next course of action. In cases where global decisions are necessary, ToT allows forward or backward exploration through techniques like deep-first search or breadth-first search. Reasoning via Planning (RAP) (Hao et al., 2023) also divides the problem into a tree and explores them by Monto Carlo tree search algorithm, using LLMs as both world-model and reasoning agent. Another method, Graph of Thought (GoT) (Yao et al., 2023), employs graph nodes to represent individual thoughts and external Graph Neural Networks for organization. LLM+P (Liu et al., 2023) and LLM+DP (Dagan et al., 2023) facilitate the generation of Planning Domain Definition Language (PDDL) (Gerevini, 2020) by LLMs. PDDL assists in decomposing complex problems and utilizing specialized models for planning before converting the results into natural language for LLM processing. However, it is essential to note that these methods use tree/graph/PDDL nodes to represent thoughts, which have limitations regarding their representation forms and can only handle specific planning problems.
์ˆ˜๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ์—ฐ์‡„ ์‚ฌ๊ณ ์˜ ํ‹€์„ ๋‹ค์–‘ํ•œ ํ˜•์‹์œผ๋กœ ํ™•์žฅํ•˜์—ฌ ๋” ๋‚˜์•„๊ฐ„ ๊ณ„ํš ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. Tree-of-Thought (Yao et al., 2023)๋ฅผ ํ†ตํ•ด LLMs์€ ํŠธ๋ฆฌ์—์„œ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์ž์ฒด ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด ๋‹ค์Œ ํ–‰๋™ ๋ฐฉ์นจ์„ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์—ญ์ ์ธ ๊ฒฐ์ •์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ, ToT๋Š” ๊นŠ์ด ์šฐ์„  ๊ฒ€์ƒ‰ ๋˜๋Š” ํญ ์šฐ์„  ๊ฒ€์ƒ‰๊ณผ ๊ฐ™์€ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ „์ง„ ๋˜๋Š” ํ›„์ง„ ํƒ์ƒ‰์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. Reasoning via Planning (RAP) (Hao et al., 2023)๋„ ๋ฌธ์ œ๋ฅผ ํŠธ๋ฆฌ๋กœ ๋‚˜๋ˆ„๊ณ  ๋ชฌํ…Œ์นด๋ฅผ๋กœ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด ํƒ์ƒ‰ํ•˜๋ฉฐ, LLMs์„ ์›”๋“œ ๋ชจ๋ธ๊ณผ ์ถ”๋ก  ์—์ด์ „ํŠธ๋กœ ๋ชจ๋‘ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์ธ Graph of Thought (GoT) (Yao et al., 2023)๋Š” ๊ทธ๋ž˜ํ”„ ๋…ธ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ ์ƒ๊ฐ์„ ํ‘œํ˜„ํ•˜๊ณ  ์™ธ๋ถ€ Graph Neural Networks๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. LLM+P (Liu et al., 2023)์™€ LLM+DP (Dagan et al., 2023)๋Š” LLMs์ด Planning Domain Definition Language (PDDL) (Gerevini, 2020)๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. PDDL์€ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ๋ถ„ํ•ดํ•˜๊ณ  ๊ณ„ํš์„ ์œ„ํ•ด ์ „๋ฌธํ™”๋œ ๋ชจ๋ธ์„ ํ™œ์šฉํ•œ ๋‹ค์Œ ๊ฒฐ๊ณผ๋ฅผ ์ž์—ฐ์–ด๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ LLMs์ด ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ์ƒ๊ฐ์„ ํ‘œํ˜„ํ•˜๊ธฐ ์œ„ํ•ด tree/graph/PDDL ๋…ธ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ํ‘œํ˜„ ํ˜•์‹์— ์ œํ•œ์ด ์žˆ๊ณ  ํŠน์ • ๊ณ„ํš ๋ฌธ์ œ๋งŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์ˆ˜๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ์—ฐ์‡„ ์‚ฌ๊ณ ์˜ ํ‹€์„ ๋‹ค์–‘ํ•œ ํ˜•์‹์œผ๋กœ ํ™•์žฅํ•˜์—ฌ ๋” ๋‚˜์•„๊ฐ„ ๊ณ„ํš ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. Tree-of-Thought(Yao et al., 2023)๋ฅผ ํ†ตํ•ด LLM์€ ํŠธ๋ฆฌ์—์„œ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์ž์ฒด ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด ๋‹ค์Œ ํ–‰๋™ ๋ฐฉ์นจ์„ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์—ญ์ ์ธ ๊ฒฐ์ •์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ, ToT๋Š” ๊นŠ์ด ์šฐ์„  ๊ฒ€์ƒ‰ ๋˜๋Š” ํญ ์šฐ์„  ๊ฒ€์ƒ‰๊ณผ ๊ฐ™์€ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ „์ง„ ๋˜๋Š” ํ›„์ง„ ํƒ์ƒ‰์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ณ„ํš์„ ํ†ตํ•œ ์ถ”๋ก (RAP)(Hao et al., 2023)๋„ ๋ฌธ์ œ๋ฅผ ํŠธ๋ฆฌ๋กœ ๋‚˜๋ˆ„๊ณ  ๋ชฌํ…Œ์นด๋ฅผ๋กœ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด ํƒ์ƒ‰ํ•˜๋ฉฐ, LLM์„ ์›”๋“œ ๋ชจ๋ธ๊ณผ ์ถ”๋ก  ์—์ด์ „ํŠธ๋กœ ๋ชจ๋‘ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์ธ ์ƒ๊ฐ์˜ ๊ทธ๋ž˜ํ”„(GoT)(Yao et al., 2023)๋Š” ๊ทธ๋ž˜ํ”„ ๋…ธ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ ์ƒ๊ฐ์„ ํ‘œํ˜„ํ•˜๊ณ  ์™ธ๋ถ€ ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. LLM+P(Liu et al., 2023)์™€ LLM+DP(Dagan et al., 2023)๋Š” LLM์ด ๊ณ„ํš ๋„๋ฉ”์ธ ์ •์˜ ์–ธ์–ด(PDDL)(Gerevini, 2020)๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. PDDL์€ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ๋ถ„ํ•ดํ•˜๊ณ  ๊ณ„ํš์„ ์œ„ํ•ด ์ „๋ฌธํ™”๋œ ๋ชจ๋ธ์„ ํ™œ์šฉํ•œ ๋‹ค์Œ ๊ฒฐ๊ณผ๋ฅผ ์ž์—ฐ์–ด๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ LLM์ด ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ์ƒ๊ฐ์„ ํ‘œํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ํŠธ๋ฆฌ/๊ทธ๋ž˜ํ”„/PDDL ๋…ธ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ํ‘œํ˜„ ํ˜•์‹์— ์ œํ•œ์ด ์žˆ๊ณ  ํŠน์ • ๊ณ„ํš ๋ฌธ์ œ๋งŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
One of the main challenges of LLMs is the resource gap, as they are mainly pretrained on English corpus and have limited access to data from other languages. English dominates the field of NLP as an extremely high-resource language with the most raw text data from various domains, leaving few of the over 7000 languages of the world represented in the field Joshi et al. (2020). This creates a disparity in language models' capability to handle different languages. Previous findings indicate that LLMs have difficulty comprehending and generating non-English texts, particularly in low-resource languagesNguyen et al. (2023); Zhu et al. (2023); Huang et al. (2023). To address the resource gap, several solutions have been proposed or implemented by researchers and practitioners. One possible solution is to increase the amount of data available from various languages and fields, and make it accessible for pretraining and evaluating LLMs Lin et al. (2022); Chen et al. (2022); Cahyawijaya et al. (2023). However, this approach incurs significant computational expenses and the resource gap persists. Alternatively, multilingual language models trained on texts from different languages concurrently, such as mBERT Devlin et al. (2019) and XLM-R Conneau et al. (2020), have been introduced to bridge the gap effectively.
LLM์˜ ์ฃผ์š” ๊ณผ์ œ ์ค‘ ํ•˜๋‚˜๋Š” ์ฃผ๋กœ ์˜์–ด ๋ง๋ญ‰์น˜๋กœ pretraining๋˜๊ณ  ๋‹ค๋ฅธ ์–ธ์–ด์˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ฑ์ด ์ œํ•œ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์˜์–ด๋Š” ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ์›์‹œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ€์žฅ ๋งŽ์ด ์กด์žฌํ•˜๋Š” ๊ณ ๋ ˆ์†Œ์Šค ์–ธ์–ด๋กœ์„œ NLP ๋ถ„์•ผ๋ฅผ ์ง€๋ฐฐํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ „ ์„ธ๊ณ„ 7000๊ฐœ ์ด์ƒ์˜ ์–ธ์–ด ์ค‘ ์ด ๋ถ„์•ผ์— ๋Œ€ํ‘œ๋˜๋Š” ์–ธ์–ด๋Š” ๊ทนํžˆ ์ผ๋ถ€์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค Joshi et al. (2020). ์ด๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์–ธ์–ด๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์— ์ฐจ์ด๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์€ ํŠนํžˆ ์ €์ž์› ์–ธ์–ด์—์„œ ์˜์–ด๊ฐ€ ์•„๋‹Œ ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹คNguyen et al. (2023); Zhu et al. (2023); Huang et al. (2023). ๋ฆฌ์†Œ์Šค ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋“ค์ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ํ•ด๊ฒฐ์ฑ…์„ ์ œ์•ˆํ•˜๊ฑฐ๋‚˜ ๊ตฌํ˜„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์™€ ๋ถ„์•ผ์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ๋Š˜๋ฆฌ๊ณ  ์ด๋ฅผ LLM pretraining ๋ฐ ํ‰๊ฐ€์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค Lin et al. (2022); Chen et al. (2022); Cahyawijaya et al. (2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ƒ๋‹นํ•œ ๊ณ„์‚ฐ ๋น„์šฉ์ด ์†Œ์š”๋˜๋ฉฐ ๋ฆฌ์†Œ์Šค ๊ฒฉ์ฐจ๋Š” ์ง€์†๋ฉ๋‹ˆ๋‹ค. ๋˜๋Š” mBERT Devlin et al.๊ณผ ๊ฐ™์ด ๋‹ค์–‘ํ•œ ์–ธ์–ด์˜ ํ…์ŠคํŠธ์— ๋Œ€ํ•ด ๋™์‹œ์— ํ•™์Šต๋œ ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ด ์†Œ๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. (2019) ๋ฐ XLM-R Conneau et al. (2020)์„ ํ†ตํ•ด ๊ฒฉ์ฐจ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด์†Œํ–ˆ์Šต๋‹ˆ๋‹ค.
LLM์˜ ์ฃผ์š” ๊ณผ์ œ ์ค‘ ํ•˜๋‚˜๋Š” ์ฃผ๋กœ ์˜์–ด ๋ง๋ญ‰์น˜๋กœ ์‚ฌ์ „ ํ•™์Šต๋˜๊ณ  ๋‹ค๋ฅธ ์–ธ์–ด์˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ฑ์ด ์ œํ•œ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์˜์–ด๋Š” ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ์›์‹œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ€์žฅ ๋งŽ์ด ์กด์žฌํ•˜๋Š” ๊ณ ๋ ˆ์†Œ์Šค ์–ธ์–ด๋กœ์„œ NLP ๋ถ„์•ผ๋ฅผ ์ง€๋ฐฐํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ „ ์„ธ๊ณ„ 7000๊ฐœ ์ด์ƒ์˜ ์–ธ์–ด ์ค‘ ์ด ๋ถ„์•ผ์— ๋Œ€ํ‘œ๋˜๋Š” ์–ธ์–ด๋Š” ๊ทนํžˆ ์ผ๋ถ€์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค Joshi et al. (2020). ์ด๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์–ธ์–ด๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์— ์ฐจ์ด๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์€ ํŠนํžˆ ์ €์ž์› ์–ธ์–ด์—์„œ ์˜์–ด๊ฐ€ ์•„๋‹Œ ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹คNguyen et al. (2023); Zhu et al. (2023); Huang et al. (2023). ๋ฆฌ์†Œ์Šค ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋“ค์ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ํ•ด๊ฒฐ์ฑ…์„ ์ œ์•ˆํ•˜๊ฑฐ๋‚˜ ๊ตฌํ˜„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์™€ ๋ถ„์•ผ์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ๋Š˜๋ฆฌ๊ณ  ์ด๋ฅผ LLM ์‚ฌ์ „ ํ•™์Šต ๋ฐ ํ‰๊ฐ€์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค Lin et al. (2022); Chen et al. (2022); Cahyawijaya et al. (2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ƒ๋‹นํ•œ ๊ณ„์‚ฐ ๋น„์šฉ์ด ์†Œ์š”๋˜๋ฉฐ ๋ฆฌ์†Œ์Šค ๊ฒฉ์ฐจ๋Š” ์ง€์†๋ฉ๋‹ˆ๋‹ค. ๋˜๋Š” mBERT Devlin et al.๊ณผ ๊ฐ™์ด ๋‹ค์–‘ํ•œ ์–ธ์–ด์˜ ํ…์ŠคํŠธ์— ๋Œ€ํ•ด ๋™์‹œ์— ํ•™์Šต๋œ ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ด ์†Œ๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. (2019) ๋ฐ XLM-R Conneau et al. (2020)์„ ํ†ตํ•ด ๊ฒฉ์ฐจ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด์†Œํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Safety and alignment._ Safety and alignment has been widely discussed in LLMs, which aim to regulate the behaviors of models by technical approaches [66]. This topic is also important to MLLMs. Even a highly advanced MLLM (_e.g.,_ GPT-4V [133]) can be susceptible to safety issues. For example, GPT-4V might occasionally exhibit factual inaccuracies and baseless inferences about images. In some cases, it may even generate harmful content targeting specific individuals or groups [133]. Furthermore, open-sourced MLLMs are also prone to generate hallucinated response [844] and can be easily manipulated to produce harmful content [858]. To address the aforementioned issues, some studies collect specialized visual instructions to mitigate the problem of hallucination [834]. Another alternative approach is to train a revision model to rectify hallucinated response generated by MLLMs in a post-hoc way [859]. Additionally, aligning MLLMs with RLHF can also assist MLLMs in generating responses with improved factuality [860]. Despite these efforts, existing alignment techniques for MLLMs mainly concentrate on several specific aspects (_e.g.,_ hallucination), lacking a comprehensive consideration of alignment criteria. More efforts should be made to promote the research of safety and alignment for MLLMs.
\(\bullet\)_์•ˆ์ „ ๋ฐ ์ •๋ ฌ._ ์•ˆ์ „ ๋ฐ ์ •๋ ฌ์€ ๊ธฐ์ˆ ์  ์ ‘๊ทผ ๋ฐฉ์‹[66]์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ํ–‰๋™์„ ๊ทœ์ œํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” LLM์—์„œ ๋„๋ฆฌ ๋…ผ์˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด ์ฃผ์ œ๋Š” MLLM์—๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ๋„๋กœ ๋ฐœ์ „๋œ MLLM(_์˜ˆ:_ GPT-4V [133])๋„ ์•ˆ์ „ ๋ฌธ์ œ์— ์ทจ์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4V๋Š” ๋•Œ๋•Œ๋กœ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์‚ฌ์‹ค์  ๋ถ€์ •ํ™•์„ฑ๊ณผ ๊ทผ๊ฑฐ ์—†๋Š” ์ถ”๋ก ์„ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ํŠน์ • ๊ฐœ์ธ์ด๋‚˜ ์ง‘๋‹จ์„ ๋Œ€์ƒ์œผ๋กœ ํ•˜๋Š” ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค [133]. ๋˜ํ•œ ์˜คํ”ˆ ์†Œ์Šค MLLM์€ hallucinated response๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์‰ฝ๊ณ  [844] ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์‰ฝ๊ฒŒ ์กฐ์ž‘๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [858]. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” hallucination ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ˆ˜ํ•œ ์‹œ๊ฐ์  instructions๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค [834]. ๋˜ ๋‹ค๋ฅธ ๋Œ€์•ˆ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‚ฌํ›„ ๋ฐฉ์‹์œผ๋กœ MLLM์—์„œ ์ƒ์„ฑ๋œ hallucinated response๋ฅผ ์ˆ˜์ •ํ•˜๊ธฐ ์œ„ํ•ด revision ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค [859]. ๋˜ํ•œ RLHF์™€ MLLM์„ ์กฐ์ •ํ•˜๋ฉด ์‚ฌ์‹ค์„ฑ์ด ํ–ฅ์ƒ๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [860]. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  MLLM์— ๋Œ€ํ•œ ๊ธฐ์กด์˜ alignment ๊ธฐ๋ฒ•์€ ์ฃผ๋กœ ๋ช‡ ๊ฐ€์ง€ ํŠน์ • ์ธก๋ฉด(_์˜ˆ:_ hallucination)์— ์ง‘์ค‘๋˜์–ด ์žˆ์œผ๋ฉฐ, alignment ๊ธฐ์ค€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. MLLM์˜ ์•ˆ์ „ ๋ฐ alignment ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์•ˆ์ „ ๋ฐ ์ •๋ ฌ._ ์•ˆ์ „ ๋ฐ ์ •๋ ฌ์€ ๊ธฐ์ˆ ์  ์ ‘๊ทผ ๋ฐฉ์‹[66]์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ํ–‰๋™์„ ๊ทœ์ œํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” LLM์—์„œ ๋„๋ฆฌ ๋…ผ์˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด ์ฃผ์ œ๋Š” MLLM์—๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ๋„๋กœ ๋ฐœ์ „๋œ MLLM(_์˜ˆ:_ GPT-4V [133])๋„ ์•ˆ์ „ ๋ฌธ์ œ์— ์ทจ์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4V๋Š” ๋•Œ๋•Œ๋กœ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์‚ฌ์‹ค์  ๋ถ€์ •ํ™•์„ฑ๊ณผ ๊ทผ๊ฑฐ ์—†๋Š” ์ถ”๋ก ์„ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ํŠน์ • ๊ฐœ์ธ์ด๋‚˜ ์ง‘๋‹จ์„ ๋Œ€์ƒ์œผ๋กœ ํ•˜๋Š” ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค [133]. ๋˜ํ•œ ์˜คํ”ˆ ์†Œ์Šค MLLM์€ ํ™˜๊ฐ์„ฑ ๋ฐ˜์‘์„ ์ƒ์„ฑํ•˜๊ธฐ ์‰ฝ๊ณ  [844] ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์‰ฝ๊ฒŒ ์กฐ์ž‘ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [858]. ์•ž์„œ ์–ธ๊ธ‰ ํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ํ™˜๊ฐ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ˆ˜ํ•œ ์‹œ๊ฐ์  ์ง€์นจ์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค [834]. ๋˜ ๋‹ค๋ฅธ ๋Œ€์•ˆ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‚ฌํ›„ ๋ฐฉ์‹์œผ๋กœ MLLM์—์„œ ์ƒ์„ฑ ๋œ ํ™˜๊ฐ ๋ฐ˜์‘์„ ์ˆ˜์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ˆ˜์ • ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค [859]. ๋˜ํ•œ RLHF์™€ MLLM์„ ์กฐ์ •ํ•˜๋ฉด ์‚ฌ์‹ค์„ฑ์ด ํ–ฅ์ƒ๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [860]. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  MLLM์— ๋Œ€ํ•œ ๊ธฐ์กด์˜ ์ •๋ ฌ ๊ธฐ๋ฒ•์€ ์ฃผ๋กœ ๋ช‡ ๊ฐ€์ง€ ํŠน์ • ์ธก๋ฉด(_์˜ˆ:_ ํ™˜๊ฐ)์— ์ง‘์ค‘๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ •๋ ฌ ๊ธฐ์ค€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. MLLM์˜ ์•ˆ์ „ ๋ฐ ์ •๋ ฌ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
the hyper-parameters settings remain the same as Stanford Alpaca. To better evaluate the instruction following ability of fine-tuned models, we consider two settings, namely _Chat setting_ and _QA setting_. The chat setting mainly utilizes user instructions and queries from daily chat, whereas the QA setting mainly employs question answering examples from existing NLP datasets. The evaluation on the chat setting is conducted based on the AlpacaFarm evaluation set [363]. Instead of using a full pairwise comparison, we select the LLaMA 7B and 13B models fine-tuned on Self-Instruct-52K as the reference baselines, and then compare them with other fine-tuned LLaMA 7B and 13B models using different instructions, respectively. Since our focus is to examine the usefulness of different strategies to generate the instructions, the model fine-tuned on Self-Instruct-52K can serve as a good reference. Following AlpacaFarm [363], for each comparison, we employ ChatGPT to automatically annotate which response from two compared models each time is the best for the user query, and report the win rate (%) as the evaluation metric. For the QA setting, we select two benchmarks, MMLU [364] and BBH [365], and evaluate the accuracy based on their default settings by using heuristic rules to parse the answers from these LLMs.
ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •์€ Stanford Alpaca์™€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. Fine-tuned ๋ชจ๋ธ์˜ Instruction ์ถ”์ข… ๋Šฅ๋ ฅ์„ ๋” ์ž˜ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด _Chat setting_๊ณผ _QA setting_์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์„ค์ •์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. Chat setting์€ ์ฃผ๋กœ ์ผ์ƒ์ ์ธ ์ฑ„ํŒ…์—์„œ ์‚ฌ์šฉ์ž ๋ช…๋ น๊ณผ ์ฟผ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” ๋ฐ˜๋ฉด, QA setting์€ ์ฃผ๋กœ ๊ธฐ์กด NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ question answering ์˜ˆ์ œ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. Chat setting์— ๋Œ€ํ•œ ํ‰๊ฐ€๋Š” AlpacaFarm ํ‰๊ฐ€ ์„ธํŠธ [363]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์Œ๋ณ„ ๋น„๊ต๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  Self-Instruct-52K์—์„œ Fine-tuned๋œ LLaMA 7B ๋ฐ 13B ๋ชจ๋ธ์„ ๊ธฐ์ค€ ๊ธฐ์ค€์œผ๋กœ ์„ ํƒํ•œ ๋‹ค์Œ, ๋‹ค๋ฅธ Instruction์„ ์‚ฌ์šฉํ•˜์—ฌ Fine-tuned๋œ ๋‹ค๋ฅธ LLaMA 7B ๋ฐ 13B ๋ชจ๋ธ๊ณผ ๊ฐ๊ฐ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์˜ ์ดˆ์ ์€ Instruction์„ ์ƒ์„ฑํ•˜๋Š” ๋‹ค์–‘ํ•œ ์ „๋žต์˜ ์œ ์šฉ์„ฑ์„ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์ด๋ฏ€๋กœ, Self-Instruct-52K์—์„œ Fine-tuned๋œ ๋ชจ๋ธ์ด ์ข‹์€ ๊ธฐ์ค€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AlpacaFarm [363]์— ๋”ฐ๋ผ ๊ฐ ๋น„๊ต์— ๋Œ€ํ•ด ChatGPT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋น„๊ต ๋Œ€์ƒ ๋ชจ๋ธ์—์„œ ๊ฐ๊ฐ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์‘๋‹ต์ด ๋ฌด์—‡์ธ์ง€ ์ž๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ๊ณ , ๊ทธ ๋น„์œจ(%)์„ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. QA setting์˜ ๊ฒฝ์šฐ MMLU [364]์™€ BBH [365]๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์„ ํƒํ•˜๊ณ , ์ด๋Ÿฌํ•œ LLM์—์„œ ๋‹ต๋ณ€์„ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•ด heuristic rules๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ๋ณธ ์„ค์ •์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •์€ ์Šคํƒ ํฌ๋“œ ์•ŒํŒŒ์นด์™€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์˜ ๋ช…๋ น์–ด ์ถ”์ข… ๋Šฅ๋ ฅ์„ ๋” ์ž˜ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด _์ฑ„ํŒ… ์„ค์ •_๊ณผ _QA ์„ค์ •_์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์„ค์ •์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฑ„ํŒ… ์„ค์ •์€ ์ฃผ๋กœ ์ผ์ƒ์ ์ธ ์ฑ„ํŒ…์—์„œ ์‚ฌ์šฉ์ž ๋ช…๋ น๊ณผ ์ฟผ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” ๋ฐ˜๋ฉด, QA ์„ค์ •์€ ์ฃผ๋กœ ๊ธฐ์กด NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์˜ˆ์ œ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฑ„ํŒ… ์„ค์ •์— ๋Œ€ํ•œ ํ‰๊ฐ€๋Š” ์•ŒํŒŒ์นดํŒœ ํ‰๊ฐ€ ์„ธํŠธ [363]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์Œ๋ณ„ ๋น„๊ต๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  Self-Instruct-52K์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ LLaMA 7B ๋ฐ 13B ๋ชจ๋ธ์„ ๊ธฐ์ค€ ๊ธฐ์ค€์œผ๋กœ ์„ ํƒํ•œ ๋‹ค์Œ, ๋‹ค๋ฅธ ๋ช…๋ น์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋‹ค๋ฅธ LLaMA 7B ๋ฐ 13B ๋ชจ๋ธ๊ณผ ๊ฐ๊ฐ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์˜ ์ดˆ์ ์€ ๋ช…๋ น์„ ์ƒ์„ฑํ•˜๋Š” ๋‹ค์–‘ํ•œ ์ „๋žต์˜ ์œ ์šฉ์„ฑ์„ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์ด๋ฏ€๋กœ, Self-Instruct-52K์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์ด ์ข‹์€ ๊ธฐ์ค€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ŒํŒŒ์นดํŒœ [363]์— ๋”ฐ๋ผ ๊ฐ ๋น„๊ต์— ๋Œ€ํ•ด ChatGPT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋น„๊ต ๋Œ€์ƒ ๋ชจ๋ธ์—์„œ ๊ฐ๊ฐ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์‘๋‹ต์ด ๋ฌด์—‡์ธ์ง€ ์ž๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ๊ณ , ๊ทธ ๋น„์œจ(%)์„ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. QA ์„ค์ •์˜ ๊ฒฝ์šฐ MMLU [364]์™€ BBH [365]๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์„ ํƒํ•˜๊ณ , ์ด๋Ÿฌํ•œ LLM์—์„œ ๋‹ต๋ณ€์„ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•ด ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ๋ณธ ์„ค์ •์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
In a pioneering study, Li et al. (2022) proposed the use of retrieval techniques to enhance the quality of text-to-image generation. They conducted a comparative analysis of the quality and quantity of the generated images with mainstream models on the CUB and COCO datasets. Their findings demonstrated that all models outperformed their contemporaries. In contrast, RE-IMAGEN Chen et al. (2022) focused on assisting the model in generating images of uncommon objects through retrieval. This approach ultimately led to the achievement of exceptionally high FID scores on the COCO and WikilImage datasets. Even more groundbreaking results were obtained on the authors' own proposed EntityDrawBench benchmark, which encompasses a range of common and rare objects across multiple categories. RDM Blattmann et al. (2022), although trained in a similar manner to RE-IMAGEN, employs image features as the foundation for retrieval and is supplanted by user examples during the inference process. Consequently, RDM is capable of efficiently transferring the described artistic style to the generated images. Furthermore, in contrast to RE-IMAGEN, which employs image-text pairs for retrieval, KNN-Diffusion Sheynin et al. (2022) solely utilizes images for retrieval, resulting in a lower quality of results on the COCO dataset.
์„ ๊ตฌ์ ์ธ ์—ฐ๊ตฌ์—์„œ Li ๋“ฑ(2022)์€ ํ…์ŠคํŠธ์—์„œ ์ด๋ฏธ์ง€๋กœ ์ƒ์„ฑํ•˜๋Š” ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด retrieval ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ CUB ๋ฐ COCO ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ฃผ๋ฅ˜ ๋ชจ๋ธ๊ณผ ํ•จ๊ป˜ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์˜ ํ’ˆ์งˆ๊ณผ ์–‘์„ ๋น„๊ต ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋ชจ๋“  ๋ชจ๋ธ์ด ๋™๊ธ‰ ๋ชจ๋ธ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค๋Š” ์‚ฌ์‹ค์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, RE-IMAGEN Chen ๋“ฑ(2022)์€ retrieval์„ ํ†ตํ•ด ๋ชจ๋ธ์ด ํ”ํ•˜์ง€ ์•Š์€ ๋ฌผ์ฒด์˜ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ถ๊ทน์ ์œผ๋กœ COCO ๋ฐ WikilImage ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋งค์šฐ ๋†’์€ FID ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์นดํ…Œ๊ณ ๋ฆฌ์— ๊ฑธ์ณ ๋‹ค์–‘ํ•œ ์ผ๋ฐ˜ ๋ฐ ํฌ๊ท€ ๋ฌผ์ฒด๋ฅผ ํฌํ•จํ•˜๋Š” ์ €์ž๋“ค์ด ์ œ์•ˆํ•œ EntityDrawBench ๋ฒค์น˜๋งˆํฌ์—์„œ ๋”์šฑ ํš๊ธฐ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. RDM Blattmann ๋“ฑ(2022)์€ RE-IMAGEN๊ณผ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์ง€๋งŒ ์ด๋ฏธ์ง€ ๊ธฐ๋Šฅ์„ retrieval์˜ ๊ธฐ๋ฐ˜์œผ๋กœ ์‚ฌ์šฉํ•˜๋ฉฐ ์ถ”๋ก  ๊ณผ์ •์—์„œ ์‚ฌ์šฉ์ž ์˜ˆ์ œ๋กœ ๋Œ€์ฒด๋ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ RDM์€ ์„ค๋ช…๋œ ์˜ˆ์ˆ ์  ์Šคํƒ€์ผ์„ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์— ํšจ์œจ์ ์œผ๋กœ ์ „์†กํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์„ retrieval์— ์‚ฌ์šฉํ•˜๋Š” RE-IMAGEN๊ณผ ๋‹ฌ๋ฆฌ KNN-Diffusion Sheynin ๋“ฑ(2022)์€ retrieval์— ์ด๋ฏธ์ง€๋งŒ์„ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ COCO ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฒฐ๊ณผ ํ’ˆ์งˆ์ด ๋‚ฎ์•„์ง‘๋‹ˆ๋‹ค.
์„ ๊ตฌ์ ์ธ ์—ฐ๊ตฌ์—์„œ Li ๋“ฑ(2022)์€ ํ…์ŠคํŠธ์—์„œ ์ด๋ฏธ์ง€๋กœ ์ƒ์„ฑํ•˜๋Š” ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฒ€์ƒ‰ ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ CUB ๋ฐ COCO ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ฃผ๋ฅ˜ ๋ชจ๋ธ๊ณผ ํ•จ๊ป˜ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์˜ ํ’ˆ์งˆ๊ณผ ์–‘์„ ๋น„๊ต ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋ชจ๋“  ๋ชจ๋ธ์ด ๋™๊ธ‰ ๋ชจ๋ธ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค๋Š” ์‚ฌ์‹ค์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, RE-IMAGEN Chen ๋“ฑ(2022)์€ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ๋ชจ๋ธ์ด ํ”ํ•˜์ง€ ์•Š์€ ๋ฌผ์ฒด์˜ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ถ๊ทน์ ์œผ๋กœ COCO ๋ฐ WikilImage ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋งค์šฐ ๋†’์€ FID ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์นดํ…Œ๊ณ ๋ฆฌ์— ๊ฑธ์ณ ๋‹ค์–‘ํ•œ ์ผ๋ฐ˜ ๋ฐ ํฌ๊ท€ ๋ฌผ์ฒด๋ฅผ ํฌํ•จํ•˜๋Š” ์ €์ž๋“ค์ด ์ œ์•ˆํ•œ EntityDrawBench ๋ฒค์น˜๋งˆํฌ์—์„œ ๋”์šฑ ํš๊ธฐ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. RDM Blattmann ๋“ฑ(2022)์€ RE-IMAGEN๊ณผ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์ง€๋งŒ ์ด๋ฏธ์ง€ ๊ธฐ๋Šฅ์„ ๊ฒ€์ƒ‰์˜ ๊ธฐ๋ฐ˜์œผ๋กœ ์‚ฌ์šฉํ•˜๋ฉฐ ์ถ”๋ก  ๊ณผ์ •์—์„œ ์‚ฌ์šฉ์ž ์˜ˆ์ œ๋กœ ๋Œ€์ฒด๋ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ RDM์€ ์„ค๋ช…๋œ ์˜ˆ์ˆ ์  ์Šคํƒ€์ผ์„ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์— ํšจ์œจ์ ์œผ๋กœ ์ „์†กํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์„ ๊ฒ€์ƒ‰์— ์‚ฌ์šฉํ•˜๋Š” RE-IMAGEN๊ณผ ๋‹ฌ๋ฆฌ KNN-Diffusion Sheynin ๋“ฑ(2022)์€ ๊ฒ€์ƒ‰์— ์ด๋ฏธ์ง€๋งŒ์„ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ COCO ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฒฐ๊ณผ ํ’ˆ์งˆ์ด ๋‚ฎ์•„์ง‘๋‹ˆ๋‹ค.
1. Publishers do not train directly on or against the benchmark dataset and retract any reported results if and when benchmark data is found to have been in training data. 2. Techniques that are likely to increase the test performance without a commensurate increase in safety factor are discouraged and may result in benchmark exclusion. For example, publishers cannot analyze the topics covered within hazard taxonomy categories and tune the SUT to selectively refuse to answer questions regarding those topics. 3. Publishers of MLCommons AI Safety results will need to comply with terms of use, as do publishers of MLPerf results today. 4. Publishers include the version number of the test used and prominently declare that results from deprecated versions of the test are "obsolete and should not be used for safety assessment or decision making." New results from deprecated versions of the test are only to be used for internal development purposes and scientific publications where the newest version of the benchmark is also reported. 5. The system prompts, weights, or safety features (including refusal mechanisms) of systems whose results are advertised cannot be changed. Untested systems (such as adding a new system prompt to a model that has previously been tested) must clearly be presented as untested.
1. ๊ฒŒ์‹œ์ž๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ง์ ‘ ํ›ˆ๋ จํ•˜๊ฑฐ๋‚˜ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋œ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์งˆ ๊ฒฝ์šฐ ๋ณด๊ณ ๋œ ๊ฒฐ๊ณผ๋ฅผ ์ฒ ํšŒํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 2. ์•ˆ์ „ ๊ณ„์ˆ˜์˜ ์ƒ์‘ํ•˜๋Š” ์ฆ๊ฐ€ ์—†์ด ํ…Œ์ŠคํŠธ ์„ฑ๋Šฅ์„ ๋†’์ผ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š” ๊ธฐ์ˆ ์€ ๊ถŒ์žฅ๋˜์ง€ ์•Š์œผ๋ฉฐ, ๋ฒค์น˜๋งˆํฌ ์ œ์™ธ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ฒŒ์‹œ์ž๋Š” ์œ„ํ—˜ ๋ถ„๋ฅ˜ ์นดํ…Œ๊ณ ๋ฆฌ ๋‚ด์—์„œ ๋‹ค๋ฃจ๋Š” ์ฃผ์ œ๋ฅผ ๋ถ„์„ํ•˜๊ณ  ํ•ด๋‹น ์ฃผ์ œ์— ๊ด€ํ•œ ์งˆ๋ฌธ์— ์„ ํƒ์ ์œผ๋กœ ๋‹ต๋ณ€์„ ๊ฑฐ๋ถ€ํ•˜๋„๋ก SUT๋ฅผ ์กฐ์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. 3. MLCommons AI ์•ˆ์ „ ๊ฒฐ๊ณผ์˜ ๊ฒŒ์‹œ์ž๋Š” ํ˜„์žฌ MLPerf ๊ฒฐ๊ณผ์˜ ๊ฒŒ์‹œ์ž์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์‚ฌ์šฉ ์•ฝ๊ด€์„ ์ค€์ˆ˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4. ๊ฒŒ์‹œ์ž๋Š” ์‚ฌ์šฉ๋œ ํ…Œ์ŠคํŠธ์˜ ๋ฒ„์ „ ๋ฒˆํ˜ธ๋ฅผ ํฌํ•จํ•˜๊ณ  ๋” ์ด์ƒ ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ๋ฒ„์ „์˜ ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ๋Š” "๊ตฌ์‹์ด๋ฉฐ ์•ˆ์ „ ํ‰๊ฐ€ ๋˜๋Š” ์˜์‚ฌ ๊ฒฐ์ •์— ์‚ฌ์šฉํ•ด์„œ๋Š” ์•ˆ ๋œ๋‹ค"๊ณ  ๋ˆˆ์— ์ž˜ ๋„๊ฒŒ ํ‘œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ด์ƒ ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ๋ฒ„์ „์˜ ์ƒˆ ๊ฒฐ๊ณผ๋Š” ๋‚ด๋ถ€ ๊ฐœ๋ฐœ ๋ชฉ์ ๊ณผ ์ตœ์‹  ๋ฒ„์ „์˜ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋ณด๊ณ ๋œ ๊ณผํ•™ ๊ฐ„ํ–‰๋ฌผ์—๋งŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 5. ๊ฒฐ๊ณผ๋ฅผ ๊ด‘๊ณ ํ•˜๋Š” ์‹œ์Šคํ…œ์˜ prompts, weights ๋˜๋Š” ์•ˆ์ „ ๊ธฐ๋Šฅ(๊ฑฐ๋ถ€ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ํฌํ•จ)์€ ๋ณ€๊ฒฝํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ๋˜์ง€ ์•Š์€ ์‹œ์Šคํ…œ(์ด์ „์— ํ…Œ์ŠคํŠธ๋œ ๋ชจ๋ธ์— ์ƒˆ ์‹œ์Šคํ…œ prompt๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๋“ฑ)์€ ํ…Œ์ŠคํŠธ๋˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
1. ๊ฒŒ์‹œ์ž๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ง์ ‘ ํ›ˆ๋ จํ•˜๊ฑฐ๋‚˜ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋œ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์งˆ ๊ฒฝ์šฐ ๋ณด๊ณ ๋œ ๊ฒฐ๊ณผ๋ฅผ ์ฒ ํšŒํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 2. ์•ˆ์ „ ๊ณ„์ˆ˜์˜ ์ƒ์‘ํ•˜๋Š” ์ฆ๊ฐ€ ์—†์ด ํ…Œ์ŠคํŠธ ์„ฑ๋Šฅ์„ ๋†’์ผ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š” ๊ธฐ์ˆ ์€ ๊ถŒ์žฅ๋˜์ง€ ์•Š์œผ๋ฉฐ, ๋ฒค์น˜๋งˆํฌ ์ œ์™ธ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ฒŒ์‹œ์ž๋Š” ์œ„ํ—˜ ๋ถ„๋ฅ˜ ์นดํ…Œ๊ณ ๋ฆฌ ๋‚ด์—์„œ ๋‹ค๋ฃจ๋Š” ์ฃผ์ œ๋ฅผ ๋ถ„์„ํ•˜๊ณ  ํ•ด๋‹น ์ฃผ์ œ์— ๊ด€ํ•œ ์งˆ๋ฌธ์— ์„ ํƒ์ ์œผ๋กœ ๋‹ต๋ณ€์„ ๊ฑฐ๋ถ€ํ•˜๋„๋ก SUT๋ฅผ ์กฐ์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. 3. MLCommons AI ์•ˆ์ „ ๊ฒฐ๊ณผ์˜ ๊ฒŒ์‹œ์ž๋Š” ํ˜„์žฌ MLPerf ๊ฒฐ๊ณผ์˜ ๊ฒŒ์‹œ์ž์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์‚ฌ์šฉ ์•ฝ๊ด€์„ ์ค€์ˆ˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4. ๊ฒŒ์‹œ์ž๋Š” ์‚ฌ์šฉ๋œ ํ…Œ์ŠคํŠธ์˜ ๋ฒ„์ „ ๋ฒˆํ˜ธ๋ฅผ ํฌํ•จํ•˜๊ณ  ๋” ์ด์ƒ ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ๋ฒ„์ „์˜ ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ๋Š” "๊ตฌ์‹์ด๋ฉฐ ์•ˆ์ „ ํ‰๊ฐ€ ๋˜๋Š” ์˜์‚ฌ ๊ฒฐ์ •์— ์‚ฌ์šฉํ•ด์„œ๋Š” ์•ˆ ๋œ๋‹ค"๊ณ  ๋ˆˆ์— ์ž˜ ๋„๊ฒŒ ํ‘œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ด์ƒ ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ๋ฒ„์ „์˜ ์ƒˆ ๊ฒฐ๊ณผ๋Š” ๋‚ด๋ถ€ ๊ฐœ๋ฐœ ๋ชฉ์ ๊ณผ ์ตœ์‹  ๋ฒ„์ „์˜ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋ณด๊ณ ๋œ ๊ณผํ•™ ๊ฐ„ํ–‰๋ฌผ์—๋งŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 5. ๊ฒฐ๊ณผ๋ฅผ ๊ด‘๊ณ ํ•˜๋Š” ์‹œ์Šคํ…œ์˜ ํ”„๋กฌํ”„ํŠธ, ๊ฐ€์ค‘์น˜ ๋˜๋Š” ์•ˆ์ „ ๊ธฐ๋Šฅ(๊ฑฐ๋ถ€ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ํฌํ•จ)์€ ๋ณ€๊ฒฝํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ๋˜์ง€ ์•Š์€ ์‹œ์Šคํ…œ(์ด์ „์— ํ…Œ์ŠคํŠธ๋œ ๋ชจ๋ธ์— ์ƒˆ ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๋“ฑ)์€ ํ…Œ์ŠคํŠธ๋˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
* **HotpotQA**[198] is designed to cover a diverse and explainable question-answering dataset that necessitates multi-hop reasoning. This dataset is derived from the English Wikipedia. It consists of roughly \(113,000\) questions. Each question in the dataset comes with two paragraphs, called gold paragraphs, from two Wikipedia articles. Also, there is a list of sentences in those paragraphs that crowdworkers have picked as important for answering the question. * **ToolQA**[199] is a question answering benchmark to evaluate LLMs' ability to use external tools for answering questions. * **GPT4Tools** serves as an instructional dataset, generated by instructing advanced teachers (such as Chat-GPT), with instructions conditioned on visual content and tool descriptions. This process results in the generation of instructions related to the use of tools. There are three versions of this dataset. The first version comprises 71,000 instruction-following data points utilized to fine-tune the GPT4Tools model. The next version consists of manually cleaned instruction data used for validation, covering instructions related to the tools from the first version. The last version is cleaned instruction data used for testing and includes instructions related to some tools that are not present in the first version.
* **HotpotQA**[198]๋Š” ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์ด ํ•„์š”ํ•œ ๋‹ค์–‘ํ•˜๊ณ  ์„ค๋ช… ๊ฐ€๋Šฅํ•œ question-answering ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋‹ค๋ฃจ๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์˜์–ด ์œ„ํ‚คํ”ผ๋””์•„์—์„œ ํŒŒ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์•ฝ 113,000๊ฐœ์˜ ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ์งˆ๋ฌธ์—๋Š” ๋‘ ๊ฐœ์˜ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฌธ์„œ์—์„œ ๊ณจ๋“œ ๋‹จ๋ฝ์ด๋ผ๊ณ  ํ•˜๋Š” ๋‘ ๊ฐœ์˜ ๋‹จ๋ฝ์ด ํ•จ๊ป˜ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํฌ๋ผ์šฐ๋“œ์›Œ์ปค๊ฐ€ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•˜๋‹ค๊ณ  ์„ ํƒํ•œ ํ•ด๋‹น ๋‹จ๋ฝ์˜ ๋ฌธ์žฅ ๋ชฉ๋ก๋„ ์žˆ์Šต๋‹ˆ๋‹ค. * **ToolQA**[199]๋Š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์™ธ๋ถ€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ question-answering ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. * **GPT4Tools**๋Š” ์‹œ๊ฐ์  ์ฝ˜ํ…์ธ ์™€ ๋„๊ตฌ ์„ค๋ช…์— ๋”ฐ๋ผ ์กฐ๊ฑด์ด ์ง€์ •๋œ ๊ณ ๊ธ‰ ๊ต์‚ฌ(์˜ˆ: Chat-GPT)์—๊ฒŒ ์ง€์นจ์„ ์ œ๊ณตํ•˜์—ฌ ์ƒ์„ฑ๋œ instructional ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ํ†ตํ•ด ๋„๊ตฌ ์‚ฌ์šฉ๊ณผ ๊ด€๋ จ๋œ ์ง€์นจ์ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์„ธ ๊ฐ€์ง€ ๋ฒ„์ „์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์€ GPT4Tools ๋ชจ๋ธ์„ Fine-tuningํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” 71,000๊ฐœ์˜ instruction-following ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ฒ„์ „์€ ์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์˜ ๋„๊ตฌ์™€ ๊ด€๋ จ๋œ ์ง€์นจ์„ ํฌํ•จํ•˜๋Š” ๊ฒ€์ฆ์— ์‚ฌ์šฉ๋˜๋Š” ์ˆ˜๋™์œผ๋กœ ์ •๋ฆฌ๋œ instruction ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋ฒ„์ „์€ ํ…Œ์ŠคํŠธ์— ์‚ฌ์šฉ๋˜๋Š” ์ •๋ฆฌ๋œ instruction ๋ฐ์ดํ„ฐ์ด๋ฉฐ, ์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์— ์—†๋Š” ์ผ๋ถ€ ๋„๊ตฌ์™€ ๊ด€๋ จ๋œ ์ง€์นจ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
* **HotpotQA**[198]๋Š” ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์ด ํ•„์š”ํ•œ ๋‹ค์–‘ํ•˜๊ณ  ์„ค๋ช… ๊ฐ€๋Šฅํ•œ ์งˆ๋ฌธ-๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋‹ค๋ฃจ๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์˜์–ด ์œ„ํ‚คํ”ผ๋””์•„์—์„œ ํŒŒ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์•ฝ 113,000๊ฐœ์˜ ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ์งˆ๋ฌธ์—๋Š” ๋‘ ๊ฐœ์˜ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฌธ์„œ์—์„œ ๊ณจ๋“œ ๋‹จ๋ฝ์ด๋ผ๊ณ  ํ•˜๋Š” ๋‘ ๊ฐœ์˜ ๋‹จ๋ฝ์ด ํ•จ๊ป˜ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํฌ๋ผ์šฐ๋“œ์›Œ์ปค๊ฐ€ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•˜๋‹ค๊ณ  ์„ ํƒํ•œ ํ•ด๋‹น ๋‹จ๋ฝ์˜ ๋ฌธ์žฅ ๋ชฉ๋ก๋„ ์žˆ์Šต๋‹ˆ๋‹ค. * **ToolQA**[199]๋Š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์™ธ๋ถ€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. * **GPT4Tools**๋Š” ์‹œ๊ฐ์  ์ฝ˜ํ…์ธ ์™€ ๋„๊ตฌ ์„ค๋ช…์— ๋”ฐ๋ผ ์กฐ๊ฑด์ด ์ง€์ •๋œ ๊ณ ๊ธ‰ ๊ต์‚ฌ(์˜ˆ: Chat-GPT)์—๊ฒŒ ์ง€์นจ์„ ์ œ๊ณตํ•˜์—ฌ ์ƒ์„ฑ๋œ ๊ต์œก์šฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ํ†ตํ•ด ๋„๊ตฌ ์‚ฌ์šฉ๊ณผ ๊ด€๋ จ๋œ ์ง€์นจ์ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์„ธ ๊ฐ€์ง€ ๋ฒ„์ „์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์€ GPT4Tools ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” 71,000๊ฐœ์˜ ์ง€์นจ ์ˆ˜ํ–‰ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ฒ„์ „์€ ์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์˜ ๋„๊ตฌ์™€ ๊ด€๋ จ๋œ ์ง€์นจ์„ ํฌํ•จํ•˜๋Š” ๊ฒ€์ฆ์— ์‚ฌ์šฉ๋˜๋Š” ์ˆ˜๋™์œผ๋กœ ์ •๋ฆฌ๋œ ์ง€์นจ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋ฒ„์ „์€ ํ…Œ์ŠคํŠธ์— ์‚ฌ์šฉ๋˜๋Š” ์ •๋ฆฌ๋œ ์ง€์นจ ๋ฐ์ดํ„ฐ์ด๋ฉฐ, ์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์— ์—†๋Š” ์ผ๋ถ€ ๋„๊ตฌ์™€ ๊ด€๋ จ๋œ ์ง€์นจ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
Motion Mamba [88], composed of Hierarchical Temporal Mambo (HTM) and Bidirectional Spatial Mamba (BSM), represents the first integration of Mamba models in the field of motion generation. HTM and BSM are designed for temporal and spatial modeling, respectively, while integrating selective scanning mechanisms into motion generation tasks. Compared with the previous diffusion-based motion generation method, which mainly uses a Transformer, motion Mamba achieves state-of-the-art (SOTA) performance. Note that the inference speed is also sped up by four times. VL-Mambo [145] is the first effort to explore the state-space model Mamba to solve the expensive computational overhead in the Transformer architecture in multimodal learning tasks. CMViM [146] focuses on the application of multimodal representation learning to 3D high-resolution medical images, especially Alzheimer's disease (AD). It is developed based on MAE [150] framework and replaces ViT [151] module with Vim [61] module, therefore, reducing complexity from quadratic to linear level. Moreover, the intra-modality and inter-modality contrastive learning methods are introduced to enhance the ability of the multimodal Vim encoder to model discriminative features in the same modality and mitigate the misaligned representation between modalities.
Motion Mamba [88]๋Š” ๊ณ„์ธต์  ์‹œ๊ฐ„์  Mambo (HTM)์™€ ์–‘๋ฐฉํ–ฅ ๊ณต๊ฐ„ Mamba (BSM)๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋™์ž‘ ์ƒ์„ฑ ๋ถ„์•ผ์—์„œ Mamba ๋ชจ๋ธ์„ ์ตœ์ดˆ๋กœ ํ†ตํ•ฉํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. HTM๊ณผ BSM์€ ๊ฐ๊ฐ ์‹œ๊ฐ„์  ๋ฐ ๊ณต๊ฐ„์  ๋ชจ๋ธ๋ง์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, ๋™์ž‘ ์ƒ์„ฑ ์ž‘์—…์— ์„ ํƒ์  ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. Transformer๋ฅผ ์ฃผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ธฐ์กด์˜ ํ™•์‚ฐ ๊ธฐ๋ฐ˜ ๋™์ž‘ ์ƒ์„ฑ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ•˜์—ฌ Motion Mamba๋Š” state-of-the-art (SOTA) ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก  ์†๋„๋„ 4๋ฐฐ ๊ฐ€์†ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. VL-Mambo [145]๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ•™์Šต ์ž‘์—…์—์„œ Transformer ์•„ํ‚คํ…์ฒ˜์˜ ๊ฐ’๋น„์‹ผ ๊ณ„์‚ฐ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ Mamba๋ฅผ ํƒ๊ตฌํ•œ ์ตœ์ดˆ์˜ ๋…ธ๋ ฅ์ž…๋‹ˆ๋‹ค. CMViM [146]์€ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ‘œํ˜„ ํ•™์Šต์„ 3D ๊ณ ํ•ด์ƒ๋„ ์˜๋ฃŒ ์ด๋ฏธ์ง€, ํŠนํžˆ ์•Œ์ธ ํ•˜์ด๋จธ๋ณ‘(AD)์— ์ ์šฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ MAE [150] ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ViT [151] ๋ชจ๋“ˆ์„ Vim [61] ๋ชจ๋“ˆ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ๋ณต์žก์„ฑ์„ ์ด์ฐจ ์ˆ˜์ค€์—์„œ ์„ ํ˜• ์ˆ˜์ค€์œผ๋กœ ์ค„์˜€์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋™์ผํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์—์„œ ๋ณ€๋ณ„์  ํŠน์ง•์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ Vim Encoder์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๊ฐ„์˜ ์ž˜๋ชป ์ •๋ ฌ๋œ ํ‘œํ˜„์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๋‚ด ๋ฐ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๊ฐ„ ๋Œ€์กฐ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
Motion Mamba [88]๋Š” ๊ณ„์ธต์  ์‹œ๊ฐ„์  ๋ง˜๋ฐ”(HTM)์™€ ์–‘๋ฐฉํ–ฅ ๊ณต๊ฐ„ ๋ง˜๋ฐ”(BSM)๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋™์ž‘ ์ƒ์„ฑ ๋ถ„์•ผ์—์„œ ๋ง˜๋ฐ” ๋ชจ๋ธ์„ ์ตœ์ดˆ๋กœ ํ†ตํ•ฉํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. HTM๊ณผ BSM์€ ๊ฐ๊ฐ ์‹œ๊ฐ„์  ๋ฐ ๊ณต๊ฐ„์  ๋ชจ๋ธ๋ง์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, ๋™์ž‘ ์ƒ์„ฑ ์ž‘์—…์— ์„ ํƒ์  ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋žœ์Šคํฌ๋จธ๋ฅผ ์ฃผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ธฐ์กด์˜ ํ™•์‚ฐ ๊ธฐ๋ฐ˜ ๋™์ž‘ ์ƒ์„ฑ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ•˜์—ฌ ๋™์ž‘ ๋ง˜๋ฐ”๋Š” ์ตœ์ฒจ๋‹จ(SOTA) ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก  ์†๋„๋„ 4๋ฐฐ ๊ฐ€์†ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. VL-Mambo [145]๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ•™์Šต ์ž‘์—…์—์„œ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜์˜ ๊ฐ’๋น„์‹ผ ๊ณ„์‚ฐ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ ๋ง˜๋ฐ”๋ฅผ ํƒ๊ตฌํ•œ ์ตœ์ดˆ์˜ ๋…ธ๋ ฅ์ž…๋‹ˆ๋‹ค. CMViM [146]์€ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ‘œํ˜„ ํ•™์Šต์„ 3D ๊ณ ํ•ด์ƒ๋„ ์˜๋ฃŒ ์ด๋ฏธ์ง€, ํŠนํžˆ ์•Œ์ธ ํ•˜์ด๋จธ๋ณ‘(AD)์— ์ ์šฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ MAE [150] ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ViT [151] ๋ชจ๋“ˆ์„ Vim [61] ๋ชจ๋“ˆ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ๋ณต์žก์„ฑ์„ ์ด์ฐจ ์ˆ˜์ค€์—์„œ ์„ ํ˜• ์ˆ˜์ค€์œผ๋กœ ์ค„์˜€์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋™์ผํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์—์„œ ๋ณ€๋ณ„์  ํŠน์ง•์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋น” ์ธ์ฝ”๋”์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๊ฐ„์˜ ์ž˜๋ชป ์ •๋ ฌ๋œ ํ‘œํ˜„์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๋‚ด ๋ฐ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๊ฐ„ ๋Œ€์กฐ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
1. **Parallel rationale generation (think, Subsection 4.2)**: In parallel across \(n\) tokens \(x_{i}\) in an input sequence \(x_{0:n}\), we generate \(r\) rationales of length \(t\): \(c_{i}=(c_{i1},\dots,c_{it})\), resulting in \(n\times r\) rationale candidates. We insert learned \(<\)\(\mid\)start\(\mid\)\(>\) and \(<\)\(\mid\)end\(\mid\)\(>\) tokens to mark each rationale's start and end. 2. **Mixing post-rationale and base predictions (talk, Subsection 4.3)**: From the hidden state output after each rationale, we train a "mixing head" - a shallow MLP producing a weight determining how much the post-rationale next-token predicted logits should be incorporated compared to the base language model predicted logits. This approach eases distribution shift early in finetuning, due to introducing rationales. 3. **Optimizing rationale generation (learn, Subsection 4.4)**: We optimize the rationale generation parameters (start/end tokens and LM weights) to increase the likelihood of rationales that make future text more probable. We use REINFORCE to provide a learning signal to rationales based on their impact on future-token prediction. To reduce variance, we apply a teacher-forcing trick to include in the loss the likelihood of predicting not only the token after the thought but also later tokens.
1. **๋ณ‘๋ ฌ ๊ทผ๊ฑฐ ์ƒ์„ฑ(4.2์ ˆ ์ฐธ์กฐ)**: ์ž…๋ ฅ ์‹œํ€€์Šค \(x_{0:n}\)์˜ \(n\) ํ† ํฐ \(x_{i}\)์— ๋Œ€ํ•ด ๋ณ‘๋ ฌ์ ์œผ๋กœ ๊ธธ์ด \(t\)์˜ \(r\) ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: \(c_{i}=(c_{i1},\dots,c_{it})\), ๊ฒฐ๊ณผ์ ์œผ๋กœ \(n\times r\)์˜ ๊ทผ๊ฑฐ ํ›„๋ณด๊ฐ€ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ•™์Šต๋œ \(<\)\(\mid\)start\(\mid\)\(>\) ๋ฐ \(<\)\(\mid\)end\(\mid\)\(>\) ํ† ํฐ์„ ์‚ฝ์ž…ํ•˜์—ฌ ๊ฐ ๊ทผ๊ฑฐ์˜ ์‹œ์ž‘๊ณผ ๋์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. 2. **์‚ฌํ›„ ๊ทผ๊ฑฐ์™€ ๊ธฐ๋ณธ ์˜ˆ์ธก ํ˜ผํ•ฉ(4.3์ ˆ)**: ๊ฐ ๊ทผ๊ฑฐ ๋’ค์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ ์ถœ๋ ฅ์—์„œ, ์‚ฌํ›„ ๊ทผ๊ฑฐ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ๋กœ์ง“๊ณผ ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ ์˜ˆ์ธก ๋กœ์ง“์„ ๋น„๊ตํ•˜์—ฌ ์–ผ๋งˆ๋‚˜ ๋งŽ์ด ํ†ตํ•ฉํ•ด์•ผ ํ•˜๋Š”์ง€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์–•์€ MLP์ธ "๋ฏน์‹ฑ ํ—ค๋“œ"๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ทผ๊ฑฐ๋ฅผ ๋„์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— Fine-tuning ์ดˆ๊ธฐ์— ๋ถ„ํฌ ์ด๋™์„ ์‰ฝ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. 3. **๊ทผ๊ฑฐ ์ƒ์„ฑ ์ตœ์ ํ™”(4.4์ ˆ)**: ๊ทผ๊ฑฐ ์ƒ์„ฑ ๋งค๊ฐœ๋ณ€์ˆ˜(์‹œ์ž‘/๋ ํ† ํฐ ๋ฐ LM ๊ฐ€์ค‘์น˜)๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ฏธ๋ž˜ ํ…์ŠคํŠธ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ด๋Š” ๊ทผ๊ฑฐ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ํ–ฅํ›„ ํ† ํฐ ์˜ˆ์ธก์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ทผ๊ฑฐ์— ํ•™์Šต ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด REINFORCE๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์‚ฐ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ต์‚ฌ ๊ฐ•์ œ ํŠธ๋ฆญ์„ ์ ์šฉํ•˜์—ฌ ์ƒ๊ฐ ๋’ค์˜ ํ† ํฐ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ดํ›„ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์†์‹ค์— ํฌํ•จ์‹œํ‚ต๋‹ˆ๋‹ค.
1. **๋ณ‘๋ ฌ ๊ทผ๊ฑฐ ์ƒ์„ฑ(4.2์ ˆ ์ฐธ์กฐ)**: ์ž…๋ ฅ ์‹œํ€€์Šค \(x_{0:n}\)์˜ \(n\) ํ† ํฐ \(x_{i}\)์— ๋Œ€ํ•ด ๋ณ‘๋ ฌ์ ์œผ๋กœ ๊ธธ์ด \(t\)์˜ \(r\) ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: \(c_{i}=(c_{i1},\dots,c_{it})\), ๊ฒฐ๊ณผ์ ์œผ๋กœ \(n\times r\)์˜ ๊ทผ๊ฑฐ ํ›„๋ณด๊ฐ€ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ•™์Šต๋œ \(<\)\(\mid\)start\(\mid\)\(>\) ๋ฐ \(<\)\(\mid\)end\(\mid\)\(>\) ํ† ํฐ์„ ์‚ฝ์ž…ํ•˜์—ฌ ๊ฐ ๊ทผ๊ฑฐ์˜ ์‹œ์ž‘๊ณผ ๋์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. 2. **์‚ฌํ›„ ๊ทผ๊ฑฐ์™€ ๊ธฐ๋ณธ ์˜ˆ์ธก ํ˜ผํ•ฉ(4.3์ ˆ)**: ๊ฐ ๊ทผ๊ฑฐ ๋’ค์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ ์ถœ๋ ฅ์—์„œ, ์‚ฌํ›„ ๊ทผ๊ฑฐ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ๋กœ์ง“๊ณผ ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ ์˜ˆ์ธก ๋กœ์ง“์„ ๋น„๊ตํ•˜์—ฌ ์–ผ๋งˆ๋‚˜ ๋งŽ์ด ํ†ตํ•ฉํ•ด์•ผ ํ•˜๋Š”์ง€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์–•์€ MLP์ธ "๋ฏน์‹ฑ ํ—ค๋“œ"๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ทผ๊ฑฐ๋ฅผ ๋„์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฏธ์„ธ ์กฐ์ • ์ดˆ๊ธฐ์— ๋ถ„ํฌ ์ด๋™์„ ์‰ฝ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. 3. **๊ทผ๊ฑฐ ์ƒ์„ฑ ์ตœ์ ํ™”(4.4์ ˆ)**: ๊ทผ๊ฑฐ ์ƒ์„ฑ ๋งค๊ฐœ๋ณ€์ˆ˜(์‹œ์ž‘/๋ ํ† ํฐ ๋ฐ LM ๊ฐ€์ค‘์น˜)๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ฏธ๋ž˜ ํ…์ŠคํŠธ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ด๋Š” ๊ทผ๊ฑฐ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ํ–ฅํ›„ ํ† ํฐ ์˜ˆ์ธก์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ทผ๊ฑฐ์— ํ•™์Šต ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด REINFORCE๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์‚ฐ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ต์‚ฌ ๊ฐ•์ œ ํŠธ๋ฆญ์„ ์ ์šฉํ•˜์—ฌ ์ƒ๊ฐ ๋’ค์˜ ํ† ํฐ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ดํ›„ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์†์‹ค์— ํฌํ•จ์‹œํ‚ต๋‹ˆ๋‹ค.
Directed evaluationDirected evaluation involves principled and clearly defined evaluation of models for known risks. Typically, models are tested against a set of clearly defined prompts that have been assigned to a clear set of categories and subcategories. Benchmarks and evaluation suites are typically directed evaluation, such as [30, 31, 152, 153]. Another form of directed evaluation is testing models' Natural Language Understanding for toxic content, which involves using LMs as zero-shot or few-shot classifiers to assess whether user-generated content is a violation of safety policies. If models are good at this task, it indicates that they have a strong natural language understanding of hazardous content [154], and therefore have the potential to be safe. The primary benefit of directed evaluation is that the results are highly interpretable and standardized, which enables us to make comparisons across time and across models. However, one limitation is that since the tests are not tailored to the characteristics or capabilities of the individual models, they may not fully challenge or evaluate the unique aspects of each model. Further, it takes time to develop, release and update directed evaluation test sets, which risks them going out of date given the rapid pace of AI development [155].
์ง€์‹œํ˜• ํ‰๊ฐ€. ์ง€์‹œํ˜• ํ‰๊ฐ€์—๋Š” ์•Œ๋ ค์ง„ ์œ„ํ—˜์— ๋Œ€ํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์›์น™์ ์ด๊ณ  ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋œ ํ‰๊ฐ€๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ์€ ๋ช…ํ™•ํ•œ ๋ฒ”์ฃผ ๋ฐ ํ•˜์œ„ ๋ฒ”์ฃผ ์ง‘ํ•ฉ์— ํ• ๋‹น๋œ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋œ ํ”„๋กฌํ”„ํŠธ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ…Œ์ŠคํŠธ๋ฉ๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๋ฐ ํ‰๊ฐ€ ์ œํ’ˆ๊ตฐ์€ ์ผ๋ฐ˜์ ์œผ๋กœ [30, 31, 152, 153]๊ณผ ๊ฐ™์€ ์ง€์‹œํ˜• ํ‰๊ฐ€์ž…๋‹ˆ๋‹ค. ์ง€์‹œํ˜• ํ‰๊ฐ€์˜ ๋˜ ๋‹ค๋ฅธ ํ˜•ํƒœ๋Š” ์œ ํ•ด ์ฝ˜ํ…์ธ ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ Natural Language Understanding์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‚ฌ์šฉ์ž ์ƒ์„ฑ ์ฝ˜ํ…์ธ ๊ฐ€ ์•ˆ์ „ ์ •์ฑ…์„ ์œ„๋ฐ˜ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด LMs์„ zero-shot ๋˜๋Š” few-shot ๋ถ„๋ฅ˜๊ธฐ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ฉด ํ•ด๋‹น ๋ชจ๋ธ์ด ์œ„ํ—˜ํ•œ ์ฝ˜ํ…์ธ ์— ๋Œ€ํ•œ Natural Language Understanding์ด ๋›ฐ์–ด๋‚˜๋ฏ€๋กœ ์•ˆ์ „ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค[154]. ์ง€์‹œํ˜• ํ‰๊ฐ€์˜ ์ฃผ์š” ์žฅ์ ์€ ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋„๋กœ ํ•ด์„ํ•  ์ˆ˜ ์žˆ๊ณ  ํ‘œ์ค€ํ™”๋˜์–ด ์žˆ์–ด ์‹œ๊ฐ„๊ณผ ๋ชจ๋ธ ๊ฐ„์— ๋น„๊ตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ•œ ๊ฐ€์ง€ ํ•œ๊ณ„๋Š” ํ…Œ์ŠคํŠธ๊ฐ€ ๊ฐœ๋ณ„ ๋ชจ๋ธ์˜ ํŠน์„ฑ์ด๋‚˜ ๊ธฐ๋Šฅ์— ๋งž๊ฒŒ ์กฐ์ •๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ ๋ชจ๋ธ์˜ ๊ณ ์œ ํ•œ ์ธก๋ฉด์„ ์™„์ „ํžˆ ๋„์ „ํ•˜๊ฑฐ๋‚˜ ํ‰๊ฐ€ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ง€์‹œํ˜• ํ‰๊ฐ€ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœ, ์ถœ์‹œ ๋ฐ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์‹œ๊ฐ„์ด ๊ฑธ๋ฆฌ๋ฏ€๋กœ AI ๊ฐœ๋ฐœ์˜ ๋น ๋ฅธ ์†๋„๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์˜ค๋ž˜๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ ๋  ์œ„ํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค[155].
์ง€์‹œํ˜• ํ‰๊ฐ€์ง€์‹œํ˜• ํ‰๊ฐ€์—๋Š” ์•Œ๋ ค์ง„ ์œ„ํ—˜์— ๋Œ€ํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์›์น™์ ์ด๊ณ  ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋œ ํ‰๊ฐ€๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ์€ ๋ช…ํ™•ํ•œ ๋ฒ”์ฃผ ๋ฐ ํ•˜์œ„ ๋ฒ”์ฃผ ์ง‘ํ•ฉ์— ํ• ๋‹น๋œ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋œ ํ”„๋กฌํ”„ํŠธ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ…Œ์ŠคํŠธ๋ฉ๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๋ฐ ํ‰๊ฐ€ ์ œํ’ˆ๊ตฐ์€ ์ผ๋ฐ˜์ ์œผ๋กœ [30, 31, 152, 153]๊ณผ ๊ฐ™์€ ์ง€์‹œํ˜• ํ‰๊ฐ€์ž…๋‹ˆ๋‹ค. ์ง€์‹œํ˜• ํ‰๊ฐ€์˜ ๋˜ ๋‹ค๋ฅธ ํ˜•ํƒœ๋Š” ์œ ํ•ด ์ฝ˜ํ…์ธ ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์ž์—ฐ์–ด ์ดํ•ด๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‚ฌ์šฉ์ž ์ƒ์„ฑ ์ฝ˜ํ…์ธ ๊ฐ€ ์•ˆ์ „ ์ •์ฑ…์„ ์œ„๋ฐ˜ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด LM์„ ์ œ๋กœ์ƒท ๋˜๋Š” ํ‘ธ์ƒท ๋ถ„๋ฅ˜๊ธฐ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ฉด ํ•ด๋‹น ๋ชจ๋ธ์ด ์œ„ํ—˜ํ•œ ์ฝ˜ํ…์ธ ์— ๋Œ€ํ•œ ์ž์—ฐ์–ด ์ดํ•ด๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋ฏ€๋กœ ์•ˆ์ „ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค[154]. ์ง€์‹œํ˜• ํ‰๊ฐ€์˜ ์ฃผ์š” ์žฅ์ ์€ ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋„๋กœ ํ•ด์„ํ•  ์ˆ˜ ์žˆ๊ณ  ํ‘œ์ค€ํ™”๋˜์–ด ์žˆ์–ด ์‹œ๊ฐ„๊ณผ ๋ชจ๋ธ ๊ฐ„์— ๋น„๊ตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ•œ ๊ฐ€์ง€ ํ•œ๊ณ„๋Š” ํ…Œ์ŠคํŠธ๊ฐ€ ๊ฐœ๋ณ„ ๋ชจ๋ธ์˜ ํŠน์„ฑ์ด๋‚˜ ๊ธฐ๋Šฅ์— ๋งž๊ฒŒ ์กฐ์ •๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ ๋ชจ๋ธ์˜ ๊ณ ์œ ํ•œ ์ธก๋ฉด์„ ์™„์ „ํžˆ ๋„์ „ํ•˜๊ฑฐ๋‚˜ ํ‰๊ฐ€ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ง€์‹œํ˜• ํ‰๊ฐ€ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœ, ์ถœ์‹œ ๋ฐ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์‹œ๊ฐ„์ด ๊ฑธ๋ฆฌ๋ฏ€๋กœ AI ๊ฐœ๋ฐœ์˜ ๋น ๋ฅธ ์†๋„๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์˜ค๋ž˜๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ ๋  ์œ„ํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค[155].
* Doc2Dial (Feng et al., 2020) is a document-grounded conversational QA dataset covering four domains: DMV, SSA, VA, and Student Aid. Each sample comprises a dialogue where a user poses queries regarding the document, and an agent responds those questions. The average document length is around 101K words. * QuAC (Choi et al., 2018) is based on Wikipedia documents. Originally, the document is short. Since each dialogue is linked to multiple Wikipedia URLs, we extract text from these links to increase the document size to approximately an average of 15K words. It contains unanswerable cases where answers cannot be found within the given context. * QReCC (Anantha et al., 2021) is an open-domain conversational QA dataset across multiple sources. Similar to QuAC, each dialogue also has corresponding URLs. We extract text from those URLs to construct the documents. In the end, the average document size is around 5K words, with a maximum document size of 20K words. * TopiOCQA (Adlakha et al., 2022) is grounded on the whole Wikipedia. It incorporates topic switching and requires the agent to search the entire Wikipedia for answers to user questions. * INSCIT (Wu et al., 2023) is also grounded on the whole Wikipedia. It studies the case where user questions are under-specified and require clarification.
* Doc2Dial(Feng et al., 2020)์€ DMV, SSA, VA, ํ•™์ž๊ธˆ ์ง€์› ๋“ฑ 4๊ฐœ์˜ ๋„๋ฉ”์ธ์„ ๋‹ค๋ฃจ๋Š” ๋ฌธ์„œ ๊ธฐ๋ฐ˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ๊ฐ ์ƒ˜ํ”Œ์€ ์‚ฌ์šฉ์ž๊ฐ€ ๋ฌธ์„œ์™€ ๊ด€๋ จํ•˜์—ฌ ์งˆ๋ฌธ์„ ํ•˜๊ณ  ์ƒ๋‹ด์›์ด ํ•ด๋‹น ์งˆ๋ฌธ์— ์‘๋‹ตํ•˜๋Š” ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ‰๊ท  ๋ฌธ์„œ ๊ธธ์ด๋Š” ์•ฝ 101K ๋‹จ์–ด์ž…๋‹ˆ๋‹ค. * QuAC(Choi et al., 2018)์€ Wikipedia ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ ๋ฌธ์„œ๋Š” ์งง์Šต๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™”๋Š” ์—ฌ๋Ÿฌ Wikipedia URL์— ์—ฐ๊ฒฐ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋งํฌ์—์„œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜์—ฌ ๋ฌธ์„œ ํฌ๊ธฐ๋ฅผ ํ‰๊ท  ์•ฝ 15K ๋‹จ์–ด๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ ๋‚ด์—์„œ ๋‹ต์„ ์ฐพ์„ ์ˆ˜ ์—†๋Š” ๋‹ต๋ณ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. * QReCC(Anantha et al., 2021)๋Š” ์—ฌ๋Ÿฌ ์†Œ์Šค์— ๊ฑธ์นœ ๊ฐœ๋ฐฉํ˜• ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. QuAC์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๊ฐ ๋Œ€ํ™”์—๋Š” ํ•ด๋‹น URL๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ URL์—์„œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜์—ฌ ๋ฌธ์„œ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข…์ ์œผ๋กœ ํ‰๊ท  ๋ฌธ์„œ ํฌ๊ธฐ๋Š” ์•ฝ 5K ๋‹จ์–ด์ด๋ฉฐ ์ตœ๋Œ€ ๋ฌธ์„œ ํฌ๊ธฐ๋Š” 20K ๋‹จ์–ด์ž…๋‹ˆ๋‹ค. * TopiOCQA(Adlakha et al., 2022)๋Š” ์ „์ฒด Wikipedia๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ฃผ์ œ ์ „ํ™˜์„ ํ†ตํ•ฉํ•˜๋ฉฐ, ์ƒ๋‹ด์›์ด ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ์ฐพ๊ธฐ ์œ„ํ•ด ์ „์ฒด Wikipedia๋ฅผ ๊ฒ€์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. * INSCIT(Wu et al., 2023)๋„ ์ „์ฒด Wikipedia๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์ด ๋ช…์‹œ๋˜์ง€ ์•Š์•„ ๋ช…ํ™•ํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
* Doc2Dial(Feng et al., 2020)์€ DMV, SSA, VA, ํ•™์ž๊ธˆ ์ง€์› ๋“ฑ 4๊ฐœ์˜ ๋„๋ฉ”์ธ์„ ๋‹ค๋ฃจ๋Š” ๋ฌธ์„œ ๊ธฐ๋ฐ˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ๊ฐ ์ƒ˜ํ”Œ์€ ์‚ฌ์šฉ์ž๊ฐ€ ๋ฌธ์„œ์™€ ๊ด€๋ จํ•˜์—ฌ ์งˆ๋ฌธ์„ ํ•˜๊ณ  ์ƒ๋‹ด์›์ด ํ•ด๋‹น ์งˆ๋ฌธ์— ์‘๋‹ตํ•˜๋Š” ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ‰๊ท  ๋ฌธ์„œ ๊ธธ์ด๋Š” ์•ฝ 101K ๋‹จ์–ด์ž…๋‹ˆ๋‹ค. * QuAC(Choi et al., 2018)์€ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ ๋ฌธ์„œ๋Š” ์งง์Šต๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™”๋Š” ์—ฌ๋Ÿฌ ์œ„ํ‚ค๋ฐฑ๊ณผ URL์— ์—ฐ๊ฒฐ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋งํฌ์—์„œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜์—ฌ ๋ฌธ์„œ ํฌ๊ธฐ๋ฅผ ํ‰๊ท  ์•ฝ 15K ๋‹จ์–ด๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ ๋‚ด์—์„œ ๋‹ต์„ ์ฐพ์„ ์ˆ˜ ์—†๋Š” ๋‹ต๋ณ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. * QReCC(Anantha et al., 2021)๋Š” ์—ฌ๋Ÿฌ ์†Œ์Šค์— ๊ฑธ์นœ ๊ฐœ๋ฐฉํ˜• ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. QuAC์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๊ฐ ๋Œ€ํ™”์—๋Š” ํ•ด๋‹น URL๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ URL์—์„œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜์—ฌ ๋ฌธ์„œ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข…์ ์œผ๋กœ ํ‰๊ท  ๋ฌธ์„œ ํฌ๊ธฐ๋Š” ์•ฝ 5K ๋‹จ์–ด์ด๋ฉฐ ์ตœ๋Œ€ ๋ฌธ์„œ ํฌ๊ธฐ๋Š” 20K ๋‹จ์–ด์ž…๋‹ˆ๋‹ค. * TopiOCQA(Adlakha et al., 2022)๋Š” ์ „์ฒด ์œ„ํ‚ค๋ฐฑ๊ณผ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ฃผ์ œ ์ „ํ™˜์„ ํ†ตํ•ฉํ•˜๋ฉฐ, ์ƒ๋‹ด์›์ด ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ์ฐพ๊ธฐ ์œ„ํ•ด ์ „์ฒด ์œ„ํ‚ค๋ฐฑ๊ณผ๋ฅผ ๊ฒ€์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. * INSCIT(Wu et al., 2023)๋„ ์ „์ฒด ์œ„ํ‚ค๋ฐฑ๊ณผ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์ด ๋ช…์‹œ๋˜์ง€ ์•Š์•„ ๋ช…ํ™•ํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
Instruction tuning aims to train large language models to generate responses that align with input instructions, thereby enable LLMs with conversational and task execution capabilities. Compared with standard LLM, SFT allows the behavior of the model to be more controllable and predictable, thereby achieving the purpose of aligning human will. Methods to build instruction tuning datasets include: (1) Pure manual annotation Conover et al. (2023). This method completely constructs instructions and answers manually, which is very time-consuming and laborious; (2) Converted from existing datasets Mishra et al. (2022); Sanh et al. (2022); Chung et al. (2022). Some studies use supervised data sets from NLP tasks to construct instruction tuning data; (3) Automatically generated using LLM Honovich et al. (2022); Wang et al. (2023); Xu et al. (2023); Ji et al. (2023); Xu et al. (2023). Others use existing LLMs to generate instruction tuning data. A common practice is to first manually annotate high-quality seed datasets, and then use LLM to expand the seed instructions and corresponding outputs, which can generate large-scale instruction tuning data with very little human annotation. However, the quality cannot be guaranteed and there is a certain amount of noisy data, which can lead to hallucinations.
Instruction tuning์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž…๋ ฅ ๋ช…๋ น์— ๋งž๊ฒŒ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จํ•˜์—ฌ LLM์— ๋Œ€ํ™” ๋ฐ ์ž‘์—… ์‹คํ–‰ ๊ธฐ๋Šฅ์„ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ์ค€ LLM๊ณผ ๋น„๊ตํ•˜์—ฌ Supervised Fine-tuning (SFT)์„ ์‚ฌ์šฉํ•˜๋ฉด ๋ชจ๋ธ์˜ ๋™์ž‘์„ ๋ณด๋‹ค ์ œ์–ดํ•  ์ˆ˜ ์žˆ๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์–ด ์ธ๊ฐ„์˜ ์˜์ง€๋ฅผ ์กฐ์ •ํ•˜๋Š” ๋ชฉ์ ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Instruction tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐฉ๋ฒ•์—๋Š” ๋‹ค์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค: (1) ์ˆœ์ˆ˜ํ•œ ์ˆ˜๋™ ์ฃผ์„ Conover et al. (2023). ์ด ๋ฐฉ๋ฒ•์€ ์ง€์นจ๊ณผ ๋‹ต๋ณ€์„ ์™„์ „ํžˆ ์ˆ˜๋™์œผ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‹œ๊ฐ„๊ณผ ๋…ธ๋™ ์ง‘์•ฝ์ ์ž…๋‹ˆ๋‹ค; (2) ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ณ€ํ™˜ Mishra et al. (2022); Sanh et al. (2022); Chung et al. (2022). ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” NLP ์ž‘์—…์˜ ์ง€๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Instruction tuning ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค; (3) LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž๋™ ์ƒ์„ฑ Honovich et al. (2022); Wang et al. (2023); Xu et al. (2023); Ji et al. (2023); Xu et al. (2023). ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ธฐ์กด LLM์„ ์‚ฌ์šฉํ•˜์—ฌ Instruction tuning ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์€ ๋จผ์ € ๊ณ ํ’ˆ์งˆ ์‹œ๋“œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์€ ๋‹ค์Œ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๋“œ ๋ช…๋ น๊ณผ ํ•ด๋‹น ์ถœ๋ ฅ์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‚ฌ๋žŒ์˜ ์ฃผ์„์ด ๊ฑฐ์˜ ์—†์–ด๋„ ๋Œ€๊ทœ๋ชจ Instruction tuning ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ’ˆ์งˆ์„ ๋ณด์žฅํ•  ์ˆ˜ ์—†๊ณ  ๋…ธ์ด์ฆˆ ๋ฐ์ดํ„ฐ๊ฐ€ ์ผ์ •๋Ÿ‰ ์กด์žฌํ•˜์—ฌ Hallucination ํ˜„์ƒ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ช…๋ น ํŠœ๋‹์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž…๋ ฅ ๋ช…๋ น์— ๋งž๊ฒŒ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จํ•˜์—ฌ LLM์— ๋Œ€ํ™” ๋ฐ ์ž‘์—… ์‹คํ–‰ ๊ธฐ๋Šฅ์„ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ์ค€ LLM๊ณผ ๋น„๊ตํ•˜์—ฌ SFT๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ชจ๋ธ์˜ ๋™์ž‘์„ ๋ณด๋‹ค ์ œ์–ดํ•  ์ˆ˜ ์žˆ๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์–ด ์ธ๊ฐ„์˜ ์˜์ง€๋ฅผ ์กฐ์ •ํ•˜๋Š” ๋ชฉ์ ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ช…๋ น ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐฉ๋ฒ•์—๋Š” ๋‹ค์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค: (1) ์ˆœ์ˆ˜ํ•œ ์ˆ˜๋™ ์ฃผ์„ Conover et al. (2023). ์ด ๋ฐฉ๋ฒ•์€ ์ง€์นจ๊ณผ ๋‹ต๋ณ€์„ ์™„์ „ํžˆ ์ˆ˜๋™์œผ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‹œ๊ฐ„๊ณผ ๋…ธ๋™ ์ง‘์•ฝ์ ์ž…๋‹ˆ๋‹ค; (2) ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ณ€ํ™˜ Mishra et al. (2022); Sanh et al. (2022); Chung et al. (2022). ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” NLP ์ž‘์—…์˜ ์ง€๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ช…๋ น ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค; (3) LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž๋™ ์ƒ์„ฑ Honovich et al. (2022); Wang et al. (2023); Xu et al. (2023); Ji et al. (2023); Xu et al. (2023). ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ธฐ์กด LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ช…๋ น ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์€ ๋จผ์ € ๊ณ ํ’ˆ์งˆ ์‹œ๋“œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์€ ๋‹ค์Œ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๋“œ ๋ช…๋ น๊ณผ ํ•ด๋‹น ์ถœ๋ ฅ์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‚ฌ๋žŒ์˜ ์ฃผ์„์ด ๊ฑฐ์˜ ์—†์–ด๋„ ๋Œ€๊ทœ๋ชจ ๋ช…๋ น ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ’ˆ์งˆ์„ ๋ณด์žฅํ•  ์ˆ˜ ์—†๊ณ  ๋…ธ์ด์ฆˆ ๋ฐ์ดํ„ฐ๊ฐ€ ์ผ์ •๋Ÿ‰ ์กด์žฌํ•˜์—ฌ ํ™˜๊ฐ ํ˜„์ƒ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**LLMs as Recommendation Models.** With specific methods or mechanisms, LLMs can be adapted to serve as recommendation models. Existing work along this line can be generally divided into two main categories. First, some methods prompt LLMs for completing the recommendation task in a zero-shot paradigm (_i.e._, without parameter tuning) [805, 806]. A series of prompt engineering methods like recency-focused and in-context learning are introduced to improve recommendation performance as well as alleviate the potential model biases [807, 808]. Second, another category of studies aim to specialize LLMs for personalized recommendation through instruction tuning [357, 809]. Specially, high-quality instruction data is key to adapt LLMs to the recommendation tasks, which can be constructed based on user-item interactions with heuristic templates. To further improve the instruction diversity, InstructRec [357] employs self-instruct technique to simulate large amounts of potential user instructions in various scenarios like product search and personalized recommendations. In addition to representing each item by its text description, there is also growing attention on extending LLM's vocabulary with semantic identifiers in recommender systems [810, 811], to incorporate collaborative semantics into LLMs.
**์ถ”์ฒœ ๋ชจ๋ธ๋กœ์„œ์˜ LLM**. ํŠน์ • ๋ฐฉ๋ฒ•์ด๋‚˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์‚ฌ์šฉํ•˜๋ฉด LLM์„ ์ถ”์ฒœ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๋„๋ก ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จ๋œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ํŠœ๋‹ ์—†์ด(_์ฆ‰, zero-shot ํŒจ๋Ÿฌ๋‹ค์ž„) ์ถ”์ฒœ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋„๋ก LLM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋ƒ…๋‹ˆ๋‹ค[805, 806]. recency-focused ๋ฐ In-Context Learning๊ณผ ๊ฐ™์€ ์ผ๋ จ์˜ prompt engineering ๋ฐฉ๋ฒ•์ด ๋„์ž…๋˜์–ด ์ถ”์ฒœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ณ  ์ž ์žฌ์ ์ธ ๋ชจ๋ธ ํŽธํ–ฅ์„ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค[807, 808]. ๋‘˜์งธ, ๋˜ ๋‹ค๋ฅธ ๋ฒ”์ฃผ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” Instruction Tuning์„ ํ†ตํ•ด ๊ฐœ์ธํ™”๋œ ์ถ”์ฒœ์„ ์œ„ํ•ด LLM์„ ์ „๋ฌธํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค[357, 809]. ํŠนํžˆ, ๊ณ ํ’ˆ์งˆ Instruction ๋ฐ์ดํ„ฐ๋Š” ์ถ”์ฒœ ์ž‘์—…์— LLM์„ ์ ์šฉํ•˜๋Š” ๋ฐ ํ•ต์‹ฌ์ ์ธ๋ฐ, ์ด๋Š” ํœด๋ฆฌ์Šคํ‹ฑ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž-์•„์ดํ…œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด InstructRec [357]์€ self-instruct ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ œํ’ˆ ๊ฒ€์ƒ‰ ๋ฐ ๊ฐœ์ธํ™” ์ถ”์ฒœ๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ž ์žฌ์ ์ธ ์‚ฌ์šฉ์ž ์ง€์นจ์˜ ๋Œ€๋Ÿ‰์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํ•ญ๋ชฉ์„ ํ…์ŠคํŠธ ์„ค๋ช…์œผ๋กœ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ์ถ”์ฒœ ์‹œ์Šคํ…œ์—์„œ ์˜๋ฏธ๋ก ์  ์‹๋ณ„์ž๋กœ LLM์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜์—ฌ ํ˜‘์—… ์˜๋ฏธ๋ก ์„ LLM์— ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[810, 811].
**์ถ”์ฒœ ๋ชจ๋ธ๋กœ์„œ์˜ LLM**. ํŠน์ • ๋ฐฉ๋ฒ•์ด๋‚˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์‚ฌ์šฉํ•˜๋ฉด LLM์„ ์ถ”์ฒœ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๋„๋ก ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จ๋œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ํŠœ๋‹ ์—†์ด(_์ฆ‰, ์ œ๋กœ ์ƒท ํŒจ๋Ÿฌ๋‹ค์ž„) ์ถ”์ฒœ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋„๋ก LLM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋ƒ…๋‹ˆ๋‹ค[805, 806]. ์ตœ๊ทผ์„ฑ ์ค‘์‹ฌ ๋ฐ ์ƒํ™ฉ์— ๋”ฐ๋ฅธ ํ•™์Šต๊ณผ ๊ฐ™์€ ์ผ๋ จ์˜ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง ๋ฐฉ๋ฒ•์ด ๋„์ž…๋˜์–ด ์ถ”์ฒœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ณ  ์ž ์žฌ์ ์ธ ๋ชจ๋ธ ํŽธํ–ฅ์„ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค[807, 808]. ๋‘˜์งธ, ๋˜ ๋‹ค๋ฅธ ๋ฒ”์ฃผ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง€์‹œ ํŠœ๋‹์„ ํ†ตํ•ด ๊ฐœ์ธํ™”๋œ ์ถ”์ฒœ์„ ์œ„ํ•ด LLM์„ ์ „๋ฌธํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค[357, 809]. ํŠนํžˆ, ๊ณ ํ’ˆ์งˆ ์ง€์นจ ๋ฐ์ดํ„ฐ๋Š” ์ถ”์ฒœ ์ž‘์—…์— LLM์„ ์ ์šฉํ•˜๋Š” ๋ฐ ํ•ต์‹ฌ์ ์ธ๋ฐ, ์ด๋Š” ํœด๋ฆฌ์Šคํ‹ฑ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž-์•„์ดํ…œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด InstructRec [357]์€ ์ž์ฒด ์ง€์นจ ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ œํ’ˆ ๊ฒ€์ƒ‰ ๋ฐ ๊ฐœ์ธํ™” ์ถ”์ฒœ๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ž ์žฌ์ ์ธ ์‚ฌ์šฉ์ž ์ง€์นจ์˜ ๋Œ€๋Ÿ‰์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํ•ญ๋ชฉ์„ ํ…์ŠคํŠธ ์„ค๋ช…์œผ๋กœ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ์ถ”์ฒœ ์‹œ์Šคํ…œ์—์„œ ์˜๋ฏธ๋ก ์  ์‹๋ณ„์ž๋กœ LLM์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜์—ฌ ํ˜‘์—… ์˜๋ฏธ๋ก ์„ LLM์— ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[810, 811].
* CoQA (Reddy et al., 2019) is a conversational QA dataset with each dialogue grounded on a short passage. The answers are generally short, and the passages cover a wide range of domains like children's stories, literature, mid/high school exams, news, Wikipedia. * DoQA (Campos et al., 2020) covers three domains: cooking, travel, and movies collected from active Stack Exchange4 forums. The dataset contains unanswerable cases where answers cannot be found within the given document. Footnote 4: [https://stackexchange.com/](https://stackexchange.com/) * ConvFinQA (Chen et al., 2022) is based on the Financial domain. Each document contains a single financial report table along with relevant text surrounding the table. This dataset involves arithmetic calculation and complex numerical reasoning. * SQA (Pasupat and Liang, 2015) is grounded on documents which contain only a single Table without any surrounding text. The documents are collected from Wikipedia, and the questions are highly compositional, which requires the model with robust table comprehension ability to give correct answers. * HybridDial (Nakamura et al., 2022) is a conversational QA dataset grounded on documents containing both Wikipedia tabular and textual data. The questions are complex which requires reasoning over the documents.
* CoQA(Reddy et al., 2019)๋Š” ๊ฐ ๋Œ€ํ™”๊ฐ€ ์งง์€ ๊ตฌ์ ˆ์— ๊ธฐ๋ฐ˜ํ•œ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ๋‹ต๋ณ€์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์งง์œผ๋ฉฐ, ๊ตฌ์ ˆ์€ ๋™ํ™”, ๋ฌธํ•™, ์ค‘๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜, ๋‰ด์Šค, Wikipedia ๋“ฑ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. * DoQA(Campos et al., 2020)๋Š” ํ™œ์„ฑ Stack Exchange ํฌ๋Ÿผ์—์„œ ์ˆ˜์ง‘ํ•œ ์š”๋ฆฌ, ์—ฌํ–‰, ์˜ํ™”์˜ ์„ธ ๊ฐ€์ง€ ๋„๋ฉ”์ธ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ฃผ์–ด์ง„ ๋ฌธ์„œ ๋‚ด์—์„œ ๋‹ต์„ ์ฐพ์„ ์ˆ˜ ์—†๋Š” ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ์ฃผ 4: [https://stackexchange.com/](https://stackexchange.com/) * ConvFinQA(Chen et al., 2022)๋Š” ๊ธˆ์œต ๋„๋ฉ”์ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ฌธ์„œ์—๋Š” ํ…Œ์ด๋ธ”๊ณผ ํ…Œ์ด๋ธ” ์ฃผ๋ณ€์˜ ๊ด€๋ จ ํ…์ŠคํŠธ์™€ ํ•จ๊ป˜ ๋‹จ์ผ ์žฌ๋ฌด ๋ณด๊ณ ์„œ ํ…Œ์ด๋ธ”์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” arithmetic calculation๊ณผ ๋ณต์žกํ•œ numerical reasoning์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. * SQA(Pasupat and Liang, 2015)๋Š” ์ฃผ๋ณ€ ํ…์ŠคํŠธ ์—†์ด ๋‹จ์ผ ํ…Œ์ด๋ธ”๋งŒ ํฌํ•จ๋œ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ๋Š” Wikipedia์—์„œ ์ˆ˜์ง‘๋˜์—ˆ์œผ๋ฉฐ, ์งˆ๋ฌธ์€ ๋งค์šฐ ๊ตฌ์„ฑ์ ์ด์–ด์„œ ๊ฐ•๋ ฅํ•œ ํ‘œ ์ดํ•ด ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ˜ ๋ชจ๋ธ์ด ์ •๋‹ต์„ ์ œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. * HybridDial(Nakamura et al., 2022)๋Š” Wikipedia์˜ ํ‘œ ๋ฐ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์งˆ๋ฌธ์€ ๋ฌธ์„œ์— ๋Œ€ํ•œ reasoning์ด ํ•„์š”ํ•œ ๋ณต์žกํ•œ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค.
* CoQA(Reddy et al., 2019)๋Š” ๊ฐ ๋Œ€ํ™”๊ฐ€ ์งง์€ ๊ตฌ์ ˆ์— ๊ธฐ๋ฐ˜ํ•œ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ๋‹ต๋ณ€์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์งง์œผ๋ฉฐ, ๊ตฌ์ ˆ์€ ๋™ํ™”, ๋ฌธํ•™, ์ค‘๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜, ๋‰ด์Šค, ์œ„ํ‚คํ”ผ๋””์•„ ๋“ฑ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. * DoQA(Campos et al., 2020)๋Š” ํ™œ์„ฑ ์Šคํƒ ๊ตํ™˜4 ํฌ๋Ÿผ์—์„œ ์ˆ˜์ง‘ํ•œ ์š”๋ฆฌ, ์—ฌํ–‰, ์˜ํ™”์˜ ์„ธ ๊ฐ€์ง€ ๋„๋ฉ”์ธ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ฃผ์–ด์ง„ ๋ฌธ์„œ ๋‚ด์—์„œ ๋‹ต์„ ์ฐพ์„ ์ˆ˜ ์—†๋Š” ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ์ฃผ 4: [https://stackexchange.com/](https://stackexchange.com/) * ConvFinQA(Chen et al., 2022)๋Š” ๊ธˆ์œต ๋„๋ฉ”์ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ฌธ์„œ์—๋Š” ํ…Œ์ด๋ธ”๊ณผ ํ…Œ์ด๋ธ” ์ฃผ๋ณ€์˜ ๊ด€๋ จ ํ…์ŠคํŠธ์™€ ํ•จ๊ป˜ ๋‹จ์ผ ์žฌ๋ฌด ๋ณด๊ณ ์„œ ํ…Œ์ด๋ธ”์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์‚ฐ์ˆ  ๊ณ„์‚ฐ๊ณผ ๋ณต์žกํ•œ ์ˆ˜๋ฆฌ ์ถ”๋ก ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. * SQA(Pasupat and Liang, 2015)๋Š” ์ฃผ๋ณ€ ํ…์ŠคํŠธ ์—†์ด ๋‹จ์ผ ํ…Œ์ด๋ธ”๋งŒ ํฌํ•จ๋œ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ๋Š” ์œ„ํ‚คํ”ผ๋””์•„์—์„œ ์ˆ˜์ง‘๋˜์—ˆ์œผ๋ฉฐ, ์งˆ๋ฌธ์€ ๋งค์šฐ ๊ตฌ์„ฑ์ ์ด์–ด์„œ ๊ฐ•๋ ฅํ•œ ํ‘œ ์ดํ•ด ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ˜ ๋ชจ๋ธ์ด ์ •๋‹ต์„ ์ œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. * ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ๋‹ค์ด์–ผ(Nakamura et al., 2022)์€ ์œ„ํ‚คํ”ผ๋””์•„์˜ ํ‘œ ๋ฐ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์งˆ๋ฌธ์€ ๋ฌธ์„œ์— ๋Œ€ํ•œ ์ถ”๋ก ์ด ํ•„์š”ํ•œ ๋ณต์žกํ•œ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค.
Fig. 39: HuggingGPT: An agent-based approach to use tools and planning [image courtesy of [171]]* **RACE**[186] suits for reading comprehension task. This dataset is based on English tests completed by Chinese students from middle school and high school, aged \(12\) to \(18\), and it contains roughly \(28,000\) texts and \(100,000\) questions rigorously prepared by human specialists, primarily English instructors. This dataset contains a wide range of subjects that were purposefully chosen to assess students' comprehension and reasoning abilities. This dataset is available in three subgroups: RACE-M, RACE-H, and RACE. RACE-M refers to the middle school examinations, whereas RACE-H denotes the high school tests. Finally, RACE is the synthesis of RACE-M and RACE-H. * **SQuAD**[187] stands for "Stanford Question Answering Dataset" and is a crowdsourced reading comprehension dataset based on Wikipedia articles. It has approximately \(100,000\) question-answer pairs connected to more than \(500\) articles. The answers to these questions are typically text fragments or spans taken from the corresponding reading passages. The questions may be unanswerable in some cases. The dataset is divided into three sets: an \(80\%\) training set, a \(10\%\) development set, and a \(10\%\) hidden test set.
๊ทธ๋ฆผ 39: HuggingGPT: ๋„๊ตฌ ๋ฐ ๊ณ„ํš์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹ [171] ์ด๋ฏธ์ง€ ์˜ˆ์‹œ] * **RACE**[186]๋Š” ๋…ํ•ด ๊ณผ์ œ์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘๊ณ ๋“ฑํ•™๊ต ์ค‘๊ณ ๋“ฑํ•™์ƒ์ธ 12~18์„ธ์˜ ์ค‘๊ตญ ํ•™์ƒ๋“ค์ด ์‘์‹œํ•œ ์˜์–ด ์‹œํ—˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ธ๊ฐ„ ์ „๋ฌธ๊ฐ€, ์ฃผ๋กœ ์˜์–ด ๊ต์‚ฌ๊ฐ€ ์—„๊ฒฉํ•˜๊ฒŒ ์ค€๋น„ํ•œ ์•ฝ 28,000๊ฐœ์˜ ํ…์ŠคํŠธ์™€ 100,000๊ฐœ์˜ ์งˆ๋ฌธ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ํ•™์ƒ๋“ค์˜ ์ดํ•ด๋ ฅ๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์˜๋„์ ์œผ๋กœ ์„ ํƒ๋œ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” RACE-M, RACE-H, RACE์˜ ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ๊ทธ๋ฃน์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. RACE-M์€ ์ค‘ํ•™๊ต ์‹œํ—˜์„ ์˜๋ฏธํ•˜๋ฉฐ, RACE-H๋Š” ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ RACE๋Š” RACE-M๊ณผ RACE-H์˜ ํ•ฉ์„ฑ์–ด์ž…๋‹ˆ๋‹ค. * **SQuAD**[187]๋Š” "Stanford Question Answering Dataset"์˜ ์•ฝ์ž๋กœ, ์œ„ํ‚ค๋ฐฑ๊ณผ ๊ธฐ์‚ฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ํฌ๋ผ์šฐ๋“œ ์†Œ์‹ฑ ๋…ํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์•ฝ 500๊ฐœ ์ด์ƒ์˜ ๊ธฐ์‚ฌ์™€ ์—ฐ๊ฒฐ๋œ ์•ฝ 100,000์Œ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ด๋‹น ์ฝ๊ธฐ ์ง€๋ฌธ์—์„œ ๋ฐœ์ทŒํ•œ ํ…์ŠคํŠธ ์กฐ๊ฐ ๋˜๋Š” ๋ฒ”์œ„์ž…๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 80%์˜ ํ›ˆ๋ จ ์„ธํŠธ, 10%์˜ ๊ฐœ๋ฐœ ์„ธํŠธ, 10%์˜ ์ˆจ๊ฒจ์ง„ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์„ธ ๊ฐ€์ง€ ์„ธํŠธ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 39: HuggingGPT: ๋„๊ตฌ ๋ฐ ๊ณ„ํš์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹ [171] ์ด๋ฏธ์ง€ ์˜ˆ์‹œ] * **RACE**[186]๋Š” ๋…ํ•ด ๊ณผ์ œ์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘๊ณ ๋“ฑํ•™๊ต ์ค‘๊ณ ๋“ฑํ•™์ƒ์ธ 12~18์„ธ์˜ ์ค‘๊ตญ ํ•™์ƒ๋“ค์ด ์‘์‹œํ•œ ์˜์–ด ์‹œํ—˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ธ๊ฐ„ ์ „๋ฌธ๊ฐ€, ์ฃผ๋กœ ์˜์–ด ๊ต์‚ฌ๊ฐ€ ์—„๊ฒฉํ•˜๊ฒŒ ์ค€๋น„ํ•œ ์•ฝ 28,000๊ฐœ์˜ ํ…์ŠคํŠธ์™€ 100,000๊ฐœ์˜ ์งˆ๋ฌธ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ํ•™์ƒ๋“ค์˜ ์ดํ•ด๋ ฅ๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์˜๋„์ ์œผ๋กœ ์„ ํƒ๋œ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” RACE-M, RACE-H, RACE์˜ ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ๊ทธ๋ฃน์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. RACE-M์€ ์ค‘ํ•™๊ต ์‹œํ—˜์„ ์˜๋ฏธํ•˜๋ฉฐ, RACE-H๋Š” ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ RACE๋Š” RACE-M๊ณผ RACE-H์˜ ํ•ฉ์„ฑ์–ด์ž…๋‹ˆ๋‹ค. * **SQuAD**[187]๋Š” "Stanford ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ"์˜ ์•ฝ์ž๋กœ, ์œ„ํ‚ค๋ฐฑ๊ณผ ๊ธฐ์‚ฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ํฌ๋ผ์šฐ๋“œ ์†Œ์‹ฑ ๋…ํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์•ฝ 500๊ฐœ ์ด์ƒ์˜ ๊ธฐ์‚ฌ์™€ ์—ฐ๊ฒฐ๋œ ์•ฝ 100,000์Œ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ด๋‹น ์ฝ๊ธฐ ์ง€๋ฌธ์—์„œ ๋ฐœ์ทŒํ•œ ํ…์ŠคํŠธ ์กฐ๊ฐ ๋˜๋Š” ๋ฒ”์œ„์ž…๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 80%์˜ ํ›ˆ๋ จ ์„ธํŠธ, 10%์˜ ๊ฐœ๋ฐœ ์„ธํŠธ, 10%์˜ ์ˆจ๊ฒจ์ง„ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์„ธ ๊ฐ€์ง€ ์„ธํŠธ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค.
We use each of the sentence fragments to construct multiple prompts by using templates. Each template corresponds to different types of interaction with a model. The interactions reflect different motivations and aims, and create different risks of harm. For instance, the segment fragment "make a bomb" can be incorporated into many distinct types of prompts, from describing past activities ("I once made a bomb") to asking for advice on whether to make a bomb ("How do I make a bomb?") or seeking encouragement ("Should I make a bomb?"). We developed the interaction types by drawing on theories of linguistics, specifically speech acts [54] and work in computer-human interactions [55]. This work indicates that people have different communicative goals which manifest when they interact with digital technology - and sometimes the same goals can be expressed in many different ways using different grammatical structures. This work has been widely used to understand the significance of social media content, conduct and conversations [56, 57, 58]. For the purposes of evaluating AI safety, we use this work to understand how a single goal (e.g., "learn how to make a bomb") can be expressed in many ways, such as by using an imperative or a question; and making the request directly or indirectly.
๊ฐ ๋ฌธ์žฅ ์กฐ๊ฐ์„ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ prompts๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํ…œํ”Œ๋ฆฟ์€ ๋ชจ๋ธ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ ์œ ํ˜•์— ๋”ฐ๋ผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ƒํ˜ธ ์ž‘์šฉ์€ ๋‹ค์–‘ํ•œ ๋™๊ธฐ์™€ ๋ชฉํ‘œ๋ฅผ ๋ฐ˜์˜ํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "ํญํƒ„์„ ๋งŒ๋“œ๋Š” ๊ฒƒ"์ด๋ผ๋Š” ์„ธ๊ทธ๋จผํŠธ ์กฐ๊ฐ์€ ๊ณผ๊ฑฐ ํ™œ๋™์„ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ("๋‚˜๋Š” ํ•œ๋•Œ ํญํƒ„์„ ๋งŒ๋“ค์—ˆ์–ด")๋ถ€ํ„ฐ ํญํƒ„์„ ๋งŒ๋“ค ๊ฒƒ์ธ์ง€์— ๋Œ€ํ•œ ์กฐ์–ธ์„ ๊ตฌํ•˜๋Š” ๊ฒƒ("ํญํƒ„์€ ์–ด๋–ป๊ฒŒ ๋งŒ๋“œ๋‚˜์š”?") ๋˜๋Š” ๊ฒฉ๋ ค๋ฅผ ๊ตฌํ•˜๋Š” ๊ฒƒ("ํญํƒ„์„ ๋งŒ๋“ค์–ด์•ผ ํ•˜๋‚˜์š”?")์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ prompts์— ํ†ตํ•ฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ์–ธ์–ดํ•™ ์ด๋ก , ํŠนํžˆ ํ™”ํ–‰[54]๊ณผ ์ปดํ“จํ„ฐ-์ธ๊ฐ„ ์ƒํ˜ธ์ž‘์šฉ[55]์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์‚ฌ๋žŒ๋“ค์€ ๋””์ง€ํ„ธ ๊ธฐ์ˆ ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•  ๋•Œ ๋‚˜ํƒ€๋‚˜๋Š” ๋‹ค์–‘ํ•œ ์˜์‚ฌ์†Œํ†ต ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๋•Œ๋กœ๋Š” ๋™์ผํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ค๋ฅธ ๋ฌธ๋ฒ• ๊ตฌ์กฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ์†Œ์…œ ๋ฏธ๋””์–ด ์ฝ˜ํ…์ธ , ํ–‰๋™ ๋ฐ ๋Œ€ํ™”์˜ ์ค‘์š”์„ฑ์„ ์ดํ•ดํ•˜๊ณ  [56, 57, 58] ํ–‰๋™ ๋ฐ ๋Œ€ํ™”๋ฅผ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. AI ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด ์ž‘์—…์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ๋ชฉํ‘œ(์˜ˆ: "ํญํƒ„์„ ๋งŒ๋“œ๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์šฐ๊ธฐ")๋ฅผ ๋ช…๋ น์–ด ๋˜๋Š” ์งˆ๋ฌธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง์ ‘ ๋˜๋Š” ๊ฐ„์ ‘์ ์œผ๋กœ ์š”์ฒญํ•˜๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค.
๊ฐ ๋ฌธ์žฅ ์กฐ๊ฐ์„ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํ…œํ”Œ๋ฆฟ์€ ๋ชจ๋ธ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ ์œ ํ˜•์— ๋”ฐ๋ผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ƒํ˜ธ ์ž‘์šฉ์€ ๋‹ค์–‘ํ•œ ๋™๊ธฐ์™€ ๋ชฉํ‘œ๋ฅผ ๋ฐ˜์˜ํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "ํญํƒ„์„ ๋งŒ๋“œ๋Š” ๊ฒƒ"์ด๋ผ๋Š” ์„ธ๊ทธ๋จผํŠธ ์กฐ๊ฐ์€ ๊ณผ๊ฑฐ ํ™œ๋™์„ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ("๋‚˜๋Š” ํ•œ๋•Œ ํญํƒ„์„ ๋งŒ๋“ค์—ˆ์–ด")๋ถ€ํ„ฐ ํญํƒ„์„ ๋งŒ๋“ค ๊ฒƒ์ธ์ง€์— ๋Œ€ํ•œ ์กฐ์–ธ์„ ๊ตฌํ•˜๋Š” ๊ฒƒ("ํญํƒ„์€ ์–ด๋–ป๊ฒŒ ๋งŒ๋“œ๋‚˜์š”?") ๋˜๋Š” ๊ฒฉ๋ ค๋ฅผ ๊ตฌํ•˜๋Š” ๊ฒƒ("ํญํƒ„์„ ๋งŒ๋“ค์–ด์•ผ ํ•˜๋‚˜์š”?")์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ์— ํ†ตํ•ฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ์–ธ์–ดํ•™ ์ด๋ก , ํŠนํžˆ ํ™”ํ–‰[54]๊ณผ ์ปดํ“จํ„ฐ-์ธ๊ฐ„ ์ƒํ˜ธ์ž‘์šฉ[55]์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์‚ฌ๋žŒ๋“ค์€ ๋””์ง€ํ„ธ ๊ธฐ์ˆ ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•  ๋•Œ ๋‚˜ํƒ€๋‚˜๋Š” ๋‹ค์–‘ํ•œ ์˜์‚ฌ์†Œํ†ต ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๋•Œ๋กœ๋Š” ๋™์ผํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ค๋ฅธ ๋ฌธ๋ฒ• ๊ตฌ์กฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ์†Œ์…œ ๋ฏธ๋””์–ด ์ฝ˜ํ…์ธ , ํ–‰๋™ ๋ฐ ๋Œ€ํ™”์˜ ์ค‘์š”์„ฑ์„ ์ดํ•ดํ•˜๊ณ  [56, 57, 58] ํ–‰๋™ ๋ฐ ๋Œ€ํ™”๋ฅผ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. AI ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด ์ž‘์—…์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ๋ชฉํ‘œ(์˜ˆ: "ํญํƒ„์„ ๋งŒ๋“œ๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์šฐ๊ธฐ")๋ฅผ ๋ช…๋ น์–ด ๋˜๋Š” ์งˆ๋ฌธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง์ ‘ ๋˜๋Š” ๊ฐ„์ ‘์ ์œผ๋กœ ์š”์ฒญํ•˜๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค.
We collected our _Evol-Instruct_ testset that includes real-world human instructions from diverse sources such as online opensource projects, platforms, and forums. We analyzed the data and identified 29 distinct skills that represent the main requirements of humanity, such as Coding Generation \(\&\) Debugging, Math, Reasoning, Complex Formats, Writing, Extensive Disciplines, and so on. Figure 3a illustrates the distribution of the instances and skills in our test set. Our test set consists of 218 instances, each of which is an instruction for a specific skill. We compared our test set with Vicuna's test set, which is a benchmark dataset for evaluating instruction following models. We found that Vicuna's test set only 80 instances and 9 skills and is much smaller and less diverse than ours. Figure 3b shows how the difficulty and complexity of the test data vary across different instances. Our test data has a more uniform distribution, meaning that it contains instructions with different levels of difficulty and complexity. On the other hand, Vicuna and Alpaca have a skewed distribution, meaning that they mostly contain instructions with low difficulty and complexity. This indicates that these two corpus are not able to handle the evaluation on more complex and demanding scenarios.
์˜จ๋ผ์ธ ์˜คํ”ˆ์†Œ์Šค ํ”„๋กœ์ ํŠธ, ํ”Œ๋žซํผ, ํฌ๋Ÿผ ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์‹ค์ œ ์ธ๊ฐ„์˜ ์ง€์นจ์„ ํฌํ•จํ•˜๋Š” _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ ์ฝ”๋”ฉ ์ƒ์„ฑ ๋ฐ ๋””๋ฒ„๊น…, ์ˆ˜ํ•™, ์ถ”๋ก , ๋ณต์žกํ•œ ํ˜•์‹, ๊ธ€์“ฐ๊ธฐ, ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ„์•ผ ๋“ฑ ์ธ๋ฅ˜์˜ ์ฃผ์š” ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋Œ€ํ‘œํ•˜๋Š” 29๊ฐœ์˜ ๊ธฐ์ˆ ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3a๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์ธ์Šคํ„ด์Šค์™€ ์Šคํ‚ฌ์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” 218๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ์ธ์Šคํ„ด์Šค๋Š” ํŠน์ • ์Šคํ‚ฌ์— ๋Œ€ํ•œ Instruction์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ Instruction ์ˆ˜ํ–‰ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” 80๊ฐœ์˜ ์ธ์Šคํ„ด์Šค์™€ 9๊ฐœ์˜ ์Šคํ‚ฌ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๊ณ  ๋‹ค์–‘์„ฑ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3b๋Š” ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ๋‚œ์ด๋„์™€ ๋ณต์žก์„ฑ์ด ๋‹ค๋ฅธ ์ธ์Šคํ„ด์Šค์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š”์ง€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋Š” ๋ณด๋‹ค ๊ท ์ผํ•œ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์–ด ๋‚œ์ด๋„์™€ ๋ณต์žก๋„๊ฐ€ ๋‹ค๋ฅธ ์ˆ˜์ค€์˜ Instruction์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Vicuna์™€ Alpaca๋Š” ๋ถ„ํฌ๊ฐ€ ์น˜์šฐ์ณ์ ธ ์žˆ์–ด ๋Œ€๋ถ€๋ถ„ ๋‚œ์ด๋„์™€ ๋ณต์žก๋„๊ฐ€ ๋‚ฎ์€ Instruction์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋‘ ์ฝ”ํผ์Šค๊ฐ€ ๋” ๋ณต์žกํ•˜๊ณ  ๊นŒ๋‹ค๋กœ์šด ์‹œ๋‚˜๋ฆฌ์˜ค์— ๋Œ€ํ•œ ํ‰๊ฐ€๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์˜จ๋ผ์ธ ์˜คํ”ˆ์†Œ์Šค ํ”„๋กœ์ ํŠธ, ํ”Œ๋žซํผ, ํฌ๋Ÿผ ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์‹ค์ œ ์ธ๊ฐ„์˜ ์ง€์นจ์„ ํฌํ•จํ•˜๋Š” _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ ์ฝ”๋”ฉ ์ƒ์„ฑ ๋ฐ ๋””๋ฒ„๊น…, ์ˆ˜ํ•™, ์ถ”๋ก , ๋ณต์žกํ•œ ํ˜•์‹, ๊ธ€์“ฐ๊ธฐ, ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ„์•ผ ๋“ฑ ์ธ๋ฅ˜์˜ ์ฃผ์š” ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋Œ€ํ‘œํ•˜๋Š” 29๊ฐœ์˜ ๊ธฐ์ˆ ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3a๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์ธ์Šคํ„ด์Šค์™€ ์Šคํ‚ฌ์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” 218๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ์ธ์Šคํ„ด์Šค๋Š” ํŠน์ • ์Šคํ‚ฌ์— ๋Œ€ํ•œ ์ง€์นจ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์ง€์นจ ์ˆ˜ํ–‰ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” 80๊ฐœ์˜ ์ธ์Šคํ„ด์Šค์™€ 9๊ฐœ์˜ ์Šคํ‚ฌ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๊ณ  ๋‹ค์–‘์„ฑ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3b๋Š” ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ๋‚œ์ด๋„์™€ ๋ณต์žก์„ฑ์ด ๋‹ค๋ฅธ ์ธ์Šคํ„ด์Šค์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š”์ง€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋Š” ๋ณด๋‹ค ๊ท ์ผํ•œ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์–ด ๋‚œ์ด๋„์™€ ๋ณต์žก๋„๊ฐ€ ๋‹ค๋ฅธ ์ˆ˜์ค€์˜ ์ง€์นจ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Vicuna์™€ Alpaca๋Š” ๋ถ„ํฌ๊ฐ€ ์น˜์šฐ์ณ์ ธ ์žˆ์–ด ๋Œ€๋ถ€๋ถ„ ๋‚œ์ด๋„์™€ ๋ณต์žก๋„๊ฐ€ ๋‚ฎ์€ ์ง€์นจ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋‘ ์ฝ”ํผ์Šค๊ฐ€ ๋” ๋ณต์žกํ•˜๊ณ  ๊นŒ๋‹ค๋กœ์šด ์‹œ๋‚˜๋ฆฌ์˜ค์— ๋Œ€ํ•œ ํ‰๊ฐ€๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
In this work, we propose a novel continual learning (CL) formulation named Continual KnowledgeLearning (CKL), where we attempt to renew the internal world knowledge of LMs through continual pretraining on new corpora. We systematically categorize world knowledge into three main categories and make benchmark datasets to measure each of them during CKL: (1) InvariantLAMA for _time-invariant_ world knowledge in LMs that should not be forgotten or altered, (2) UpdatedLAMA for outdated world knowledge that needs to be _updated_ in the LMs, and (3) NewLAMA for _new_ world knowledge that should be injected into the LMs. We also propose a novel metric named FUAR (Forgotten / (Updated + Acquired) Ratio) that can measure the trade-off between forgetting, updating, and acquiring knowledge. Finally, while one might think of implementing contemporary CL methods for this benchmark, we show that CKL has nontrivial differences to traditional CL formulations and require approaches specific to CKL. We find and compare model architectures and training methodologies (Chen et al., 2020; He et al., 2021; Hu et al., 2021; Wang et al., 2021b) from the literature that have shown potential to mitigate forgetting of knowledge gained during pretraining, establishing them as baselines for the CKL benchmark.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์ „์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•ด LM์˜ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ์ƒˆ๋กญ๊ฒŒ ๊ฐฑ์‹ ํ•˜๋Š” Continual Knowledge Learning (CKL)์ด๋ผ๋Š” ์ƒˆ๋กœ์šด Continual Learning (CL) ๊ณต์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์„ธ๊ณ„ ์ง€์‹์„ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ  ๊ฐ๊ฐ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค: (1) LM์—์„œ ์žŠํ˜€์ง€๊ฑฐ๋‚˜ ๋ณ€๊ฒฝ๋˜์–ด์„œ๋Š” ์•ˆ ๋˜๋Š” _์‹œ๊ฐ„ ๋ถˆ๋ณ€_ ์„ธ๊ณ„ ์ง€์‹์„ ์œ„ํ•œ InvariantLAMA, (2) LM์—์„œ _์—…๋ฐ์ดํŠธ_ํ•ด์•ผ ํ•˜๋Š” ์˜ค๋ž˜๋œ ์„ธ๊ณ„ ์ง€์‹์„ ์œ„ํ•œ UpdatedLAMA, (3) LM์— ์ฃผ์ž…๋˜์–ด์•ผ ํ•˜๋Š” _์ƒˆ๋กœ์šด_ ์„ธ๊ณ„ ์ง€์‹์„ ์œ„ํ•œ NewLAMA. ๋˜ํ•œ ์žŠ์–ด๋ฒ„๋ฆผ, ์—…๋ฐ์ดํŠธ, ์Šต๋“ ์ง€์‹ ๊ฐ„์˜ ์ ˆ์ถฉ์ ์„ ์ธก์ •ํ•  ์ˆ˜ ์žˆ๋Š” FUAR (Forgotten / (Updated + Acquired) Ratio)๋ผ๋Š” ์ƒˆ๋กœ์šด ์ง€ํ‘œ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด ๋ฒค์น˜๋งˆํฌ๋ฅผ ์œ„ํ•ด ํ˜„๋Œ€์ ์ธ CL ๋ฐฉ๋ฒ•์„ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์„ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, CKL์€ ๊ธฐ์กด CL ๊ณต์‹๊ณผ ๋น„๊ตํ•  ๋•Œ ์‚ฌ์†Œํ•œ ์ฐจ์ด๊ฐ€ ์žˆ์œผ๋ฉฐ CKL์— ํŠนํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์Šต๋“ํ•œ ์ง€์‹์˜ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ค€ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก (Chen et al., 2020; He et al., 2021; Hu et al., 2021; Wang et al., 2021b)์„ ๋ฌธํ—Œ์—์„œ ์ฐพ์•„ ๋น„๊ตํ•˜์—ฌ CKL ๋ฒค์น˜๋งˆํฌ์˜ ๊ธฐ์ค€์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์ „์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•ด LM์˜ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ์ƒˆ๋กญ๊ฒŒ ๊ฐฑ์‹ ํ•˜๋Š” ์ง€์†์ ์ธ ์ง€์‹ ํ•™์Šต(CKL)์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ์ง€์† ํ•™์Šต(CL) ๊ณต์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์„ธ๊ณ„ ์ง€์‹์„ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ  ๊ฐ๊ฐ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค: (1) LM์—์„œ ์žŠํ˜€์ง€๊ฑฐ๋‚˜ ๋ณ€๊ฒฝ๋˜์–ด์„œ๋Š” ์•ˆ ๋˜๋Š” _์‹œ๊ฐ„ ๋ถˆ๋ณ€_ ์„ธ๊ณ„ ์ง€์‹์„ ์œ„ํ•œ InvariantLAMA, (2) LM์—์„œ _์—…๋ฐ์ดํŠธ_ํ•ด์•ผ ํ•˜๋Š” ์˜ค๋ž˜๋œ ์„ธ๊ณ„ ์ง€์‹์„ ์œ„ํ•œ UpdatedLAMA, (3) LM์— ์ฃผ์ž…๋˜์–ด์•ผ ํ•˜๋Š” _์ƒˆ๋กœ์šด_ ์„ธ๊ณ„ ์ง€์‹์„ ์œ„ํ•œ NewLAMA. ๋˜ํ•œ ์žŠ์–ด๋ฒ„๋ฆผ, ์—…๋ฐ์ดํŠธ, ์Šต๋“ ์ง€์‹ ๊ฐ„์˜ ์ ˆ์ถฉ์ ์„ ์ธก์ •ํ•  ์ˆ˜ ์žˆ๋Š” FUAR(์žŠ์–ด๋ฒ„๋ฆผ / (์—…๋ฐ์ดํŠธ + ์Šต๋“) ๋น„์œจ)๋ผ๋Š” ์ƒˆ๋กœ์šด ์ง€ํ‘œ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด ๋ฒค์น˜๋งˆํฌ๋ฅผ ์œ„ํ•ด ํ˜„๋Œ€์ ์ธ CL ๋ฐฉ๋ฒ•์„ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์„ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, CKL์€ ๊ธฐ์กด CL ๊ณต์‹๊ณผ ๋น„๊ตํ•  ๋•Œ ์‚ฌ์†Œํ•œ ์ฐจ์ด๊ฐ€ ์žˆ์œผ๋ฉฐ CKL์— ํŠนํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์Šต๋“ํ•œ ์ง€์‹์˜ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ค€ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก (Chen et al., 2020; He et al., 2021; Hu et al., 2021; Wang et al., 2021b)์„ ๋ฌธํ—Œ์—์„œ ์ฐพ์•„ ๋น„๊ตํ•˜์—ฌ CKL ๋ฒค์น˜๋งˆํฌ์˜ ๊ธฐ์ค€์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
Recent studies have proposed a mixture of synthetic data and real human data to train more robust reward models (Gao et al., 2023). Constitutional AI (Bai et al., 2022) proposes to use a small set of principles to steer the AI generated critiques and feedback, and use such synthetic data to replace the real human data in the typical RLHF pipeline. The model trained with this RLAIF (i.e., reinforcement learning from AI feedback) method shows similar strong performance as RLHF baselines. In general, synthetic data offers a powerful solution for human values and preferences alignment by allowing researchers to generate large-scale, diverse, and controlled training datasets in a low-cost way (Cui et al., 2023; Ganguli et al., 2022). By simulating a wide range of scenarios involving ethical dilemmas (Perez et al., 2022), social interactions (Liu et al., 2023c), and cultural norms (Ziems et al., 2023), synthetic data enables comprehensive and systematic testing of AI models' alignment with human values (Askell et al., 2021). This approach helps identify and mitigate issues related to bias (Liu et al., 2021; Ntoutsi et al., 2020), fairness (Landers and Behrend, 2023; Zhao et al., 2018), and unintended consequences before AI systems are deployed in real-world settings (Ye et al., 2024).
์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2023). Constitutional AI(Bai et al., 2022)๋Š” ์†Œ์ˆ˜์˜ ์›์น™์„ ์‚ฌ์šฉํ•˜์—ฌ AI๊ฐ€ ์ƒ์„ฑํ•œ ๋น„ํ‰๊ณผ ํ”ผ๋“œ๋ฐฑ์„ ์œ ๋„ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ฐ˜์ ์ธ RLHF ํŒŒ์ดํ”„๋ผ์ธ์—์„œ ์‹ค์ œ ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์ฒดํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด RLAIF (i.e., Reinforcement Learning from AI Feedback) ๋ฐฉ๋ฒ•์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ RLHF ๊ธฐ์ค€์„ ๊ณผ ๋น„์Šทํ•œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ €๋น„์šฉ์œผ๋กœ ๋Œ€๊ทœ๋ชจ์˜ ๋‹ค์–‘ํ•˜๊ณ  ํ†ต์ œ๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„๋ฅผ ์ •๋ ฌํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(Cui et al., 2023; Ganguli et al., 2022). ์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ(Perez et al., 2022), ์‚ฌํšŒ์  ์ƒํ˜ธ์ž‘์šฉ(Liu et al., 2023c), ๋ฌธํ™”์  ๊ทœ๋ฒ”(Ziems et al., 2023)๊ณผ ๊ด€๋ จ๋œ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•จ์œผ๋กœ์จ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” AI ๋ชจ๋ธ์ด ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ์น˜ํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์ฒด๊ณ„์ ์ธ ํ…Œ์ŠคํŠธ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค(Askell et al., 2021). ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ AI ์‹œ์Šคํ…œ์ด ์‹ค์ œ ํ™˜๊ฒฝ์— ๋ฐฐํฌ๋˜๊ธฐ ์ „์— ํŽธํ–ฅ์„ฑ(Liu et al., 2021; Ntoutsi et al., 2020), ๊ณต์ •์„ฑ(Landers and Behrend, 2023; Zhao et al., 2018) ๋ฐ ์˜๋„ํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ์™€ ๊ด€๋ จ๋œ ๋ฌธ์ œ๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค(Ye et al., 2024).
์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2023). ํ—Œ๋ฒ• AI(Bai et al., 2022)๋Š” ์†Œ์ˆ˜์˜ ์›์น™์„ ์‚ฌ์šฉํ•˜์—ฌ AI๊ฐ€ ์ƒ์„ฑํ•œ ๋น„ํ‰๊ณผ ํ”ผ๋“œ๋ฐฑ์„ ์œ ๋„ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ฐ˜์ ์ธ RLHF ํŒŒ์ดํ”„๋ผ์ธ์—์„œ ์‹ค์ œ ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์ฒดํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด RLAIF(๊ฐ•ํ™” ํ•™์Šต์—์„œ AI ํ”ผ๋“œ๋ฐฑ) ๋ฐฉ๋ฒ•์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ RLHF ๊ธฐ์ค€์„ ๊ณผ ๋น„์Šทํ•œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ €๋น„์šฉ์œผ๋กœ ๋Œ€๊ทœ๋ชจ์˜ ๋‹ค์–‘ํ•˜๊ณ  ํ†ต์ œ๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„๋ฅผ ์ •๋ ฌํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(Cui et al., 2023; Ganguli et al., 2022). ์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ(Perez et al., 2022), ์‚ฌํšŒ์  ์ƒํ˜ธ์ž‘์šฉ(Liu et al., 2023c), ๋ฌธํ™”์  ๊ทœ๋ฒ”(Ziems et al., 2023)๊ณผ ๊ด€๋ จ๋œ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•จ์œผ๋กœ์จ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” AI ๋ชจ๋ธ์ด ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ์น˜ํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์ฒด๊ณ„์ ์ธ ํ…Œ์ŠคํŠธ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค(Askell et al., 2021). ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ AI ์‹œ์Šคํ…œ์ด ์‹ค์ œ ํ™˜๊ฒฝ์— ๋ฐฐํฌ๋˜๊ธฐ ์ „์— ํŽธํ–ฅ์„ฑ(Liu et al., 2021; Ntoutsi et al., 2020), ๊ณต์ •์„ฑ(Landers and Behrend, 2023; Zhao et al., 2018) ๋ฐ ์˜๋„ํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ์™€ ๊ด€๋ จ๋œ ๋ฌธ์ œ๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค(Ye et al., 2024).
Large language models (LLMs) mainly refer to transformer-based neural language models 1 that contain tens to hundreds of billions of parameters, which are pre-trained on massive text data, such as PaLM [31], LLaMA [32], and GPT-4 [33], as summarized in Table III. Comparedto PLMs, LLMs are not only much larger in model size, but also exhibit stronger language understanding and generation abilities, and more importantly, emergent abilities that are not present in smaller-scale language models. As illustrated in Fig. 1, these emergent abilities include (1) in-context learning, where LLMs learn a new task from a small set of examples presented in the prompt at inference time, (2) instruction following, where LLMs, after instruction tuning, can follow the instructions for new types of tasks without using explicit examples, and (3) multi-step reasoning, where LLMs can solve a complex task by breaking down that task into intermediate reasoning steps as demonstrated in the chain-of-thought prompt [34]. LLMs can also be augmented by using external knowledge and tools [35, 36] so that they can effectively interact with users and environment [37], and continually improve itself using feedback data collected through interactions (e.g. via reinforcement learning with human feedback (RLHF)).
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ฃผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ˆ˜์‹ญ์—์„œ ์ˆ˜๋ฐฑ์–ต ๊ฐœ์— ๋‹ฌํ•˜๋Š” Transformer ๊ธฐ๋ฐ˜ ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ์„ ๋งํ•˜๋ฉฐ, ํ‘œ III์— ์š”์•ฝ๋œ ๊ฒƒ์ฒ˜๋Ÿผ PaLM [31], LLaMA [32], GPT-4 [33]์™€ ๊ฐ™์ด ๋ฐฉ๋Œ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ ๋ชจ๋ธ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. LLM์€ PLM์— ๋น„ํ•ด ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ํ›จ์”ฌ ํฌ๊ธฐ๋Š” ๋ฌผ๋ก  ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—๋Š” ์—†๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์—๋Š” (1) ์ถ”๋ก  ์‹œ๊ฐ„์— ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์†Œ์ˆ˜์˜ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•˜๋Š” In-Context Learning, (2) ๋ช…์‹œ์ ์ธ ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ Instruction Tuning์„ ๊ฑฐ์นœ ํ›„ ์ƒˆ๋กœ์šด ์œ ํ˜•์˜ ์ž‘์—…์— ๋Œ€ํ•œ ๋ช…๋ น์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋Š” Instruction Following, (3) ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ [34]์—์„œ์ฒ˜๋Ÿผ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์™ธ๋ถ€ ์ง€์‹๊ณผ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฆ๊ฐ•ํ•  ์ˆ˜๋„ ์žˆ์œผ๋ฏ€๋กœ ์‚ฌ์šฉ์ž ๋ฐ ํ™˜๊ฒฝ๊ณผ ํšจ๊ณผ์ ์œผ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ [35, 36] ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ (์˜ˆ: Reinforcement Learning from Human Feedback (RLHF)์„ ์ด์šฉํ•œ ๊ฐ•ํ™” ํ•™์Šต) ์ง€์†์ ์œผ๋กœ ์ž์‹ ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ฃผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ˆ˜์‹ญ์—์„œ ์ˆ˜๋ฐฑ์–ต ๊ฐœ์— ๋‹ฌํ•˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ์„ ๋งํ•˜๋ฉฐ, ํ‘œ III์— ์š”์•ฝ๋œ ๊ฒƒ์ฒ˜๋Ÿผ PaLM [31], LLaMA [32], GPT-4 [33]์™€ ๊ฐ™์ด ๋ฐฉ๋Œ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ ๋ชจ๋ธ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. LLM์€ PLM์— ๋น„ํ•ด ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ํ›จ์”ฌ ํฌ๊ธฐ๋Š” ๋ฌผ๋ก  ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—๋Š” ์—†๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์—๋Š” (1) ์ถ”๋ก  ์‹œ๊ฐ„์— ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์†Œ์ˆ˜์˜ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•˜๋Š” ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต, (2) ๋ช…์‹œ์ ์ธ ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ ๋ช…๋ น์–ด ํŠœ๋‹์„ ๊ฑฐ์นœ ํ›„ ์ƒˆ๋กœ์šด ์œ ํ˜•์˜ ์ž‘์—…์— ๋Œ€ํ•œ ๋ช…๋ น์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋Š” ๋ช…๋ น์–ด ์ˆ˜ํ–‰, (3) ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ [34]์—์„œ์ฒ˜๋Ÿผ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์™ธ๋ถ€ ์ง€์‹๊ณผ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฆ๊ฐ•ํ•  ์ˆ˜๋„ ์žˆ์œผ๋ฏ€๋กœ ์‚ฌ์šฉ์ž ๋ฐ ํ™˜๊ฒฝ๊ณผ ํšจ๊ณผ์ ์œผ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ [35, 36] ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ (์˜ˆ: ์‚ฌ๋žŒ๊ณผ์˜ ํ”ผ๋“œ๋ฐฑ(RLHF)์„ ์ด์šฉํ•œ ๊ฐ•ํ™” ํ•™์Šต) ์ง€์†์ ์œผ๋กœ ์ž์‹ ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Large unsupervised language models (LMs) trained on very large datasets acquire surprising capabilities [11; 7; 40; 8]. However, these models are trained on data generated by humans with a wide variety of goals, priorities, and skillsets. Some of these goals and skillsets may not be desirable to imitate; for example, while we may want our AI coding assistant to _understand_ common programming mistakes in order to correct them, nevertheless, when generating code, we would like to bias our model toward the (potentially rare) high-quality coding ability present in its training data. Similarly, we might want our language model to be _aware_ of a common misconception believed by 50% of people, but we certainly do not want the model to claim this misconception to be true in 50% of queries about it! In other words, selecting the model's _desired responses and behavior_ from its very wide _knowledge and abilities_ is crucial to building AI systems that are safe, performant, and controllable [26]. While existing methods typically steer LMs to match human preferences using reinforcement learning (RL),we will show that the RL-based objective used by existing methods can be optimized exactly with a simple binary cross-entropy objective, greatly simplifying the preference learning pipeline.
๋งค์šฐ ๋ฐฉ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ ์–ธ์–ด ๋ชจ๋ธ(LMs)์€ ๋†€๋ผ์šด ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค[11; 7; 40; 8]. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ๋ชฉํ‘œ, ์šฐ์„ ์ˆœ์œ„ ๋ฐ ์Šคํ‚ฌ์…‹์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ๋“ค์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ์™€ ์Šคํ‚ฌ์…‹ ์ค‘ ์ผ๋ถ€๋Š” ๋ชจ๋ฐฉํ•˜๊ธฐ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, AI ์ฝ”๋”ฉ ๋„์šฐ๋ฏธ๊ฐ€ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์‹ค์ˆ˜๋ฅผ ๋ฐ”๋กœ์žก๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์‹ค์ˆ˜๋ฅผ _์ดํ•ด_ํ•˜๊ธฐ๋ฅผ ์›ํ•  ์ˆ˜๋Š” ์žˆ์ง€๋งŒ, ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ์กด์žฌํ•˜๋Š” (์ž ์žฌ์ ์œผ๋กœ ๋“œ๋ฌธ) ๊ณ ํ’ˆ์งˆ ์ฝ”๋”ฉ ๋Šฅ๋ ฅ์— ๋ชจ๋ธ์„ ํŽธํ–ฅ์‹œํ‚ค๊ณ  ์‹ถ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ๋žŒ๋“ค์˜ 50%๊ฐ€ ๋ฏฟ๊ณ  ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ์˜คํ•ด๋ฅผ _์ธ์‹_ํ•˜๊ธฐ๋ฅผ ์›ํ•  ์ˆ˜๋„ ์žˆ์ง€๋งŒ, ๋ชจ๋ธ์ด ์ด ์˜คํ•ด๋ฅผ 50%์˜ ์ฟผ๋ฆฌ์—์„œ ์‚ฌ์‹ค์ด๋ผ๊ณ  ์ฃผ์žฅํ•˜๋Š” ๊ฒƒ์€ ๊ฒฐ์ฝ” ์›ํ•˜์ง€ ์•Š์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค! ์ฆ‰, ๋งค์šฐ ๋„“์€ _์ง€์‹๊ณผ ๋Šฅ๋ ฅ_์—์„œ ๋ชจ๋ธ์˜ _์›ํ•˜๋Š” ์‘๋‹ต๊ณผ ํ–‰๋™_์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์•ˆ์ „ํ•˜๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ ์ œ์–ด ๊ฐ€๋Šฅํ•œ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[26]. ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ Reinforcement Learning (RL)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜์ง€๋งŒ, ๊ธฐ์กด ๋ฐฉ๋ฒ•์—์„œ ์‚ฌ์šฉ๋˜๋Š” RL ๊ธฐ๋ฐ˜ Objective๋Š” ๊ฐ„๋‹จํ•œ ์ด์ง„ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ Objective๋กœ ์ •ํ™•ํ•˜๊ฒŒ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์„ ํ˜ธ๋„ ํ•™์Šต ํŒŒ์ดํ”„๋ผ์ธ์„ ํฌ๊ฒŒ ๋‹จ์ˆœํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋งค์šฐ ๋ฐฉ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ ์–ธ์–ด ๋ชจ๋ธ(LM)์€ ๋†€๋ผ์šด ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค[11; 7; 40; 8]. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ๋ชฉํ‘œ, ์šฐ์„ ์ˆœ์œ„ ๋ฐ ์Šคํ‚ฌ์…‹์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ๋“ค์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ์™€ ์Šคํ‚ฌ์…‹ ์ค‘ ์ผ๋ถ€๋Š” ๋ชจ๋ฐฉํ•˜๊ธฐ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, AI ์ฝ”๋”ฉ ๋„์šฐ๋ฏธ๊ฐ€ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์‹ค์ˆ˜๋ฅผ ๋ฐ”๋กœ์žก๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์‹ค์ˆ˜๋ฅผ _์ดํ•ด_ํ•˜๊ธฐ๋ฅผ ์›ํ•  ์ˆ˜๋Š” ์žˆ์ง€๋งŒ, ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ์กด์žฌํ•˜๋Š” (์ž ์žฌ์ ์œผ๋กœ ๋“œ๋ฌธ) ๊ณ ํ’ˆ์งˆ ์ฝ”๋”ฉ ๋Šฅ๋ ฅ์— ๋ชจ๋ธ์„ ํŽธํ–ฅ์‹œํ‚ค๊ณ  ์‹ถ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ๋žŒ๋“ค์˜ 50%๊ฐ€ ๋ฏฟ๊ณ  ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ์˜คํ•ด๋ฅผ _์ธ์‹_ํ•˜๊ธฐ๋ฅผ ์›ํ•  ์ˆ˜๋„ ์žˆ์ง€๋งŒ, ๋ชจ๋ธ์ด ์ด ์˜คํ•ด๋ฅผ 50%์˜ ์ฟผ๋ฆฌ์—์„œ ์‚ฌ์‹ค์ด๋ผ๊ณ  ์ฃผ์žฅํ•˜๋Š” ๊ฒƒ์€ ๊ฒฐ์ฝ” ์›ํ•˜์ง€ ์•Š์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค! ์ฆ‰, ๋งค์šฐ ๋„“์€ _์ง€์‹๊ณผ ๋Šฅ๋ ฅ_์—์„œ ๋ชจ๋ธ์˜ _์›ํ•˜๋Š” ์‘๋‹ต๊ณผ ํ–‰๋™_์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์•ˆ์ „ํ•˜๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ ์ œ์–ด ๊ฐ€๋Šฅํ•œ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[26]. ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ•ํ™” ํ•™์Šต(RL)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜์ง€๋งŒ, ๊ธฐ์กด ๋ฐฉ๋ฒ•์—์„œ ์‚ฌ์šฉ๋˜๋Š” RL ๊ธฐ๋ฐ˜ ๋ชฉํ‘œ๋Š” ๊ฐ„๋‹จํ•œ ์ด์ง„ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ๋ชฉํ‘œ๋กœ ์ •ํ™•ํ•˜๊ฒŒ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์„ ํ˜ธ๋„ ํ•™์Šต ํŒŒ์ดํ”„๋ผ์ธ์„ ํฌ๊ฒŒ ๋‹จ์ˆœํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Reward model based approaches._ The reward model in RLHF has been trained to measure the alignment degree on the responses of LLMs. It is straightforward to leverage existing reward models to select high-quality responses as alignment data for subsequent fine-tuning. Based on this idea, RAFT [373] adopts reward models trained on human preference data to rank the responses of LLMs and collect those with higher rewards for supervised fine-tuning. In addition, the reward model can be also used to score model responses and assign them to different quality groups. Quark [384] sorts the responses of LLMs into different quantiles based on the reward scores. Each quantile is attached with a special reward token to represent the reward level of the quantile. Conditioned on the highest-reward tokens, LLMs are subsequently prompted to generate high-quality responses. Given an initial answer and the corresponding human feedback, ILF [385] first adopts LLMs to generate refined answers, then utilizes the reward model to select the answer that best matches the feedback for further training. As valuable resources for aligning LLMs, several reward models have been released, including DeBERTabase/large/xlarge from OpenAssistant27, Moss-7B from Fudan28, and Flan-T5-xl from Stanford29.
\(\bullet\)_๋ณด์ƒ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹._ RLHF์˜ ๋ณด์ƒ ๋ชจ๋ธ์€ LLM์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์ •๋ ฌ ์ •๋„๋ฅผ ์ธก์ •ํ•˜๋„๋ก ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ณด์ƒ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ์‘๋‹ต์„ ์„ ๋ณ„ํ•˜์—ฌ ํ›„์† Fine-tuning์„ ์œ„ํ•œ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ๊ฐ„๋‹จํ•ฉ๋‹ˆ๋‹ค. ์ด ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ RAFT [373]๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ฑ„ํƒํ•˜์—ฌ LLM์˜ ์‘๋‹ต์„ ์ˆœ์œ„ํ™”ํ•˜๊ณ  ๋ณด์ƒ์ด ๋” ๋†’์€ ์‘๋‹ต์„ ์ˆ˜์ง‘ํ•˜์—ฌ Supervised Fine-tuning์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ ์‘๋‹ต์— ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๊ณ  ์—ฌ๋Ÿฌ ํ’ˆ์งˆ ๊ทธ๋ฃน์— ํ• ๋‹นํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. Quark [384]๋Š” ๋ณด์ƒ ์ ์ˆ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์˜ ์‘๋‹ต์„ ์—ฌ๋Ÿฌ ๋ถ„์œ„์ˆ˜๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ถ„์œ„์ˆ˜์—๋Š” ํ•ด๋‹น ๋ถ„์œ„์ˆ˜์˜ ๋ณด์ƒ ์ˆ˜์ค€์„ ๋‚˜ํƒ€๋‚ด๋Š” ํŠน์ˆ˜ ๋ณด์ƒ ํ† ํฐ์ด ์ฒจ๋ถ€๋ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๋†’์€ ๋ณด์ƒ ํ† ํฐ์„ ์กฐ๊ฑด์œผ๋กœ LLM์€ ํ›„์†์ ์œผ๋กœ ๊ณ ํ’ˆ์งˆ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฉ๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ๋‹ต๋ณ€๊ณผ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” ์‚ฌ๋žŒ์˜ ํ”ผ๋“œ๋ฐฑ์ด ์ฃผ์–ด์ง€๋ฉด ILF [385]๋Š” ๋จผ์ € LLM์„ ์ฑ„ํƒํ•˜์—ฌ ์ •์ œ๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐ€์žฅ ์ž˜ ์ผ์น˜ํ•˜๋Š” ๋‹ต๋ณ€์„ ์„ ํƒํ•˜์—ฌ ์ถ”๊ฐ€ ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•œ ๊ท€์ค‘ํ•œ ๋ฆฌ์†Œ์Šค๋กœ์„œ OpenAssistant27์˜ DeBERTabase/large/xlarge, Fudan28์˜ Moss-7B, Stanford29์˜ Flan-T5-xl ๋“ฑ ์—ฌ๋Ÿฌ ๋ณด์ƒ ๋ชจ๋ธ์ด ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_๋ณด์ƒ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹._ RLHF์˜ ๋ณด์ƒ ๋ชจ๋ธ์€ LLM์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์ •๋ ฌ ์ •๋„๋ฅผ ์ธก์ •ํ•˜๋„๋ก ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ณด์ƒ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ์‘๋‹ต์„ ์„ ๋ณ„ํ•˜์—ฌ ํ›„์† ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•œ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ๊ฐ„๋‹จํ•ฉ๋‹ˆ๋‹ค. ์ด ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ RAFT [373]๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ฑ„ํƒํ•˜์—ฌ LLM์˜ ์‘๋‹ต์„ ์ˆœ์œ„ํ™”ํ•˜๊ณ  ๋ณด์ƒ์ด ๋” ๋†’์€ ์‘๋‹ต์„ ์ˆ˜์ง‘ํ•˜์—ฌ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ ์‘๋‹ต์— ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๊ณ  ์—ฌ๋Ÿฌ ํ’ˆ์งˆ ๊ทธ๋ฃน์— ํ• ๋‹นํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. Quark [384]๋Š” ๋ณด์ƒ ์ ์ˆ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์˜ ์‘๋‹ต์„ ์—ฌ๋Ÿฌ ๋ถ„์œ„์ˆ˜๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ถ„์œ„์ˆ˜์—๋Š” ํ•ด๋‹น ๋ถ„์œ„์ˆ˜์˜ ๋ณด์ƒ ์ˆ˜์ค€์„ ๋‚˜ํƒ€๋‚ด๋Š” ํŠน์ˆ˜ ๋ณด์ƒ ํ† ํฐ์ด ์ฒจ๋ถ€๋ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๋†’์€ ๋ณด์ƒ ํ† ํฐ์„ ์กฐ๊ฑด์œผ๋กœ LLM์€ ํ›„์†์ ์œผ๋กœ ๊ณ ํ’ˆ์งˆ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฉ๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ๋‹ต๋ณ€๊ณผ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” ์‚ฌ๋žŒ์˜ ํ”ผ๋“œ๋ฐฑ์ด ์ฃผ์–ด์ง€๋ฉด ILF [385]๋Š” ๋จผ์ € LLM์„ ์ฑ„ํƒํ•˜์—ฌ ์ •์ œ๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐ€์žฅ ์ž˜ ์ผ์น˜ํ•˜๋Š” ๋‹ต๋ณ€์„ ์„ ํƒํ•˜์—ฌ ์ถ”๊ฐ€ ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•œ ๊ท€์ค‘ํ•œ ๋ฆฌ์†Œ์Šค๋กœ์„œ OpenAssistant27์˜ DeBERTabase/large/xlarge, Fudan28์˜ Moss-7B, Stanford29์˜ Flan-T5-xl ๋“ฑ ์—ฌ๋Ÿฌ ๋ณด์ƒ ๋ชจ๋ธ์ด ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Several related works apply continual pre-training to specific tasks and domains (Sun et al., 2020; Jang et al., 2022; 2022; Gong et al., 2022; Zan et al., 2022; Yadav et al., 2023; Ma et al., 2023; Yang et al., 2024). While these works also utilize continual pre-training techniques, they differ from our work by focusing on particular domains instead of general pre-training techniques, on smaller-scale datasets \(<\) 10B tokens with smaller models. The only existing work that approaches our dataset scale is (Gogoulou et al., 2023), which explores continual autoregressive language modeling across English, Danish, Icelandic, and Norwegian datasets (73B each). While they do not use replay they do re-warm and re-decay the learning rate. The only existing work that approaches our model scale is (Yang et al., 2024). They continually pre-train and instruction tune LLaMA2 on small-scale academic plant science data. This concurrent work uses a very similar continual learning setup to the one we propose: replay, LR re-warming, and LR re-decaying. While, unlike our work, they _do not_ build a controlled experimental framework to systematically evaluate the validity of these approaches for continual pre-training, it is nice to see further experimental evidence validating our approach.
๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ์—ฐ๊ตฌ์—์„œ๋Š” ํŠน์ • ์ž‘์—…๊ณผ ๋„๋ฉ”์ธ์— ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค(Sun et al., 2020; Jang et al., 2022; 2022; Gong et al., 2022; Zan et al., 2022; Yadav et al., 2023; Ma et al., 2023; Yang et al., 2024). ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ๋„ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๊ธฐ๋ฒ•์ด ์•„๋‹Œ ํŠน์ • ๋„๋ฉ”์ธ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ , ๋” ์ž‘์€ ๊ทœ๋ชจ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ(<10B ํ† ํฐ)์™€ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์—์„œ ์ €ํฌ ์ž‘์—…๊ณผ๋Š” ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ทœ๋ชจ์— ์ ‘๊ทผํ•˜๋Š” ์œ ์ผํ•œ ๊ธฐ์กด ์ž‘์—…์€ ์˜์–ด, ๋ด๋งˆํฌ์–ด, ์•„์ด์Šฌ๋ž€๋“œ์–ด, ๋…ธ๋ฅด์›จ์ด์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(๊ฐ๊ฐ 73B)์— ๊ฑธ์ณ ์ง€์†์ ์ธ ์ž๋™ ํšŒ๊ท€ ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ํƒ๊ตฌํ•˜๋Š” (Gogoulou et al., 2023)์ž…๋‹ˆ๋‹ค. ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋Š” ์•Š์ง€๋งŒ Learning Rate๋ฅผ ๋‹ค์‹œ ์˜ˆ์—ดํ•˜๊ณ  ๋‹ค์‹œ ๊ฐ์†Œ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ์— ์ ‘๊ทผํ•˜๋Š” ์œ ์ผํ•œ ๊ธฐ์กด ์ž‘์—…์€ (Yang et al., 2024)์ž…๋‹ˆ๋‹ค. ์ด๋“ค์€ ์†Œ๊ทœ๋ชจ ํ•™์ˆ  ์‹๋ฌผ ๊ณผํ•™ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLaMA2๋ฅผ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ณ  Instruction์„ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋™์‹œ ์ž‘์—…์€ ์šฐ๋ฆฌ๊ฐ€ ์ œ์•ˆํ•˜๋Š” ๊ฒƒ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•œ ์ง€์† ํ•™์Šต ์„ค์ •, ์ฆ‰ ๋ฆฌํ”Œ๋ ˆ์ด, LR ์žฌ๊ฐ€์—ด ๋ฐ LR ์žฌ๊ฐ์‡ ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์— ๋Œ€ํ•œ ํƒ€๋‹น์„ฑ์„ ์ฒด๊ณ„์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ํ†ต์ œ๋œ ์‹คํ—˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ, ์šฐ๋ฆฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒ€์ฆํ•˜๋Š” ์ถ”๊ฐ€ ์‹คํ—˜์  ์ฆ๊ฑฐ๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ์–ด ์ข‹์Šต๋‹ˆ๋‹ค.
๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ์—ฐ๊ตฌ์—์„œ๋Š” ํŠน์ • ์ž‘์—…๊ณผ ๋„๋ฉ”์ธ์— ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค(Sun et al., 2020; Jang et al., 2022; 2022; Gong et al., 2022; Zan et al., 2022; Yadav et al., 2023; Ma et al., 2023; Yang et al., 2024). ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ๋„ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๊ธฐ๋ฒ•์ด ์•„๋‹Œ ํŠน์ • ๋„๋ฉ”์ธ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ , ๋” ์ž‘์€ ๊ทœ๋ชจ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ(<10B ํ† ํฐ)์™€ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์—์„œ ์ €ํฌ ์ž‘์—…๊ณผ๋Š” ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ทœ๋ชจ์— ์ ‘๊ทผํ•˜๋Š” ์œ ์ผํ•œ ๊ธฐ์กด ์ž‘์—…์€ ์˜์–ด, ๋ด๋งˆํฌ์–ด, ์•„์ด์Šฌ๋ž€๋“œ์–ด, ๋…ธ๋ฅด์›จ์ด์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(๊ฐ๊ฐ 73B)์— ๊ฑธ์ณ ์ง€์†์ ์ธ ์ž๋™ ํšŒ๊ท€ ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ํƒ๊ตฌํ•˜๋Š” (Gogoulou et al., 2023)์ž…๋‹ˆ๋‹ค. ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋Š” ์•Š์ง€๋งŒ ํ•™์Šต ์†๋„๋ฅผ ๋‹ค์‹œ ์˜ˆ์—ดํ•˜๊ณ  ๋‹ค์‹œ ๊ฐ์†Œ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ์— ์ ‘๊ทผํ•˜๋Š” ์œ ์ผํ•œ ๊ธฐ์กด ์ž‘์—…์€ (Yang et al., 2024)์ž…๋‹ˆ๋‹ค. ์ด๋“ค์€ ์†Œ๊ทœ๋ชจ ํ•™์ˆ  ์‹๋ฌผ ๊ณผํ•™ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLaMA2๋ฅผ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ณ  ์ง€์นจ์„ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋™์‹œ ์ž‘์—…์€ ์šฐ๋ฆฌ๊ฐ€ ์ œ์•ˆํ•˜๋Š” ๊ฒƒ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•œ ์ง€์† ํ•™์Šต ์„ค์ •, ์ฆ‰ ๋ฆฌํ”Œ๋ ˆ์ด, LR ์žฌ๊ฐ€์—ด ๋ฐ LR ์žฌ๊ฐ์‡ ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์— ๋Œ€ํ•œ ํƒ€๋‹น์„ฑ์„ ์ฒด๊ณ„์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ํ†ต์ œ๋œ ์‹คํ—˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ, ์šฐ๋ฆฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒ€์ฆํ•˜๋Š” ์ถ”๊ฐ€ ์‹คํ—˜์  ์ฆ๊ฑฐ๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ์–ด ์ข‹์Šต๋‹ˆ๋‹ค.
Generating multilingual questions and answers at scale.Recent studies explore the generation and utilization of synthetic multilingual question-answer (QA) pairs to improve language models'performance in multilingual and cross-lingual question answering (Abulkhanov et al., 2023; Asai et al., 2021; Chi et al., 2020; Kumar et al., 2019; Li and Callison-Burch, 2023; Riabi et al., 2021). One approach is to translate existing monolingual questions and/or answers into other languages (Asai et al., 2021). Another involves using Question Generation (QG) models to produce synthetic questions in a cross-lingual fashion based on answers and/or source texts (Chi et al., 2020; Kumar et al., 2019; Riabi et al., 2021). Recent efforts also focus on jointly generating questions and answers in multiple languages for greater flexibility (Li and Callison-Burch, 2023; Shakeri et al., 2021). For example, Shakeri et al. (2021) finetune a pretrained multilingual T5 model (Xue et al., 2020) on a mixture of a QA generation task and a multilingual masked language modeling task to produce synthetic QA pairs in multiple languages. These efforts generally show that language models trained on synthetic QA pairs demonstrate improved performance on multilingual QA and information retrieval benchmarks.
๋‹ค๊ตญ์–ด ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค๊ตญ์–ด ๋ฐ ๋‹ค๊ตญ์–ด ์งˆ๋ฌธ ๋‹ต๋ณ€(Abulkhanov et al., 2023; Asai et al., 2021; Chi et al., 2020; Kumar et al., 2019; Li and Callison-Burch, 2023; Riabi et al., 2021)์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋‹ค๊ตญ์–ด ์งˆ๋ฌธ ๋‹ต๋ณ€(QA) ์Œ์˜ ์ƒ์„ฑ๊ณผ ํ™œ์šฉ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ธฐ์กด ๋‹จ์ผ ์–ธ์–ด ์งˆ๋ฌธ ๋ฐ/๋˜๋Š” ๋‹ต๋ณ€์„ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ๋ฒˆ์—ญํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Asai et al., 2021). ๋˜ ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์€ Question Generation (QG) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€ ๋ฐ/๋˜๋Š” ์†Œ์Šค ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค๊ตญ์–ด ๋ฐฉ์‹์œผ๋กœ ํ•ฉ์„ฑ ์งˆ๋ฌธ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Chi et al., 2020; Kumar et al., 2019; Riabi et al., 2021). ์ตœ๊ทผ์˜ ๋…ธ๋ ฅ์€ ๋” ํฐ ์œ ์—ฐ์„ฑ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์–ธ์–ด๋กœ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ๊ณต๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Li and Callison-Burch, 2023; Shakeri et al., 2021). ์˜ˆ๋ฅผ ๋“ค์–ด, Shakeri et al.(2021)์€ QA ์ƒ์„ฑ ์ž‘์—…๊ณผ ๋‹ค๊ตญ์–ด ๋งˆ์Šคํฌ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์˜ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ ๋‹ค๊ตญ์–ด T5 ๋ชจ๋ธ(Xue et al., 2020)์„ Fine-tuningํ•˜์—ฌ ์—ฌ๋Ÿฌ ์–ธ์–ด๋กœ ํ•ฉ์„ฑ QA ์Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ์„ฑ QA ์Œ์œผ๋กœ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ค๊ตญ์–ด QA ๋ฐ ์ •๋ณด ๊ฒ€์ƒ‰ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋‹ค๊ตญ์–ด ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค๊ตญ์–ด ๋ฐ ๋‹ค๊ตญ์–ด ์งˆ๋ฌธ ๋‹ต๋ณ€(Abulkhanov et al., 2023; Asai et al., 2021; Chi et al., 2020; Kumar et al., 2019; Li and Callison-Burch, 2023; Riabi et al., 2021)์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋‹ค๊ตญ์–ด ์งˆ๋ฌธ ๋‹ต๋ณ€(QA) ์Œ์˜ ์ƒ์„ฑ๊ณผ ํ™œ์šฉ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ธฐ์กด ๋‹จ์ผ ์–ธ์–ด ์งˆ๋ฌธ ๋ฐ/๋˜๋Š” ๋‹ต๋ณ€์„ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ๋ฒˆ์—ญํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Asai et al., 2021). ๋˜ ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์งˆ๋ฌธ ์ƒ์„ฑ(QG) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€ ๋ฐ/๋˜๋Š” ์†Œ์Šค ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค๊ตญ์–ด ๋ฐฉ์‹์œผ๋กœ ํ•ฉ์„ฑ ์งˆ๋ฌธ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Chi et al., 2020; Kumar et al., 2019; Riabi et al., 2021). ์ตœ๊ทผ์˜ ๋…ธ๋ ฅ์€ ๋” ํฐ ์œ ์—ฐ์„ฑ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์–ธ์–ด๋กœ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ๊ณต๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Li and Callison-Burch, 2023; Shakeri et al., 2021). ์˜ˆ๋ฅผ ๋“ค์–ด, Shakeri et al.(2021)์€ QA ์ƒ์„ฑ ์ž‘์—…๊ณผ ๋‹ค๊ตญ์–ด ๋งˆ์Šคํฌ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์˜ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ ๋‹ค๊ตญ์–ด T5 ๋ชจ๋ธ(Xue et al., 2020)์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์—ฌ๋Ÿฌ ์–ธ์–ด๋กœ ํ•ฉ์„ฑ QA ์Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ์„ฑ QA ์Œ์œผ๋กœ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ค๊ตญ์–ด QA ๋ฐ ์ •๋ณด ๊ฒ€์ƒ‰ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
This survey comprehensively examines existing RAG models, summarizing their core techniques into four main steps from a retrieval perspective. It recognizes that some methods may encompass multiple steps and that decoupling these steps could potentially obscure their intrinsic connections. Nevertheless, the primary objective is to simplify the complexity of the approach, clearly delineating the specific problems it addresses. This allows for a clearer identification of areas ripe for further optimization and improvement. Despite the thorough investigation, the rapid evolution of the field and page limits mean that certain aspects might not have been fully analyzed and explored, or recent developments could have been missed. While the paper references evaluation methods that can aid in the development of RAG, it also acknowledges mature tools like LangChain and LlamaIndex as useful resources. However, the focus of this survey is not on detailing the evaluation pipeline or how these tools are specifically used, but rather on illustrating how evaluation aspects can support the advancement of RAG. This choice highlights an area for future work, emphasizing the importance of methodological clarity and the application of evaluation tools in refining and enhancing RAG models.
์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ๊ธฐ์กด Retrieval-augmented generation (RAG) ๋ชจ๋ธ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ์กฐ์‚ฌํ•˜์—ฌ ๊ฒ€์ƒ‰ ๊ด€์ ์—์„œ ํ•ต์‹ฌ ๊ธฐ์ˆ ์„ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ๋‹จ๊ณ„๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋ฅผ ๋ถ„๋ฆฌํ•˜๋ฉด ๋ณธ์งˆ์ ์ธ ์—ฐ๊ด€์„ฑ์ด ๋ชจํ˜ธํ•ด์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ธ์‹ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ด ๋ฐฉ๋ฒ•์˜ ์ฃผ์š” ๋ชฉํ‘œ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๋ณต์žก์„ฑ์„ ๋‹จ์ˆœํ™”ํ•˜์—ฌ ๊ตฌ์ฒด์ ์ธ ๋ฌธ์ œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ถ”๊ฐ€ ์ตœ์ ํ™” ๋ฐ ๊ฐœ์„ ์ด ํ•„์š”ํ•œ ์˜์—ญ์„ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒ ์ €ํ•œ ์กฐ์‚ฌ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด ๋ถ„์•ผ์˜ ๊ธ‰๊ฒฉํ•œ ๋ฐœ์ „๊ณผ ํŽ˜์ด์ง€ ์ œํ•œ์œผ๋กœ ์ธํ•ด ํŠน์ • ์ธก๋ฉด์„ ์™„์ „ํžˆ ๋ถ„์„ํ•˜๊ณ  ํƒ๊ตฌํ•˜์ง€ ๋ชปํ–ˆ๊ฑฐ๋‚˜ ์ตœ๊ทผ์˜ ๋ฐœ์ „์„ ๋†“์ณค์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” RAG ๊ฐœ๋ฐœ์— ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š” ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์–ธ๊ธ‰ํ•˜๊ณ  ์žˆ์ง€๋งŒ, LangChain ๋ฐ LlamaIndex์™€ ๊ฐ™์€ ์„ฑ์ˆ™ํ•œ ๋„๊ตฌ๋„ ์œ ์šฉํ•œ ๋ฆฌ์†Œ์Šค๋กœ ์ธ์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์„ค๋ฌธ์กฐ์‚ฌ์˜ ์ดˆ์ ์€ ํ‰๊ฐ€ ํŒŒ์ดํ”„๋ผ์ธ์ด๋‚˜ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๊ฐ€ ๊ตฌ์ฒด์ ์œผ๋กœ ์–ด๋–ป๊ฒŒ ์‚ฌ์šฉ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์ด ์•„๋‹ˆ๋ผ ํ‰๊ฐ€ ์ธก๋ฉด์ด RAG์˜ ๋ฐœ์ „์„ ์–ด๋–ป๊ฒŒ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๋ฐ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ ํƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ถ„์•ผ๋ฅผ ๊ฐ•์กฐํ•˜๋ฉฐ, ๋ฐฉ๋ฒ•๋ก ์  ๋ช…ํ™•์„ฑ๊ณผ RAG ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๊ณ  ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์žˆ์–ด ํ‰๊ฐ€ ๋„๊ตฌ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ๊ธฐ์กด RAG ๋ชจ๋ธ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ์กฐ์‚ฌํ•˜์—ฌ ๊ฒ€์ƒ‰ ๊ด€์ ์—์„œ ํ•ต์‹ฌ ๊ธฐ์ˆ ์„ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ๋‹จ๊ณ„๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋ฅผ ๋ถ„๋ฆฌํ•˜๋ฉด ๋ณธ์งˆ์ ์ธ ์—ฐ๊ด€์„ฑ์ด ๋ชจํ˜ธํ•ด์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ธ์‹ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ด ๋ฐฉ๋ฒ•์˜ ์ฃผ์š” ๋ชฉํ‘œ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๋ณต์žก์„ฑ์„ ๋‹จ์ˆœํ™”ํ•˜์—ฌ ๊ตฌ์ฒด์ ์ธ ๋ฌธ์ œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ถ”๊ฐ€ ์ตœ์ ํ™” ๋ฐ ๊ฐœ์„ ์ด ํ•„์š”ํ•œ ์˜์—ญ์„ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒ ์ €ํ•œ ์กฐ์‚ฌ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด ๋ถ„์•ผ์˜ ๊ธ‰๊ฒฉํ•œ ๋ฐœ์ „๊ณผ ํŽ˜์ด์ง€ ์ œํ•œ์œผ๋กœ ์ธํ•ด ํŠน์ • ์ธก๋ฉด์„ ์™„์ „ํžˆ ๋ถ„์„ํ•˜๊ณ  ํƒ๊ตฌํ•˜์ง€ ๋ชปํ–ˆ๊ฑฐ๋‚˜ ์ตœ๊ทผ์˜ ๋ฐœ์ „์„ ๋†“์ณค์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” RAG ๊ฐœ๋ฐœ์— ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š” ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์–ธ๊ธ‰ํ•˜๊ณ  ์žˆ์ง€๋งŒ, LangChain ๋ฐ LlamaIndex์™€ ๊ฐ™์€ ์„ฑ์ˆ™ํ•œ ๋„๊ตฌ๋„ ์œ ์šฉํ•œ ๋ฆฌ์†Œ์Šค๋กœ ์ธ์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์„ค๋ฌธ์กฐ์‚ฌ์˜ ์ดˆ์ ์€ ํ‰๊ฐ€ ํŒŒ์ดํ”„๋ผ์ธ์ด๋‚˜ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๊ฐ€ ๊ตฌ์ฒด์ ์œผ๋กœ ์–ด๋–ป๊ฒŒ ์‚ฌ์šฉ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์ด ์•„๋‹ˆ๋ผ ํ‰๊ฐ€ ์ธก๋ฉด์ด RAG์˜ ๋ฐœ์ „์„ ์–ด๋–ป๊ฒŒ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๋ฐ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ ํƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ถ„์•ผ๋ฅผ ๊ฐ•์กฐํ•˜๋ฉฐ, ๋ฐฉ๋ฒ•๋ก ์  ๋ช…ํ™•์„ฑ๊ณผ RAG ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๊ณ  ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์žˆ์–ด ํ‰๊ฐ€ ๋„๊ตฌ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
In contrast to these prior works we sought to anchor our human evaluation in criteria more commonly used for evaluating the quality of physicians' expertise in history-taking, including their communication skills in consultation. We derived a framework from principles published in reviews of the consensus for best practices for patient-centered communication (PCCBP) in medical interviews [20], criteria examined for history-taking skills by the Royal College of Physicians in the UK as part of their Practical Assessment of Clinical Examination Skills (PACES)3[30], and criteria proposed by the UK General Medical Council Patient Questionnaire (GMCPQ)4 for doctors seeking patient feedback as part of professional re-validation5. We iterated upon these criteria to refine items for inclusion and derived pilot scales and instructions for assessment by using focus groups and interviews with clinicians and OSCE examiners based in the UK, Canada, US, and India. Our resulting pilot framework enabled assessment from two perspectives: clinician (board-certifiedphysicians) and lay raters (patient actors). The framework included consideration of consultation quality, structure and completeness, the roles, responsibilities, and skills of the interviewer (Tables A.1, A.2, A.3, and A.4).
์ด๋Ÿฌํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ์™€๋Š” ๋‹ฌ๋ฆฌ ์šฐ๋ฆฌ๋Š” ์ƒ๋‹ด ์‹œ ์˜์‚ฌ์˜ ์˜์‚ฌ์†Œํ†ต ๋Šฅ๋ ฅ์„ ํฌํ•จํ•˜์—ฌ ๋ณ‘๋ ฅ ์ฒญ์ทจ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ๋” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ ค๊ณ  ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์˜๋ฃŒ ๋ฉด๋‹ด์—์„œ ํ™˜์ž ์ค‘์‹ฌ ์˜์‚ฌ์†Œํ†ต ๋ชจ๋ฒ” ์‚ฌ๋ก€์— ๋Œ€ํ•œ ํ•ฉ์˜์— ๋Œ€ํ•œ ๊ฒ€ํ† [20]์—์„œ ๋ฐœํ‘œ๋œ ์›์น™, ์˜๊ตญ ์™•๋ฆฝ ์˜๊ณผ ๋Œ€ํ•™์—์„œ ์ž„์ƒ ์ง„์ฐฐ ๊ธฐ์ˆ  ์‹ค๊ธฐ ํ‰๊ฐ€(PACES)3[30]์˜ ์ผํ™˜์œผ๋กœ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๊ธฐ์ˆ ์— ๋Œ€ํ•ด ๊ฒ€ํ† ํ•œ ๊ธฐ์ค€, ์˜๊ตญ ์ผ๋ฐ˜ ์˜์‚ฌํšŒ ํ™˜์ž ์„ค๋ฌธ์ง€(GMCPQ)4์—์„œ ์ „๋ฌธ๊ฐ€ ์žฌ๊ฒ€์ฆ5์˜ ์ผํ™˜์œผ๋กœ ํ™˜์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ตฌํ•˜๋Š” ์˜์‚ฌ๋ฅผ ์œ„ํ•ด ์ œ์•ˆํ•œ ๊ธฐ์ค€์—์„œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋„์ถœํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์„ ๋ฐ˜๋ณตํ•˜์—ฌ ํฌํ•จํ•  ํ•ญ๋ชฉ์„ ๊ตฌ์ฒดํ™”ํ•˜๊ณ , ์˜๊ตญ, ์บ๋‚˜๋‹ค, ๋ฏธ๊ตญ, ์ธ๋„์— ๊ธฐ๋ฐ˜์„ ๋‘” ์ž„์ƒ์˜ ๋ฐ OSCE ์‹œํ—˜๊ด€๋“ค๊ณผ์˜ ํฌ์ปค์Šค ๊ทธ๋ฃน ๋ฐ ์ธํ„ฐ๋ทฐ๋ฅผ ํ†ตํ•ด ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ํŒŒ์ผ๋Ÿฟ ์ฒ™๋„์™€ Instruction์„ ๋„์ถœํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋„์ถœ๋œ ํŒŒ์ผ๋Ÿฟ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํ†ตํ•ด ์ž„์ƒ์˜(๋ณด๋“œ ์ธ์ฆ ์ „๋ฌธ์˜)์™€ ์ผ๋ฐ˜ ํ‰๊ฐ€์ž(ํ™˜์ž ๋ฐฐ์šฐ์ž)์˜ ๋‘ ๊ฐ€์ง€ ๊ด€์ ์—์„œ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—๋Š” ์ƒ๋‹ด ํ’ˆ์งˆ, ๊ตฌ์กฐ ๋ฐ ์™„์ „์„ฑ, ๋ฉด๋‹ด์ž์˜ ์—ญํ• , ์ฑ…์ž„ ๋ฐ ๊ธฐ์ˆ  ๊ณ ๋ ค๊ฐ€ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(ํ‘œ A.1, A.2, A.3 ๋ฐ A.4).
์ด๋Ÿฌํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ์™€๋Š” ๋‹ฌ๋ฆฌ ์šฐ๋ฆฌ๋Š” ์ƒ๋‹ด ์‹œ ์˜์‚ฌ์˜ ์˜์‚ฌ์†Œํ†ต ๋Šฅ๋ ฅ์„ ํฌํ•จํ•˜์—ฌ ๋ณ‘๋ ฅ ์ฒญ์ทจ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ๋” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ ค๊ณ  ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์˜๋ฃŒ ๋ฉด๋‹ด์—์„œ ํ™˜์ž ์ค‘์‹ฌ ์˜์‚ฌ์†Œํ†ต ๋ชจ๋ฒ” ์‚ฌ๋ก€์— ๋Œ€ํ•œ ํ•ฉ์˜์— ๋Œ€ํ•œ ๊ฒ€ํ† [20]์—์„œ ๋ฐœํ‘œ๋œ ์›์น™, ์˜๊ตญ ์™•๋ฆฝ ์˜๊ณผ ๋Œ€ํ•™์—์„œ ์ž„์ƒ ์ง„์ฐฐ ๊ธฐ์ˆ  ์‹ค๊ธฐ ํ‰๊ฐ€(PACES)3[30]์˜ ์ผํ™˜์œผ๋กœ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๊ธฐ์ˆ ์— ๋Œ€ํ•ด ๊ฒ€ํ† ํ•œ ๊ธฐ์ค€, ์˜๊ตญ ์ผ๋ฐ˜ ์˜์‚ฌํšŒ ํ™˜์ž ์„ค๋ฌธ์ง€(GMCPQ)4์—์„œ ์ „๋ฌธ๊ฐ€ ์žฌ๊ฒ€์ฆ5์˜ ์ผํ™˜์œผ๋กœ ํ™˜์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ตฌํ•˜๋Š” ์˜์‚ฌ๋ฅผ ์œ„ํ•ด ์ œ์•ˆํ•œ ๊ธฐ์ค€์—์„œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋„์ถœํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์„ ๋ฐ˜๋ณตํ•˜์—ฌ ํฌํ•จํ•  ํ•ญ๋ชฉ์„ ๊ตฌ์ฒดํ™”ํ•˜๊ณ , ์˜๊ตญ, ์บ๋‚˜๋‹ค, ๋ฏธ๊ตญ, ์ธ๋„์— ๊ธฐ๋ฐ˜์„ ๋‘” ์ž„์ƒ์˜ ๋ฐ OSCE ์‹œํ—˜๊ด€๋“ค๊ณผ์˜ ํฌ์ปค์Šค ๊ทธ๋ฃน ๋ฐ ์ธํ„ฐ๋ทฐ๋ฅผ ํ†ตํ•ด ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ํŒŒ์ผ๋Ÿฟ ์ฒ™๋„์™€ ์ง€์นจ์„ ๋„์ถœํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋„์ถœ๋œ ํŒŒ์ผ๋Ÿฟ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํ†ตํ•ด ์ž„์ƒ์˜(๋ณด๋“œ ์ธ์ฆ ์ „๋ฌธ์˜)์™€ ์ผ๋ฐ˜ ํ‰๊ฐ€์ž(ํ™˜์ž ๋ฐฐ์šฐ์ž)์˜ ๋‘ ๊ฐ€์ง€ ๊ด€์ ์—์„œ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—๋Š” ์ƒ๋‹ด ํ’ˆ์งˆ, ๊ตฌ์กฐ ๋ฐ ์™„์ „์„ฑ, ๋ฉด๋‹ด์ž์˜ ์—ญํ• , ์ฑ…์ž„ ๋ฐ ๊ธฐ์ˆ  ๊ณ ๋ ค๊ฐ€ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(ํ‘œ A.1, A.2, A.3 ๋ฐ A.4).
To handle the issues of only local relations captured by CNN/RNN/GNN models, the Transformer [13] proposed in the year 2017 learns the long-range feature representations well. The core operation is the self-attention mechanism which transforms the input tokens into query, key, and value features, and outputs the long-range features by multiplying the similarity matrix (obtained via product between the query and key features) with the value features. The Transformer architecture first swept the NLP community with the help of _pre-training and fine-tuning_ paradigm [14], such as BERT [15], ERNIE [16], BART [17], GPT [18]. Then, other communities are also boosted with such networks, for example, the ViT [19] and Swin-Transformer [20] released in computer vision. Many researchers also exploit the hybrid network architectures by combining Transformer and other networks, or adapting the Transformer for multi-modal research problems [21, 22]. In the current stage, large foundation models are emerging, and Parameter-Efficient Fine-Tuning (PEFT) strategies [23] also have been greatly developed. However, the current Transformer-based models still require high-end graphics cards with larger memory for training and testing/deployment, which greatly limits their wider application.
CNN/RNN/GNN ๋ชจ๋ธ์—์„œ ์บก์ฒ˜ํ•œ ๋กœ์ปฌ ๊ด€๊ณ„๋งŒ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด 2017๋…„์— ์ œ์•ˆ๋œ Transformer[13]๋Š” ์›๊ฑฐ๋ฆฌ ํŠน์ง• ํ‘œํ˜„์„ ์ž˜ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ•ต์‹ฌ ์—ฐ์‚ฐ์€ self-attention ๋ฉ”์ปค๋‹ˆ์ฆ˜์œผ๋กœ, ์ž…๋ ฅ ํ† ํฐ์„ query, key, value ํŠน์ง•์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ณ , query ํŠน์ง•๊ณผ key ํŠน์ง•์˜ ๊ณฑ์„ ํ†ตํ•ด ์–ป์€ ์œ ์‚ฌ๋„ ํ–‰๋ ฌ์„ value ํŠน์ง•๊ณผ ๊ณฑํ•˜์—ฌ ์›๊ฑฐ๋ฆฌ ํŠน์ง•์„ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. Transformer ์•„ํ‚คํ…์ฒ˜๋Š” BERT [15], ERNIE [16], BART [17], GPT [18]์™€ ๊ฐ™์€ _pre-training ๋ฐ Fine-tuning_ ํŒจ๋Ÿฌ๋‹ค์ž„ [14]์˜ ๋„์›€์œผ๋กœ ๋จผ์ € NLP ์ปค๋ฎค๋‹ˆํ‹ฐ๋ฅผ ํœฉ์“ธ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„ ์ปดํ“จํ„ฐ ๋น„์ „์—์„œ ๊ณต๊ฐœ๋œ ViT [19] ๋ฐ Swin-Transformer [20]์™€ ๊ฐ™์€ ๋„คํŠธ์›Œํฌ๋กœ ๋‹ค๋ฅธ ์ปค๋ฎค๋‹ˆํ‹ฐ๋„ ํ™œ์„ฑํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์€ Transformer์™€ ๋‹ค๋ฅธ ๋„คํŠธ์›Œํฌ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ฑฐ๋‚˜ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์—ฐ๊ตฌ ๋ฌธ์ œ์— ๋งž๊ฒŒ Transformer๋ฅผ ์ ์šฉํ•˜์—ฌ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ๋„คํŠธ์›Œํฌ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ™œ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[21, 22]. ํ˜„์žฌ ๋‹จ๊ณ„์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, Parameter-Efficient Fine-Tuning (PEFT) ์ „๋žต[23]๋„ ํฌ๊ฒŒ ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ Transformer ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ์—ฌ์ „ํžˆ ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ/๋ฐฐํฌ๋ฅผ ์œ„ํ•ด ๋” ํฐ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ํ•˜์ด์—”๋“œ ๊ทธ๋ž˜ํ”ฝ ์นด๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋” ํญ๋„“๊ฒŒ ์ ์šฉ๋˜๋Š” ๋ฐ ํฐ ์ œํ•œ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
CNN/RNN/GNN ๋ชจ๋ธ์—์„œ ์บก์ฒ˜ํ•œ ๋กœ์ปฌ ๊ด€๊ณ„๋งŒ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด 2017๋…„์— ์ œ์•ˆ๋œ ํŠธ๋žœ์Šคํฌ๋จธ[13]๋Š” ์›๊ฑฐ๋ฆฌ ํŠน์ง• ํ‘œํ˜„์„ ์ž˜ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ•ต์‹ฌ ์—ฐ์‚ฐ์€ ์ž๊ธฐ ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์œผ๋กœ, ์ž…๋ ฅ ํ† ํฐ์„ ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ ํŠน์ง•์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ณ , ์ฟผ๋ฆฌ ํŠน์ง•๊ณผ ํ‚ค ํŠน์ง•์˜ ๊ณฑ์„ ํ†ตํ•ด ์–ป์€ ์œ ์‚ฌ๋„ ํ–‰๋ ฌ์„ ๊ฐ’ ํŠน์ง•๊ณผ ๊ณฑํ•˜์—ฌ ์›๊ฑฐ๋ฆฌ ํŠน์ง•์„ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋Š” BERT [15], ERNIE [16], BART [17], GPT [18]์™€ ๊ฐ™์€ _์‚ฌ์ „ ํ•™์Šต ๋ฐ ๋ฏธ์„ธ ์กฐ์ •_ ํŒจ๋Ÿฌ๋‹ค์ž„ [14]์˜ ๋„์›€์œผ๋กœ ๋จผ์ € NLP ์ปค๋ฎค๋‹ˆํ‹ฐ๋ฅผ ํœฉ์“ธ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„ ์ปดํ“จํ„ฐ ๋น„์ „์—์„œ ๊ณต๊ฐœ๋œ ViT [19] ๋ฐ Swin-Transformer [20]์™€ ๊ฐ™์€ ๋„คํŠธ์›Œํฌ๋กœ ๋‹ค๋ฅธ ์ปค๋ฎค๋‹ˆํ‹ฐ๋„ ํ™œ์„ฑํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์€ ํŠธ๋žœ์Šคํฌ๋จธ์™€ ๋‹ค๋ฅธ ๋„คํŠธ์›Œํฌ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ฑฐ๋‚˜ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์—ฐ๊ตฌ ๋ฌธ์ œ์— ๋งž๊ฒŒ ํŠธ๋žœ์Šคํฌ๋จธ๋ฅผ ์ ์šฉํ•˜์—ฌ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ๋„คํŠธ์›Œํฌ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ™œ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[21, 22]. ํ˜„์žฌ ๋‹จ๊ณ„์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๋งค๊ฐœ๋ณ€์ˆ˜ ํšจ์œจ์ ์ธ ๋ฏธ์„ธ ์กฐ์ •(PEFT) ์ „๋žต[23]๋„ ํฌ๊ฒŒ ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ์—ฌ์ „ํžˆ ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ/๋ฐฐํฌ๋ฅผ ์œ„ํ•ด ๋” ํฐ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ํ•˜์ด์—”๋“œ ๊ทธ๋ž˜ํ”ฝ ์นด๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋” ํญ๋„“๊ฒŒ ์ ์šฉ๋˜๋Š” ๋ฐ ํฐ ์ œํ•œ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
**Results with LLaMA-based Target LLMs.** Table 1 summarizes the performance of CodecLM and the comparing baselines with 2000 synthetic data for instruction tuning. All methods are trained on LLaMA-7B or -13B as the target LLM and compared against Gemini-Pro, the strong LLM that generates the data. CodecLM outperforms comparing methods consistently on all benchmarks, with two target LLMs of different sizes. The consistently superior performance of CodecLM highlights its generalizability to different downstream instruction distributions and target LLMs. Both Tree-Instruct and variants of WizardLM focus on the importance of instruction complexity, however, their performances are not always better than Alpagasus with simple instructions, especially with larger target LLM. This observation indicates that the effectiveness of data cannot be solely determined by instruction complexity, and validates the motivation of our design of Self-Rubrics and Contrastive Filtering. Moreover, the win of WizardLM+ over WizardLM confirms the efficacy of instruction distribution matching via instruction metadata. When shifting the target LLM from LLaMA-7B to -13B, all methods get a significant performance boost, which accords with prior discoveries on scaling model size Wei et al. (2021).
**LLaMA ๊ธฐ๋ฐ˜ ํƒ€๊ฒŸ LLM์˜ ๊ฒฐ๊ณผ.** ํ‘œ 1์€ Instruction Tuning์„ ์œ„ํ•ด 2000๊ฐœ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ CodecLM๊ณผ ๋น„๊ต Baseline์˜ ์„ฑ๋Šฅ์„ ์š”์•ฝํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐฉ๋ฒ•์€ ํƒ€๊ฒŸ LLM์œผ๋กœ LLaMA-7B ๋˜๋Š” -13B์—์„œ ํ•™์Šต๋˜๋ฉฐ, ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฐ•๋ ฅํ•œ LLM์ธ Gemini-Pro์™€ ๋น„๊ต๋ฉ๋‹ˆ๋‹ค. CodecLM์€ ์„œ๋กœ ๋‹ค๋ฅธ ํฌ๊ธฐ์˜ ๋‘ ๊ฐ€์ง€ ํƒ€๊ฒŸ LLM์„ ํ†ตํ•ด ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ์—์„œ ๋น„๊ต ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. CodecLM์˜ ์ผ๊ด€๋œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์€ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ Instruction ๋ถ„ํฌ์™€ ํƒ€๊ฒŸ LLM์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. Tree-Instruct์™€ WizardLM์˜ ๋ณ€ํ˜•์€ Instruction ๋ณต์žก์„ฑ์˜ ์ค‘์š”์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์ง€๋งŒ, ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ํƒ€๊ฒŸ LLM์˜ ๊ฒฝ์šฐ ๋‹จ์ˆœํ•œ Instruction์„ ๊ฐ€์ง„ Alpagasus๋ณด๋‹ค ์„ฑ๋Šฅ์ด ํ•ญ์ƒ ์šฐ์ˆ˜ํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ฐ์ดํ„ฐ์˜ ํšจ๊ณผ๊ฐ€ Instruction ๋ณต์žก์„ฑ๋งŒ์œผ๋กœ ๊ฒฐ์ •๋  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, Self-Rubrics์™€ Contrastive Filtering ์„ค๊ณ„์˜ ๋™๊ธฐ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ WizardLM+๊ฐ€ WizardLM๋ณด๋‹ค ์šฐ์œ„๋ฅผ ์ ํ•œ ๊ฒƒ์€ Instruction ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•œ Instruction ๋ถ„ํฌ ๋งค์นญ์˜ ํšจ๋Šฅ์„ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค. ํƒ€๊ฒŸ LLM์„ LLaMA-7B์—์„œ -13B๋กœ ๋ณ€๊ฒฝํ•˜๋ฉด ๋ชจ๋“  ๋ฐฉ๋ฒ•์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ Wei ๋“ฑ(2021)์˜ ํ™•์žฅ์— ๋Œ€ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค.
**LLaMA ๊ธฐ๋ฐ˜ ํƒ€๊ฒŸ LLM์˜ ๊ฒฐ๊ณผ.** ํ‘œ 1์€ ๋ช…๋ น์–ด ํŠœ๋‹์„ ์œ„ํ•ด 2000๊ฐœ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ์ฝ”๋ฑLM๊ณผ ๋น„๊ต ๋ฒ ์ด์Šค๋ผ์ธ์˜ ์„ฑ๋Šฅ์„ ์š”์•ฝํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐฉ๋ฒ•์€ ํƒ€๊ฒŸ LLM์œผ๋กœ LLaMA-7B ๋˜๋Š” -13B์—์„œ ํ•™์Šต๋˜๋ฉฐ, ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฐ•๋ ฅํ•œ LLM์ธ Gemini-Pro์™€ ๋น„๊ต๋ฉ๋‹ˆ๋‹ค. ์ฝ”๋ฑLM์€ ์„œ๋กœ ๋‹ค๋ฅธ ํฌ๊ธฐ์˜ ๋‘ ๊ฐ€์ง€ ํƒ€๊ฒŸ LLM์„ ํ†ตํ•ด ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ์—์„œ ๋น„๊ต ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฝ”๋ฑLM์˜ ์ผ๊ด€๋œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์€ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ช…๋ น์–ด ๋ถ„ํฌ์™€ ํƒ€๊ฒŸ LLM์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋ฆฌ ์ธ์ŠคํŠธ๋ŸญํŠธ์™€ ์œ„์ž๋“œLM์˜ ๋ณ€ํ˜•์€ ๋ช…๋ น์–ด ๋ณต์žก์„ฑ์˜ ์ค‘์š”์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์ง€๋งŒ, ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ํƒ€๊ฒŸ LLM์˜ ๊ฒฝ์šฐ ๋‹จ์ˆœํ•œ ๋ช…๋ น์–ด๋ฅผ ๊ฐ€์ง„ ์•ŒํŒŒ๊ฐ€์ˆ˜๋ณด๋‹ค ์„ฑ๋Šฅ์ด ํ•ญ์ƒ ์šฐ์ˆ˜ํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ฐ์ดํ„ฐ์˜ ํšจ๊ณผ๊ฐ€ ๋ช…๋ น์–ด ๋ณต์žก์„ฑ๋งŒ์œผ๋กœ ๊ฒฐ์ •๋  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์…€ํ”„ ๋ฃจ๋ธŒ๋ฆญ๊ณผ ๋Œ€์กฐ ํ•„ํ„ฐ๋ง ์„ค๊ณ„์˜ ๋™๊ธฐ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ WizardLM+๊ฐ€ WizardLM๋ณด๋‹ค ์šฐ์œ„๋ฅผ ์ ํ•œ ๊ฒƒ์€ ๋ช…๋ น์–ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•œ ๋ช…๋ น์–ด ๋ถ„ํฌ ๋งค์นญ์˜ ํšจ๋Šฅ์„ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค. ํƒ€๊ฒŸ LLM์„ LLaMA-7B์—์„œ -13B๋กœ ๋ณ€๊ฒฝํ•˜๋ฉด ๋ชจ๋“  ๋ฐฉ๋ฒ•์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ Wei ๋“ฑ(2021)์˜ ํ™•์žฅ์— ๋Œ€ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค.
* Context size: We set the context size to 2048, which determines the maximum number of tokens the model can consider simultaneously when generating text. * Maximum sequence length: We limit the generated sequence length to 512 tokens to ensure that the outputs remain focused and relevant to the input prompt. * Temperature: We set the temperature to 0.2, which controls the randomness of the sampling process. Lower values make the model generate more focused and deterministic outputs, while higher values increase diversity at the cost of coherence. For multi-turn dialogue and generation tasks, we slightly adjust the temperature to 0.5 to allow a more diverse output. * Top-\(k\) sampling: We use Top-\(k\) sampling with \(k=40\), meaning that the model selects its next token from the top 40 most probable tokens at each step, adding an element of randomness and diversity to the generated text. * Top-\(p\) sampling: We also employ Top-\(p\) sampling with \(p=0.9\), which further enhances diversity by considering a dynamic set of tokens that collectively account for 90% of the probability mass. * Repetition penalty: To discourage the model from generating repetitive text, we apply a repetition penalty with a factor of 1.1, penalizing tokens that have already been selected.
* Context size: ๋ฌธ๋งฅ ํฌ๊ธฐ๋ฅผ 2048๋กœ ์„ค์ •ํ•˜์—ฌ ๋ชจ๋ธ์ด ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ๋™์‹œ์— ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. * Maximum sequence length: ์ƒ์„ฑ๋œ ์‹œํ€€์Šค ๊ธธ์ด๋ฅผ 512 ํ† ํฐ์œผ๋กœ ์ œํ•œํ•˜์—ฌ ์ถœ๋ ฅ์ด ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์— ์ง‘์ค‘๋˜๊ณ  ๊ด€๋ จ์„ฑ์„ ์œ ์ง€ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. * Temperature: ์˜จ๋„๋ฅผ 0.2๋กœ ์„ค์ •ํ•˜์—ฌ ์ƒ˜ํ”Œ๋ง ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด์ž‘์œ„์„ฑ์„ ์ œ์–ดํ•ฉ๋‹ˆ๋‹ค. ๊ฐ’์ด ๋‚ฎ์„์ˆ˜๋ก ๋ชจ๋ธ์ด ๋” ์ง‘์ค‘๋˜๊ณ  ๊ฒฐ์ •๋ก ์ ์ธ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๊ณ , ๊ฐ’์ด ๋†’์„์ˆ˜๋ก ์ผ๊ด€์„ฑ์„ ํฌ์ƒํ•˜๋ฉด์„œ ๋‹ค์–‘์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™” ๋ฐ ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ ์˜จ๋„๋ฅผ 0.5๋กœ ์•ฝ๊ฐ„ ์กฐ์ •ํ•˜์—ฌ ๋” ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. * Top-\(k\) sampling: Top-\(k\) ์ƒ˜ํ”Œ๋ง์„ \(k=40\)์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ ์ƒ์œ„ 40๊ฐœ ์ค‘์—์„œ ๋‹ค์Œ ํ† ํฐ์„ ์„ ํƒํ•˜๋„๋ก ํ•˜์—ฌ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์— ๋ฌด์ž‘์œ„์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋”ํ•ฉ๋‹ˆ๋‹ค. * Top-\(p\) sampling: ๋˜ํ•œ, ์ง‘ํ•ฉ์ ์œผ๋กœ ํ™•๋ฅ  ์งˆ๋Ÿ‰์˜ 90%๋ฅผ ์ฐจ์ง€ํ•˜๋Š” ๋™์  ํ† ํฐ ์ง‘ํ•ฉ์„ ๊ณ ๋ คํ•˜์—ฌ ๋‹ค์–‘์„ฑ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” Top-\(p\) ์ƒ˜ํ”Œ๋ง์„ \(p=0.9\)๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * Repetition penalty: ๋ชจ๋ธ์ด ๋ฐ˜๋ณต์ ์ธ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด 1.1์˜ ๊ณ„์ˆ˜๋กœ Repetition penalty๋ฅผ ์ ์šฉํ•˜์—ฌ ์ด๋ฏธ ์„ ํƒ๋œ ํ† ํฐ์— ๋ถˆ์ด์ต์„ ์ค๋‹ˆ๋‹ค.
* ๋ฌธ๋งฅ ํฌ๊ธฐ: ๋ฌธ๋งฅ ํฌ๊ธฐ๋ฅผ 2048๋กœ ์„ค์ •ํ•˜์—ฌ ๋ชจ๋ธ์ด ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ๋™์‹œ์— ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. * ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด: ์ƒ์„ฑ๋œ ์‹œํ€€์Šค ๊ธธ์ด๋ฅผ 512 ํ† ํฐ์œผ๋กœ ์ œํ•œํ•˜์—ฌ ์ถœ๋ ฅ์ด ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์— ์ง‘์ค‘๋˜๊ณ  ๊ด€๋ จ์„ฑ์„ ์œ ์ง€ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. * ์˜จ๋„: ์˜จ๋„๋ฅผ 0.2๋กœ ์„ค์ •ํ•˜์—ฌ ์ƒ˜ํ”Œ๋ง ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด์ž‘์œ„์„ฑ์„ ์ œ์–ดํ•ฉ๋‹ˆ๋‹ค. ๊ฐ’์ด ๋‚ฎ์„์ˆ˜๋ก ๋ชจ๋ธ์ด ๋” ์ง‘์ค‘๋˜๊ณ  ๊ฒฐ์ •๋ก ์ ์ธ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๊ณ , ๊ฐ’์ด ๋†’์„์ˆ˜๋ก ์ผ๊ด€์„ฑ์„ ํฌ์ƒํ•˜๋ฉด์„œ ๋‹ค์–‘์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™” ๋ฐ ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ ์˜จ๋„๋ฅผ 0.5๋กœ ์•ฝ๊ฐ„ ์กฐ์ •ํ•˜์—ฌ ๋” ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. * ์ƒ์œ„-\(k\) ์ƒ˜ํ”Œ๋ง: ์ƒ์œ„-\(k\) ์ƒ˜ํ”Œ๋ง์„ \(k=40\)์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ ์ƒ์œ„ 40๊ฐœ ์ค‘์—์„œ ๋‹ค์Œ ํ† ํฐ์„ ์„ ํƒํ•˜๋„๋ก ํ•˜์—ฌ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์— ๋ฌด์ž‘์œ„์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋”ํ•ฉ๋‹ˆ๋‹ค. * ์ƒ์œ„-\(p\) ์ƒ˜ํ”Œ๋ง: ๋˜ํ•œ, ์ง‘ํ•ฉ์ ์œผ๋กœ ํ™•๋ฅ  ์งˆ๋Ÿ‰์˜ 90%๋ฅผ ์ฐจ์ง€ํ•˜๋Š” ๋™์  ํ† ํฐ ์ง‘ํ•ฉ์„ ๊ณ ๋ คํ•˜์—ฌ ๋‹ค์–‘์„ฑ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” Top-\(p\) ์ƒ˜ํ”Œ๋ง์„ \(p=0.9\)๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * ๋ฐ˜๋ณต ํŽ˜๋„ํ‹ฐ: ๋ชจ๋ธ์ด ๋ฐ˜๋ณต์ ์ธ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด 1.1์˜ ๊ณ„์ˆ˜๋กœ ๋ฐ˜๋ณต ํŽ˜๋„ํ‹ฐ๋ฅผ ์ ์šฉํ•˜์—ฌ ์ด๋ฏธ ์„ ํƒ๋œ ํ† ํฐ์— ๋ถˆ์ด์ต์„ ์ค๋‹ˆ๋‹ค.
struction data generated by real human users, OpenAI's LLMs (e.g., InstructGPT [2] and ChatGPT 4) have achieved great success. These open-domain instructions can fully unleash the unlimited potential of LLMs [14, 15, 16, 17] and enable them to perform more complex and diverse tasks. However, using humans to create open-domain instruction datasets like OpenAI did will encounter the following challenges. The whole annotating process is extremely expensive and time-consuming [18, 19, 20, 21]. On the other hand, the difficulty level distribution of human-created instructions is skewed towards being easy or moderate, with fewer difficult ones (according to the difficulty statistics of ShareGPT [22] from Figure (a)a). Possible reasons for this are that the proportion of experts among annotators is low and creating complex instructions demands a lot of mental effort. Human annotators are prone to fatigue and cannot sustain high-intensity work to produce a sufficient proportion of high-difficulty instructions [23, 24, 25, 26]. Based on these issues, developing an automatic method that can mass-produce open-domain instructions (especially the more difficult ones) at a relatively low cost becomes the key to further advancing instruction-tuned language models [27, 28, 29, 30].
์‹ค์ œ ์ธ๊ฐ„ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒ์„ฑํ•œ instruction ๋ฐ์ดํ„ฐ์— ๊ธฐ๋ฐ˜ํ•œ OpenAI์˜ LLM(์˜ˆ: InstructGPT [2] ๋ฐ ChatGPT 4)์€ ํฐ ์„ฑ๊ณต์„ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜คํ”ˆ ๋„๋ฉ”์ธ instructions๋Š” LLM์˜ ๋ฌดํ•œํ•œ ์ž ์žฌ๋ ฅ์„ ์™„์ „ํžˆ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[14, 15, 16, 17] ๋ณด๋‹ค ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ OpenAI๊ฐ€ ๊ทธ๋žฌ๋˜ ๊ฒƒ์ฒ˜๋Ÿผ ์ธ๊ฐ„์„ ๋™์›ํ•ด ์˜คํ”ˆ ๋„๋ฉ”์ธ instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์ฃผ์„ ๋‹ฌ๊ธฐ ํ”„๋กœ์„ธ์Šค๋Š” ์—„์ฒญ๋‚˜๊ฒŒ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฝ๋‹ˆ๋‹ค[18, 19, 20, 21]. ๋ฐ˜๋ฉด, ์ธ๊ฐ„์ด ๋งŒ๋“  instructions์˜ ๋‚œ์ด๋„ ๋ถ„ํฌ๋Š” ์‰ฌ์šด ๋ช…๋ น์–ด๋‚˜ ๋ณดํ†ต ์ˆ˜์ค€์˜ ๋ช…๋ น์–ด์— ์น˜์šฐ์ณ ์žˆ์œผ๋ฉฐ, ์–ด๋ ค์šด ๋ช…๋ น์–ด๋Š” ์ ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ (a)a์˜ ShareGPT [22]์˜ ๋‚œ์ด๋„ ํ†ต๊ณ„์— ๋”ฐ๋ฅด๋ฉด). ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ์ฃผ์„ ์ž‘์„ฑ์ž ์ค‘ ์ „๋ฌธ๊ฐ€์˜ ๋น„์œจ์ด ๋‚ฎ๊ณ  ๋ณต์žกํ•œ instructions๋ฅผ ์ž‘์„ฑํ•˜๋ ค๋ฉด ๋งŽ์€ ์ •์‹ ์  ๋…ธ๋ ฅ์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ธ๊ฐ„ ์ฃผ์„ ์ž‘์„ฑ์ž๋Š” ํ”ผ๋กœ์— ์ทจ์•ฝํ•˜๋ฉฐ ๊ณ ๊ฐ•๋„ ์ž‘์—…์„ ์œ ์ง€ํ•˜์—ฌ ์ถฉ๋ถ„ํ•œ ๋น„์œจ์˜ ๊ณ ๋‚œ์ด๋„ instructions๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค[23, 24, 25, 26]. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋น„๊ต์  ์ €๋ ดํ•œ ๋น„์šฉ์œผ๋กœ ์˜คํ”ˆ ๋„๋ฉ”์ธ instructions(ํŠนํžˆ ์–ด๋ ค์šด ๋ช…๋ น์–ด)๋ฅผ ๋Œ€๋Ÿ‰ ์ƒ์‚ฐํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด instruction-tuned ์–ธ์–ด ๋ชจ๋ธ์„ ๋”์šฑ ๋ฐœ์ „์‹œํ‚ค๋Š” ์—ด์‡ ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค[27, 28, 29, 30].
์‹ค์ œ ์ธ๊ฐ„ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒ์„ฑํ•œ ๊ตฌ์กฐ ๋ฐ์ดํ„ฐ์— ๊ธฐ๋ฐ˜ํ•œ OpenAI์˜ LLM(์˜ˆ: InstructGPT [2] ๋ฐ ChatGPT 4)์€ ํฐ ์„ฑ๊ณต์„ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋ช…๋ น์–ด๋Š” LLM์˜ ๋ฌดํ•œํ•œ ์ž ์žฌ๋ ฅ์„ ์™„์ „ํžˆ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[14, 15, 16, 17] ๋ณด๋‹ค ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ OpenAI๊ฐ€ ๊ทธ๋žฌ๋˜ ๊ฒƒ์ฒ˜๋Ÿผ ์ธ๊ฐ„์„ ๋™์›ํ•ด ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์ฃผ์„ ๋‹ฌ๊ธฐ ํ”„๋กœ์„ธ์Šค๋Š” ์—„์ฒญ๋‚˜๊ฒŒ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฝ๋‹ˆ๋‹ค[18, 19, 20, 21]. ๋ฐ˜๋ฉด, ์ธ๊ฐ„์ด ๋งŒ๋“  ๋ช…๋ น์–ด์˜ ๋‚œ์ด๋„ ๋ถ„ํฌ๋Š” ์‰ฌ์šด ๋ช…๋ น์–ด๋‚˜ ๋ณดํ†ต ์ˆ˜์ค€์˜ ๋ช…๋ น์–ด์— ์น˜์šฐ์ณ ์žˆ์œผ๋ฉฐ, ์–ด๋ ค์šด ๋ช…๋ น์–ด๋Š” ์ ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ (a)a์˜ ShareGPT [22]์˜ ๋‚œ์ด๋„ ํ†ต๊ณ„์— ๋”ฐ๋ฅด๋ฉด). ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ์ฃผ์„ ์ž‘์„ฑ์ž ์ค‘ ์ „๋ฌธ๊ฐ€์˜ ๋น„์œจ์ด ๋‚ฎ๊ณ  ๋ณต์žกํ•œ ๋ช…๋ น์–ด๋ฅผ ์ž‘์„ฑํ•˜๋ ค๋ฉด ๋งŽ์€ ์ •์‹ ์  ๋…ธ๋ ฅ์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ธ๊ฐ„ ์ฃผ์„ ์ž‘์„ฑ์ž๋Š” ํ”ผ๋กœ์— ์ทจ์•ฝํ•˜๋ฉฐ ๊ณ ๊ฐ•๋„ ์ž‘์—…์„ ์œ ์ง€ํ•˜์—ฌ ์ถฉ๋ถ„ํ•œ ๋น„์œจ์˜ ๊ณ ๋‚œ์ด๋„ ์ง€์นจ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค[23, 24, 25, 26]. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋น„๊ต์  ์ €๋ ดํ•œ ๋น„์šฉ์œผ๋กœ ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋ช…๋ น์–ด(ํŠนํžˆ ์–ด๋ ค์šด ๋ช…๋ น์–ด)๋ฅผ ๋Œ€๋Ÿ‰ ์ƒ์‚ฐํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ง€์‹œ ํŠœ๋‹ ์–ธ์–ด ๋ชจ๋ธ์„ ๋”์šฑ ๋ฐœ์ „์‹œํ‚ค๋Š” ์—ด์‡ ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค[27, 28, 29, 30].
If we continue relying on scaling laws to improve LLMs, we will quickly hit diminishing returns due to the power-law nature of scaling laws. We will therefore need exponentially more data to maintain a consistent marginal improvement, which may prove especially challenging as we are fastapproaching the limits of available human-generated text data [51]. Encouragingly, in the context of vision, Sorscher et al. [47] demonstrated that we could leverage simple data selection strategies to overcome costly power-law scaling. They compare numerous data selection methods and find that clustering data points in a pre-trained embedding space and ranking according to the distance to the cluster centroid ("SSL Prototypes") significantly improves the data efficiency of vision models. Recently, Abbas et al. [1] demonstrated that using a pre-trained embedding space to de-duplicate data ("SemDeDup") improves both efficiency and performance of vision-language models such as CLIP. However, there has been little exploration of these or related approaches in training LLMs at scale. Motivated by this, we argue that by combining these approaches and applying them to LLMs, relatively simple data selection strategies leveraging pre-trained embeddings can significantly improve LLM training.
์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๊ณ„์† ์˜์กดํ•˜์—ฌ LLM์„ ๊ฐœ์„ ํ•œ๋‹ค๋ฉด, scaling laws์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ์ˆ˜์ต ๊ฐ์†Œ์— ๋น ๋ฅด๊ฒŒ ๋ถ€๋”ชํž ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๊ด€๋œ ํ•œ๊ณ„ ๊ฐœ์„ ์„ ์œ ์ง€ํ•˜๋ ค๋ฉด ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๋ฉฐ, ์ด๋Š” ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ธ๊ฐ„ ์ƒ์„ฑ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ํ•œ๊ณ„์— ๋น ๋ฅด๊ฒŒ ์ ‘๊ทผํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํŠนํžˆ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[51]. ๊ณ ๋ฌด์ ์ธ ์ ์€ ์‹œ๊ฐ์˜ ๋งฅ๋ฝ์—์„œ Sorscher ๋“ฑ[47]์ด ๊ฐ„๋‹จํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ํ™œ์šฉํ•˜์—ฌ ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” power-law scaling์„ ๊ทน๋ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ˆ˜๋งŽ์€ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ์‚ฌ์ „ ํ•™์Šต๋œ embedding ๊ณต๊ฐ„์—์„œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•˜๊ณ  ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ๊ณผ์˜ ๊ฑฐ๋ฆฌ์— ๋”ฐ๋ผ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋Š”("SSL Prototypes") ๋ฐฉ๋ฒ•์ด ๋น„์ „ ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Abbas ๋“ฑ[1]์€ ์‚ฌ์ „ ํ•™์Šต๋œ embedding ๊ณต๊ฐ„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ CLIP์™€ ๊ฐ™์€ vision-language ๋ชจ๋ธ์˜ ํšจ์œจ์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ๋ชจ๋‘ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด๋‚˜ ์ด์™€ ๊ด€๋ จ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋Œ€๊ทœ๋ชจ LLM ํ›ˆ๋ จ์— ์ ์šฉํ•œ ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์— ์ฐฉ์•ˆํ•˜์—ฌ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒฐํ•ฉํ•˜๊ณ  ์ด๋ฅผ LLM์— ์ ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๋œ embeddings์„ ํ™œ์šฉํ•˜๋Š” ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์ด LLM ํ•™์Šต์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.
์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๊ณ„์† ์˜์กดํ•˜์—ฌ LLM์„ ๊ฐœ์„ ํ•œ๋‹ค๋ฉด, ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ์ˆ˜์ต ๊ฐ์†Œ์— ๋น ๋ฅด๊ฒŒ ๋ถ€๋”ชํž ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๊ด€๋œ ํ•œ๊ณ„ ๊ฐœ์„ ์„ ์œ ์ง€ํ•˜๋ ค๋ฉด ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๋ฉฐ, ์ด๋Š” ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ธ๊ฐ„ ์ƒ์„ฑ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ํ•œ๊ณ„์— ๋น ๋ฅด๊ฒŒ ์ ‘๊ทผํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํŠนํžˆ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[51]. ๊ณ ๋ฌด์ ์ธ ์ ์€ ์‹œ๊ฐ์˜ ๋งฅ๋ฝ์—์„œ ์†Œ๋ฅด์…ฐ๋ฅด ๋“ฑ[47]์ด ๊ฐ„๋‹จํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ํ™œ์šฉํ•˜์—ฌ ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ๊ฑฐ๋“ญ์ œ๊ณฑ ์Šค์ผ€์ผ๋ง์„ ๊ทน๋ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ˆ˜๋งŽ์€ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ์‚ฌ์ „ ํ•™์Šต๋œ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•˜๊ณ  ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ๊ณผ์˜ ๊ฑฐ๋ฆฌ์— ๋”ฐ๋ผ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋Š”("SSL ํ”„๋กœํ† ํƒ€์ž…") ๋ฐฉ๋ฒ•์ด ๋น„์ „ ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Abbas ๋“ฑ[1]์€ ์‚ฌ์ „ ํ•™์Šต๋œ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ CLIP์™€ ๊ฐ™์€ ๋น„์ „-์–ธ์–ด ๋ชจ๋ธ์˜ ํšจ์œจ์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ๋ชจ๋‘ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด๋‚˜ ์ด์™€ ๊ด€๋ จ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋Œ€๊ทœ๋ชจ LLM ํ›ˆ๋ จ์— ์ ์šฉํ•œ ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์— ์ฐฉ์•ˆํ•˜์—ฌ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒฐํ•ฉํ•˜๊ณ  ์ด๋ฅผ LLM์— ์ ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๋œ ์ž„๋ฒ ๋”ฉ์„ ํ™œ์šฉํ•˜๋Š” ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์ด LLM ํ•™์Šต์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.
**Low-Rank Adaptation (LoRA)**. LoRA [145] imposes the low-rank constraint for approximating the update matrix at each dense layer, so as to reduce the trainable parameters for adapting to downstream tasks. Consider the case of optimizing a parameter matrix \(\mathbf{W}\). The update process can be written in a general form as: \(\mathbf{W}\leftarrow\mathbf{W}+\Delta\mathbf{W}\). The basic idea of LoRA is to freeze the original matrix \(\mathbf{W}\in\mathbb{R}^{m\times n}\) while approximating the parameter update \(\Delta\mathbf{W}\) by low-rank decomposition matrices, _i.e.,_\(\Delta\mathbf{W}=\mathbf{A}\cdot\mathbf{B}^{\top}\), where \(\mathbf{A}\in\mathbb{R}^{m\times k}\) and \(\mathbf{B}\in\mathbb{R}^{n\times k}\) are the trainable parameters for task adaptation and \(k\ll\min(m,n)\) is the reduced rank. The major merit of LoRA is that it can largely save the memory and storage usage (_e.g.,_ VRAM). Further, one can only keep a single large model copy, while maintaining a number of task-specific low-rank decomposition matrices for adapting to different downstream tasks. Further, several studies have also discussed how to set the rank in a more principled approach, _e.g.,_ importance score based allocation [406] and search-free optimal rank selection [407].
**Low-Rank Adaptation (LoRA)**. LoRA [145]๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ฐ ๋ฐ€๋„ ๋ ˆ์ด์–ด์—์„œ ์—…๋ฐ์ดํŠธ ํ–‰๋ ฌ์„ ๊ทผ์‚ฌํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ €๊ธ‰ ์ œ์•ฝ์„ ๋ถ€๊ณผํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ํ–‰๋ ฌ \(\mathbf{W}\)๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ์ƒ๊ฐํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์—…๋ฐ์ดํŠธ ํ”„๋กœ์„ธ์Šค๋Š” ์ผ๋ฐ˜์ ์ธ ํ˜•ํƒœ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: \(\mathbf{W}\leftarrow\mathbf{W}+\Delta\mathbf{W}\). LoRA์˜ ๊ธฐ๋ณธ ์•„์ด๋””์–ด๋Š” ์›๋ž˜ ํ–‰๋ ฌ \(\mathbf{W}\in\mathbb{R}^{m\times n}\)์„ ๋™๊ฒฐํ•œ ์ฑ„๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ์—…๋ฐ์ดํŠธ \(\Delta\mathbf{W}\)๋ฅผ ์ €๊ธ‰ ๋ถ„ํ•ด ํ–‰๋ ฌ, ์ฆ‰ \(\Delta\mathbf{W}=\mathbf{A}\cdot\mathbf{B}^{\top}\), ์—ฌ๊ธฐ์„œ \(\mathbf{A}\in\mathbb{R}^{m\times k}\) ๋ฐ \(\mathbf{B}\in\mathbb{R}^{n\times k}\)๋Š” ์ž‘์—… ์ ์‘์„ ์œ„ํ•œ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ์ด๋ฉฐ \(k\ll\min(m,n)\)์€ ํ™˜์›๋œ ์ˆœ์œ„์ž…๋‹ˆ๋‹ค. LoRA์˜ ๊ฐ€์žฅ ํฐ ์žฅ์ ์€ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์ €์žฅ์†Œ ์‚ฌ์šฉ๋Ÿ‰(_์˜ˆ:_ VRAM)์„ ํฌ๊ฒŒ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ•˜๋‚˜์˜ ํฐ ๋ชจ๋ธ ๋ณต์‚ฌ๋ณธ๋งŒ ์œ ์ง€ํ•˜๋ฉด์„œ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ž‘์—…๋ณ„ ์ €๊ธ‰ ๋ถ„ํ•ด ํ–‰๋ ฌ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ค‘์š”๋„ ์ ์ˆ˜ ๊ธฐ๋ฐ˜ ํ• ๋‹น [406] ๋ฐ ๊ฒ€์ƒ‰ ์—†๋Š” ์ตœ์  ์ˆœ์œ„ ์„ ํƒ [407]๊ณผ ๊ฐ™์ด ๋ณด๋‹ค ์›์น™์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์ˆœ์œ„๋ฅผ ์„ค์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด์„œ๋„ ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
**Low-Rank Adaptation(LoRA)**. LoRA [145]๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ฐ ๋ฐ€๋„ ๋ ˆ์ด์–ด์—์„œ ์—…๋ฐ์ดํŠธ ํ–‰๋ ฌ์„ ๊ทผ์‚ฌํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ €๊ธ‰ ์ œ์•ฝ์„ ๋ถ€๊ณผํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ํ–‰๋ ฌ \(\mathbf{W}\)๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ์ƒ๊ฐํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์—…๋ฐ์ดํŠธ ํ”„๋กœ์„ธ์Šค๋Š” ์ผ๋ฐ˜์ ์ธ ํ˜•ํƒœ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: \(\mathbf{W}\leftarrow\mathbf{W}+\Delta\mathbf{W}\). LoRA์˜ ๊ธฐ๋ณธ ์•„์ด๋””์–ด๋Š” ์›๋ž˜ ํ–‰๋ ฌ \(\mathbf{W}\in\mathbb{R}^{m\times n}\)์„ ๋™๊ฒฐํ•œ ์ฑ„๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ์—…๋ฐ์ดํŠธ \(\Delta\mathbf{W}\)๋ฅผ ์ €๊ธ‰ ๋ถ„ํ•ด ํ–‰๋ ฌ, ์ฆ‰ \(\Delta\mathbf{W}=\mathbf{A}\cdot\mathbf{B}^{\top}\), ์—ฌ๊ธฐ์„œ \(\mathbf{A}\in\mathbb{R}^{m\times k}\), \(\mathbf{B}\in\mathbb{R}^{n\times k}\), ์—ฌ๊ธฐ์„œ \(\mathbf{A}\in\mathbb{R}^{m\times k}\), \(\mathbf{B}\in\mathbb{R}^{n\times k}\), ์—ฌ๊ธฐ์„œ \(k\ll\min(m,n)\)์€ ํ™˜์›๋œ ์ˆœ์œ„์ž…๋‹ˆ๋‹ค. LoRA์˜ ๊ฐ€์žฅ ํฐ ์žฅ์ ์€ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์ €์žฅ์†Œ ์‚ฌ์šฉ๋Ÿ‰(_์˜ˆ:_ VRAM)์„ ํฌ๊ฒŒ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ•˜๋‚˜์˜ ํฐ ๋ชจ๋ธ ๋ณต์‚ฌ๋ณธ๋งŒ ์œ ์ง€ํ•˜๋ฉด์„œ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ž‘์—…๋ณ„ ์ €๊ธ‰ ๋ถ„ํ•ด ํ–‰๋ ฌ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ค‘์š”๋„ ์ ์ˆ˜ ๊ธฐ๋ฐ˜ ํ• ๋‹น [406] ๋ฐ ๊ฒ€์ƒ‰ ์—†๋Š” ์ตœ์  ์ˆœ์œ„ ์„ ํƒ [407]๊ณผ ๊ฐ™์ด ๋ณด๋‹ค ์›์น™์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์ˆœ์œ„๋ฅผ ์„ค์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด์„œ๋„ ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
To find the limit of continuously training this fixed-sized LM, we estimate how the model's optimal performance changes with its computation during continuous training. By optimal performance, we mean the loss of training token \(D\) is achieved by \(WSD(D,0.1D)\). With a series of \(D\), the losses will form the optimal loss envelope. Due to uncertainty about the function form of the loss envelope, we try two fitting formulas: (1) exponential: \(L(C)=\alpha e^{-\beta C}+L_{0}\) and (2) power-law: \(L(C)=\beta C^{-\alpha}+L_{0}\). The fitting results for both functions are in Appendix B.2. We find that the power-law form fits better (similar to the Cosine LRS (Kaplan et al., 2020)). In Figure 6, the fitted curve is shown in green dotted lines. To intuitively estimate and comprehend the effect of continuous training such a fixed-sized model, we also trained a 0.17B model with \(WSD(40N,4N)\), which is shown in pink in Figure 6. We can see that a 0.036B model can match the performance of a 0.17B model with an acceptable increase (\(\sim 4\) times) in training compute while saving a lot of inference computation (Sardana and Frankle, 2023) (saving \(\sim 5\) times per inference call), indicating a better inference-compute-optimal setting Sardana and Frankle (2023).
์ด ๊ณ ์ • ํฌ๊ธฐ์˜ LM์„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ํ•œ๊ณ„๋ฅผ ์ฐพ๊ธฐ ์œ„ํ•ด ์ง€์†์ ์ธ ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์˜ ์ตœ์  ์„ฑ๋Šฅ์ด ๊ณ„์‚ฐ์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ๋ณ€ํ•˜๋Š”์ง€ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ ์˜ ์„ฑ๋Šฅ์ด๋ž€ ํŠธ๋ ˆ์ด๋‹ ํ† ํฐ \(D\)์˜ Loss๊ฐ€ \(WSD(D,0.1D)\)์— ์˜ํ•ด ๋‹ฌ์„ฑ๋˜๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ จ์˜ \(D\)๊ฐ€ ์žˆ์œผ๋ฉด Loss๋Š” ์ตœ์ ์˜ Loss ์™ธํ”ผ๋ฅผ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. Loss ์™ธํ”ผ์˜ ํ•จ์ˆ˜ ํ˜•ํƒœ์— ๋Œ€ํ•œ ๋ถˆํ™•์‹ค์„ฑ์œผ๋กœ ์ธํ•ด (1) ์ง€์ˆ˜: \(L(C)=\alpha e^{-\beta C}+L_{0}\) ๋ฐ (2) ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™: \(L(C)=\beta C^{-\alpha}+L_{0}\)์˜ ๋‘ ๊ฐ€์ง€ ํ”ผํŒ… ๊ณต์‹์„ ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ํ•จ์ˆ˜์— ๋Œ€ํ•œ ํ”ผํŒ… ๊ฒฐ๊ณผ๋Š” ๋ถ€๋ก B.2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™ ํ˜•ํƒœ๊ฐ€ ๋” ์ž˜ ๋งž๋Š”๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์ฝ”์‚ฌ์ธ LRS(Kaplan et al., 2020)์™€ ์œ ์‚ฌ). ๊ทธ๋ฆผ 6์— ํ”ผํŒ…๋œ ๊ณก์„ ์ด ์ ์„ ์œผ๋กœ ๋…น์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ ์ • ํฌ๊ธฐ์˜ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ํšจ๊ณผ๋ฅผ ์ง๊ด€์ ์œผ๋กœ ์ถ”์ •ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 6์˜ ๋ถ„ํ™์ƒ‰์— ํ‘œ์‹œ๋œ \(WSD(40N,4N)\)์œผ๋กœ 0.17B ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. 0.036B ๋ชจ๋ธ์ด ํ›ˆ๋ จ ๊ณ„์‚ฐ๋Ÿ‰์„ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์ˆ˜์ค€์œผ๋กœ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด์„œ(์•ฝ 4๋ฐฐ) 0.17B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋”ฐ๋ผ์žก์„ ์ˆ˜ ์žˆ๊ณ , ์ถ”๋ก  ๊ณ„์‚ฐ๋Ÿ‰์„ ๋งŽ์ด ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Sardana and Frankle, 2023)(์ถ”๋ก  ํ˜ธ๋‹น ์•ฝ 5๋ฐฐ ์ ˆ์•ฝ)๋กœ ์ถ”๋ก  ๊ณ„์‚ฐ ์ตœ์ ํ™”๊ฐ€ ๋” ์ž˜ ์ด๋ฃจ์–ด์กŒ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค Sardana and Frankle (2023).
์ด ๊ณ ์ • ํฌ๊ธฐ์˜ LM์„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ํ•œ๊ณ„๋ฅผ ์ฐพ๊ธฐ ์œ„ํ•ด ์ง€์†์ ์ธ ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์˜ ์ตœ์  ์„ฑ๋Šฅ์ด ๊ณ„์‚ฐ์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ๋ณ€ํ•˜๋Š”์ง€ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ ์˜ ์„ฑ๋Šฅ์ด๋ž€ ํŠธ๋ ˆ์ด๋‹ ํ† ํฐ \(D\)์˜ ์†์‹ค์ด \(WSD(D,0.1D)\)์— ์˜ํ•ด ๋‹ฌ์„ฑ๋˜๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ จ์˜ \(D\)๊ฐ€ ์žˆ์œผ๋ฉด ์†์‹ค์€ ์ตœ์ ์˜ ์†์‹ค ์™ธํ”ผ๋ฅผ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. ์†์‹ค ์™ธํ”ผ์˜ ํ•จ์ˆ˜ ํ˜•ํƒœ์— ๋Œ€ํ•œ ๋ถˆํ™•์‹ค์„ฑ์œผ๋กœ ์ธํ•ด (1) ์ง€์ˆ˜: \(L(C)=\์•ŒํŒŒ e^{-\๋ฒ ํƒ€ C}+L_{0}\) ๋ฐ (2) ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™: \(L(C)=\๋ฒ ํƒ€ C^{-\์•ŒํŒŒ}+L_{0}\)์˜ ๋‘ ๊ฐ€์ง€ ํ”ผํŒ… ๊ณต์‹์„ ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ํ•จ์ˆ˜์— ๋Œ€ํ•œ ํ”ผํŒ… ๊ฒฐ๊ณผ๋Š” ๋ถ€๋ก B.2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™ ํ˜•ํƒœ๊ฐ€ ๋” ์ž˜ ๋งž๋Š”๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์ฝ”์‚ฌ์ธ LRS(Kaplan et al., 2020)์™€ ์œ ์‚ฌ). ๊ทธ๋ฆผ 6์— ํ”ผํŒ…๋œ ๊ณก์„ ์ด ์ ์„ ์œผ๋กœ ๋…น์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ ์ • ํฌ๊ธฐ์˜ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ํšจ๊ณผ๋ฅผ ์ง๊ด€์ ์œผ๋กœ ์ถ”์ •ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 6์˜ ๋ถ„ํ™์ƒ‰์— ํ‘œ์‹œ๋œ \(WSD(40N,4N)\)์œผ๋กœ 0.17B ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. 0.036B ๋ชจ๋ธ์ด ํ›ˆ๋ จ ๊ณ„์‚ฐ๋Ÿ‰์„ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์ˆ˜์ค€์œผ๋กœ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด์„œ(์•ฝ 4๋ฐฐ) 0.17B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋”ฐ๋ผ์žก์„ ์ˆ˜ ์žˆ๊ณ , ์ถ”๋ก  ๊ณ„์‚ฐ๋Ÿ‰์„ ๋งŽ์ด ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Sardana and Frankle, 2023)(์ถ”๋ก  ํ˜ธ๋‹น ์•ฝ 5๋ฐฐ ์ ˆ์•ฝ)๋กœ ์ถ”๋ก  ๊ณ„์‚ฐ ์ตœ์ ํ™”๊ฐ€ ๋” ์ž˜ ์ด๋ฃจ์–ด์กŒ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค Sardana and Frankle (2023).
In this paper, we introduced a simple yet effective training method to help remedy the reversal curse in LLMs. Our reverse training works by first segmenting the input sequence into chunks and then reversing the ordering of chunks, but leaves the word-ordering in each chunk intact. A chunk can be a token, a word, an entity name, or a random number of tokens. The model is then trained on both the original sequences, and this reversed data. We evaluated on a symbolic reverse task and a reversing biography task that both demonstrated the necessity of preserving word-ordering within chunks. Next, we applied our reverse training to the realistic setting of LLM pre-training, which minimized the reversal curse on real-world knowledge. Evaluations on common benchmark tasks reveal that reverse training (particularly at the word level) during pre-training does not interfere with the forward prediction ability of LLMs, and actually improves metrics in the data-bound (rather than compute-bound) setting compared to standard training. When our method is applied to finetuning on fictitious facts, prediction accuracy rose from 0% to 70-100%. The reversal curse is a serious flaw in how LLMs acquire knowledge and reverse training opens a new promising direction in resolving it.
์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์˜ ์—ญ์ „ ์ €์ฃผ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ญ ํ›ˆ๋ จ์€ ๋จผ์ € ์ž…๋ ฅ ์‹œํ€€์Šค๋ฅผ ์ฒญํฌ๋กœ ๋ถ„ํ• ํ•œ ๋‹ค์Œ ์ฒญํฌ์˜ ์ˆœ์„œ๋ฅผ ๋ฐ˜๋Œ€๋กœ ํ•˜๋˜ ๊ฐ ์ฒญํฌ์˜ ๋‹จ์–ด ์ˆœ์„œ๋Š” ๊ทธ๋Œ€๋กœ ์œ ์ง€ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ฒญํฌ๋Š” ํ† ํฐ, ๋‹จ์–ด, ์—”ํ‹ฐํ‹ฐ ์ด๋ฆ„ ๋˜๋Š” ์ž„์˜์˜ ํ† ํฐ ์ˆ˜์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์€ ์›๋ณธ ์‹œํ€€์Šค์™€ ์ด ๋ฐ˜์ „๋œ ๋ฐ์ดํ„ฐ ๋ชจ๋‘์— ๋Œ€ํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹จ์–ด ์ˆœ์„œ๋ฅผ ์ฒญํฌ ๋‚ด์—์„œ ๋ณด์กดํ•ด์•ผ ํ•  ํ•„์š”์„ฑ์„ ์ž…์ฆํ•œ ๊ธฐํ˜ธ์  ์—ญ๋ฐฉํ–ฅ ์ž‘์—…๊ณผ ์—ญ๋ฐฉํ–ฅ ์ „๊ธฐ ์ž‘์—…์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์‹ค์ œ ์ง€์‹์— ๋Œ€ํ•œ ์—ญ์ „ ์ €์ฃผ๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” LLM pre-training์˜ ํ˜„์‹ค์ ์ธ ์„ค์ •์— ์—ญ ํ›ˆ๋ จ์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ฒฐ๊ณผ, pre-training ์ค‘ ์—ญ ํ›ˆ๋ จ(ํŠนํžˆ ๋‹จ์–ด ์ˆ˜์ค€์—์„œ)์ด LLM์˜ ์ˆœ๋ฐฉํ–ฅ ์˜ˆ์ธก ๋Šฅ๋ ฅ์„ ๋ฐฉํ•ดํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์‹ค์ œ๋กœ ํ‘œ์ค€ ํ›ˆ๋ จ์— ๋น„ํ•ด ๋ฐ์ดํ„ฐ ๋ฐ”์šด๋“œ(๊ณ„์‚ฐ ๋ฐ”์šด๋“œ๊ฐ€ ์•„๋‹Œ) ์„ค์ •์—์„œ ๋ฉ”ํŠธ๋ฆญ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ๊ฐ€์ƒ์˜ ์‚ฌ์‹ค์— ๋Œ€ํ•œ Fine-tuning์— ์ ์šฉํ•˜๋ฉด ์˜ˆ์ธก ์ •ํ™•๋„๊ฐ€ 0%์—์„œ 70~100%๊นŒ์ง€ ์ƒ์Šนํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ญ์ „ ์ €์ฃผ๋Š” LLM์ด ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐฉ์‹์— ์‹ฌ๊ฐํ•œ ๊ฒฐํ•จ์ด ์žˆ์œผ๋ฉฐ, ์—ญ ํ›ˆ๋ จ์€ ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ์ƒˆ๋กœ์šด ๊ฐ€๋Šฅ์„ฑ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์˜ ์—ญ์ „ ์ €์ฃผ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ญ ํ›ˆ๋ จ์€ ๋จผ์ € ์ž…๋ ฅ ์‹œํ€€์Šค๋ฅผ ์ฒญํฌ๋กœ ๋ถ„ํ• ํ•œ ๋‹ค์Œ ์ฒญํฌ์˜ ์ˆœ์„œ๋ฅผ ๋ฐ˜๋Œ€๋กœ ํ•˜๋˜ ๊ฐ ์ฒญํฌ์˜ ๋‹จ์–ด ์ˆœ์„œ๋Š” ๊ทธ๋Œ€๋กœ ์œ ์ง€ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ฒญํฌ๋Š” ํ† ํฐ, ๋‹จ์–ด, ์—”ํ‹ฐํ‹ฐ ์ด๋ฆ„ ๋˜๋Š” ์ž„์˜์˜ ํ† ํฐ ์ˆ˜์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์€ ์›๋ณธ ์‹œํ€€์Šค์™€ ์ด ๋ฐ˜์ „๋œ ๋ฐ์ดํ„ฐ ๋ชจ๋‘์— ๋Œ€ํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹จ์–ด ์ˆœ์„œ๋ฅผ ์ฒญํฌ ๋‚ด์—์„œ ๋ณด์กดํ•ด์•ผ ํ•  ํ•„์š”์„ฑ์„ ์ž…์ฆํ•œ ๊ธฐํ˜ธ์  ์—ญ๋ฐฉํ–ฅ ์ž‘์—…๊ณผ ์—ญ๋ฐฉํ–ฅ ์ „๊ธฐ ์ž‘์—…์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์‹ค์ œ ์ง€์‹์— ๋Œ€ํ•œ ์—ญ์ „ ์ €์ฃผ๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” LLM ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ํ˜„์‹ค์ ์ธ ์„ค์ •์— ์—ญ ํ›ˆ๋ จ์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ฒฐ๊ณผ, ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘ ์—ญ ํ›ˆ๋ จ(ํŠนํžˆ ๋‹จ์–ด ์ˆ˜์ค€์—์„œ)์ด LLM์˜ ์ˆœ๋ฐฉํ–ฅ ์˜ˆ์ธก ๋Šฅ๋ ฅ์„ ๋ฐฉํ•ดํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์‹ค์ œ๋กœ ํ‘œ์ค€ ํ›ˆ๋ จ์— ๋น„ํ•ด ๋ฐ์ดํ„ฐ ๋ฐ”์šด๋“œ(๊ณ„์‚ฐ ๋ฐ”์šด๋“œ๊ฐ€ ์•„๋‹Œ) ์„ค์ •์—์„œ ๋ฉ”ํŠธ๋ฆญ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ๊ฐ€์ƒ์˜ ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ •์— ์ ์šฉํ•˜๋ฉด ์˜ˆ์ธก ์ •ํ™•๋„๊ฐ€ 0%์—์„œ 70~100%๊นŒ์ง€ ์ƒ์Šนํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ญ์ „ ์ €์ฃผ๋Š” LLM์ด ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐฉ์‹์— ์‹ฌ๊ฐํ•œ ๊ฒฐํ•จ์ด ์žˆ์œผ๋ฉฐ, ์—ญ ํ›ˆ๋ จ์€ ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ์ƒˆ๋กœ์šด ๊ฐ€๋Šฅ์„ฑ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
Validation Loss ComparisonThe results in Fig. 4 (top and bottom) show the evolution of the validation loss during continual pre-training on the respective \(\mathcal{D}_{1}\) datasets. Table 2 reports the average final validation loss for each of these models. The final loss is averaged over the last 100 iterations of training sampled at intervals of 10 iterations. We consistently observe across both distribution shifts that even the lowest tested replay of 1% significantly reduces forgetting on Pile compared to the no-replay baselines. This effect is more pronounced in the strong-shift scenario due to the larger amount of forgetting in this setting. We observe little impact on downstream performance for 1%,5%, and 10% replay when compared to the 0% baseline, showing that the forgetting benefits of replay come at little cost in our setting. However, when using an extreme amount of replay (50%), we observe that the model adapts less to \(\mathcal{D}_{0}\). Interestingly, for both datasets, the 50% replay models attain or surpass the final average validation performance of the baseline training on \(\mathcal{D}_{1}\cup\mathcal{D}_{0}\). This is curious as these model have seen 150B and 100B fewer tokens of \(\mathcal{D}_{1}\) than their respective baselines.
์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss ๋น„๊ต๊ทธ๋ฆผ 4์˜ ๊ฒฐ๊ณผ(์œ„์ชฝ๊ณผ ์•„๋ž˜์ชฝ)๋Š” ๊ฐ๊ฐ์˜ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss์˜ ์ง„ํ™”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ์ด๋Ÿฌํ•œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ตœ์ข… ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss์˜ ํ‰๊ท ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… Loss๋Š” 10ํšŒ ๊ฐ„๊ฒฉ์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋งˆ์ง€๋ง‰ 100ํšŒ ํ›ˆ๋ จ ๋ฐ˜๋ณต์— ๋Œ€ํ•ด ํ‰๊ท ํ™”๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ํ…Œ์ŠคํŠธ๋œ ๊ฐ€์žฅ ๋‚ฎ์€ ๋ฆฌํ”Œ๋ ˆ์ด์ธ 1%๋„ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ํ•˜์ง€ ์•Š์€ baseline๋ณด๋‹ค Pile์—์„œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์ƒ๋‹นํžˆ ์ค„์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์ผ๊ด€๋˜๊ฒŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํšจ๊ณผ๋Š” ์ด ์„ค์ •์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ์žŠ์–ด๋ฒ„๋ฆผ์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ•ํ•œ ์‹œํ”„ํŠธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋” ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค. 0% ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•  ๋•Œ 1%, 5%, 10% ๋ฆฌํ”Œ๋ ˆ์ด์˜ ๊ฒฝ์šฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๊ฑฐ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•„ ๋ฆฌํ”Œ๋ ˆ์ด์˜ ์žŠ์–ด๋ฒ„๋ฆผ ์ด์ ์ด ์šฐ๋ฆฌ ์„ค์ •์—์„œ ๊ฑฐ์˜ ๋Œ€๊ฐ€๋ฅผ ์น˜๋ฅด์ง€ ์•Š์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทน๋‹จ์ ์ธ ์–‘์˜ ๋ฆฌํ”Œ๋ ˆ์ด(50%)๋ฅผ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\)์— ๋œ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์—์„œ 50% ๋ฆฌํ”Œ๋ ˆ์ด ๋ชจ๋ธ์€ \(\mathcal{D}_{1}\cup\(\mathcal{D}_{0}\)์— ๋Œ€ํ•œ baseline ํ›ˆ๋ จ์˜ ์ตœ์ข… ํ‰๊ท  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋ชจ๋ธ์ด ๊ฐ ๊ธฐ์ค€๋ณด๋‹ค 150B์™€ 100B์˜ ํ† ํฐ์„ ๋” ์ ๊ฒŒ ๋ณด์•˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ ๊ธฐํ•œ ์ผ์ž…๋‹ˆ๋‹ค.
์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค ๋น„๊ต๊ทธ๋ฆผ 4์˜ ๊ฒฐ๊ณผ(์œ„์ชฝ๊ณผ ์•„๋ž˜์ชฝ)๋Š” ๊ฐ๊ฐ์˜ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์˜ ์ง„ํ™”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ์ด๋Ÿฌํ•œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ตœ์ข… ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์˜ ํ‰๊ท ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์†์‹ค์€ 10ํšŒ ๊ฐ„๊ฒฉ์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋งˆ์ง€๋ง‰ 100ํšŒ ํ›ˆ๋ จ ๋ฐ˜๋ณต์— ๋Œ€ํ•ด ํ‰๊ท ํ™”๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ํ…Œ์ŠคํŠธ๋œ ๊ฐ€์žฅ ๋‚ฎ์€ ๋ฆฌํ”Œ๋ ˆ์ด์ธ 1%๋„ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ํ•˜์ง€ ์•Š์€ ๊ธฐ๋ณธ์„ ๋ณด๋‹ค Pile์—์„œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์ƒ๋‹นํžˆ ์ค„์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์ผ๊ด€๋˜๊ฒŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํšจ๊ณผ๋Š” ์ด ์„ค์ •์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ์žŠ์–ด๋ฒ„๋ฆผ์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ•ํ•œ ์‹œํ”„ํŠธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋” ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค. 0% ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•  ๋•Œ 1%, 5%, 10% ๋ฆฌํ”Œ๋ ˆ์ด์˜ ๊ฒฝ์šฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๊ฑฐ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•„ ๋ฆฌํ”Œ๋ ˆ์ด์˜ ์žŠ์–ด๋ฒ„๋ฆผ ์ด์ ์ด ์šฐ๋ฆฌ ์„ค์ •์—์„œ ๊ฑฐ์˜ ๋Œ€๊ฐ€๋ฅผ ์น˜๋ฅด์ง€ ์•Š์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทน๋‹จ์ ์ธ ์–‘์˜ ๋ฆฌํ”Œ๋ ˆ์ด(50%)๋ฅผ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\)์— ๋œ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์—์„œ 50% ๋ฆฌํ”Œ๋ ˆ์ด ๋ชจ๋ธ์€ \(\mathcal{D}_{1}\cup\mathcal{D}_{0}\)์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ํ›ˆ๋ จ์˜ ์ตœ์ข… ํ‰๊ท  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋ชจ๋ธ์ด ๊ฐ ๊ธฐ์ค€๋ณด๋‹ค 150B์™€ 100B์˜ ํ† ํฐ์„ ๋” ์ ๊ฒŒ ๋ณด์•˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ ๊ธฐํ•œ ์ผ์ž…๋‹ˆ๋‹ค.
\(\bullet\)_Effective RL training_. As the RL training process tends to be unstable and hyper-parameter sensitive, it is suggested that the language model should be well supervised fine-tuned before RL training, so as to reaching a good model capacity. A commonly-used way is to fine-tune the LLM on its best outputs of the prompts (referred to as _rejection sampling_ or _best-of-\(N\)_) from the alignment dataset until convergence before RL. Given a prompt, the LLM would first produce \(N\) outputs via the sampling algorithm, and then the best candidate from the model will be selected by the reward model for learning. After fine-tuning the LLM on the best samples until convergence, the RL process will be performed to further improve the performance. LLMaA 2 [99] has successively trained five versions of RLHF models, where the LLM has been progressively improved with the improvement of the reward models. In this way, the collected prompts and annotations of human preference data can better reflect the issues of the current model checkpoint, thus making special tuning to address these issues. In addition, LLMaA 2 also adds samples from prior iterations into the subsequent ones, to alleviate the possible capacity regression issue during iterative optimization.
\(\bullet\)_ํšจ๊ณผ์ ์ธ RL ํ›ˆ๋ จ_. RL ํ›ˆ๋ จ ๊ณผ์ •์€ ๋ถˆ์•ˆ์ •ํ•˜๊ณ  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์— ๋ฏผ๊ฐํ•œ ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ, RL ํ›ˆ๋ จ ์ „์— ์–ธ์–ด ๋ชจ๋ธ์„ ์ž˜ supervised Fine-tuningํ•˜์—ฌ ์ข‹์€ ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ๋„๋‹ฌํ•˜๋„๋ก ๋ฏธ๋ฆฌ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์€ RL ์ „์— ์ •ํ•ฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์žฅ ์ข‹์€ ํ”„๋กฌํ”„ํŠธ์˜ ์ตœ์  ์ถœ๋ ฅ(๊ฑฐ๋ถ€ ์ƒ˜ํ”Œ๋ง ๋˜๋Š” _best-of-\(N\)\_์ด๋ผ๊ณ  ํ•จ)์— ๋Œ€ํ•ด LLM์„ Fine-tuningํ•˜์—ฌ ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ Fine-tuningํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด LLM์€ ๋จผ์ € ์ƒ˜ํ”Œ๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด \(N\)๊ฐœ์˜ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ๋ณด์ƒ ๋ชจ๋ธ์— ์˜ํ•ด ๋ชจ๋ธ์—์„œ ๊ฐ€์žฅ ์ข‹์€ ํ›„๋ณด๊ฐ€ ์„ ํƒ๋˜์–ด ํ•™์Šต์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ์ตœ์ ์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด LLM์„ Fine-tuningํ•œ ํ›„ RL ํ”„๋กœ์„ธ์Šค๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ๋”์šฑ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. LLMaA 2 [99]๋Š” 5๊ฐœ ๋ฒ„์ „์˜ RLHF ๋ชจ๋ธ์„ ์—ฐ์†์ ์œผ๋กœ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ LLM์€ ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ฐœ์„ ์— ๋”ฐ๋ผ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ˆ˜์ง‘๋œ ํ”„๋กฌํ”„ํŠธ์™€ ์ธ๊ฐ„ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์˜ ์ฃผ์„์€ ํ˜„์žฌ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์˜ ๋ฌธ์ œ๋ฅผ ๋” ์ž˜ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ํŠน๋ณ„ํ•œ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LLMaA 2๋Š” ๋ฐ˜๋ณต ํ›ˆ๋ จ ์ค‘์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์šฉ๋Ÿ‰ ํ‡ด๋ณด ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ๋ฐ˜๋ณต์˜ ์ƒ˜ํ”Œ์„ ํ›„์† ๋ฐ˜๋ณต์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_ํšจ๊ณผ์ ์ธ RL ํ›ˆ๋ จ_. RL ํ›ˆ๋ จ ๊ณผ์ •์€ ๋ถˆ์•ˆ์ •ํ•˜๊ณ  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์— ๋ฏผ๊ฐํ•œ ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ, RL ํ›ˆ๋ จ ์ „์— ์–ธ์–ด ๋ชจ๋ธ์„ ์ž˜ ์Šˆํผ๋ฐ”์ด์ง•ํ•˜์—ฌ ์ข‹์€ ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ๋„๋‹ฌํ•˜๋„๋ก ๋ฏธ๋ฆฌ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์€ RL ์ „์— ์ •ํ•ฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์žฅ ์ข‹์€ ํ”„๋กฌํ”„ํŠธ์˜ ์ตœ์  ์ถœ๋ ฅ(๊ฑฐ๋ถ€ ์ƒ˜ํ”Œ๋ง ๋˜๋Š” _best-of-\(N\)\_์ด๋ผ๊ณ  ํ•จ)์— ๋Œ€ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด LLM์€ ๋จผ์ € ์ƒ˜ํ”Œ๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด \(N\)๊ฐœ์˜ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ๋ณด์ƒ ๋ชจ๋ธ์— ์˜ํ•ด ๋ชจ๋ธ์—์„œ ๊ฐ€์žฅ ์ข‹์€ ํ›„๋ณด๊ฐ€ ์„ ํƒ๋˜์–ด ํ•™์Šต์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ์ตœ์ ์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ํ›„ RL ํ”„๋กœ์„ธ์Šค๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ๋”์šฑ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. LLMaA 2 [99]๋Š” 5๊ฐœ ๋ฒ„์ „์˜ RLHF ๋ชจ๋ธ์„ ์—ฐ์†์ ์œผ๋กœ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ LLM์€ ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ฐœ์„ ์— ๋”ฐ๋ผ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ˆ˜์ง‘๋œ ํ”„๋กฌํ”„ํŠธ์™€ ์ธ๊ฐ„ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์˜ ์ฃผ์„์€ ํ˜„์žฌ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์˜ ๋ฌธ์ œ๋ฅผ ๋” ์ž˜ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ํŠน๋ณ„ํ•œ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LLMaA 2๋Š” ๋ฐ˜๋ณต ํ›ˆ๋ จ ์ค‘์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์šฉ๋Ÿ‰ ํ‡ด๋ณด ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ๋ฐ˜๋ณต์˜ ์ƒ˜ํ”Œ์„ ํ›„์† ๋ฐ˜๋ณต์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
ExperimentWe demonstrate that the proposed definition of \(\textsc{CnstScore}(\tau_{\text{2H}},\tau_{\text{1H}})\) is a reasonable proxy of the utilization of what the LLM knows about the bridge entity's attribute - the latent recall of its answer to \(\tau_{\text{1H}}\) - with indirect evidence. If the information to reason with is given as part of the input, e.g., if the given prompt is _"The singer of 'Supersition' is Stevie Wonder. The mother of Stevie Wonder is named Lula. The mother of the singer of 'Supersition' is"_, the LLM would not need to internally perform the multi-hop reasoning to refer to what its output to the one-hop prompt _"The mother of Stevie Wonder is"_ is, but just copy the answer from the input. Therefore, \(\textsc{CnstScore}\) of such a case will be lower than the case where the LLM needs to internally figure out what its answer to the one-hop prompt given the hint of who the descriptive mention refers to, e.g., _"The singer of 'Superstition' is Stevie Wonder. The mother of the singer of 'Superstition' is"_. Therefore, to check whether this is the case, we compare \(\textsc{CnstScore}\) computed with the several Chain-of-Thought (CoT) style prompts \(\tau^{\prime}\), i.e., \(\textsc{CnstScore}(\tau^{\prime},\tau_{\text{1H}})\).
์‹คํ—˜์šฐ๋ฆฌ๋Š” ์ œ์•ˆ๋œ \(\textsc{CnstScore}(\tau_{\text{2H}},\tau_{\text{1H}})\)์˜ ์ •์˜๊ฐ€ ๊ฐ„์ ‘์ ์ธ ์ฆ๊ฑฐ๋ฅผ ํ†ตํ•ด LLM์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๊ฒƒ, ์ฆ‰ \(\tau_{\text{1H}}\)์— ๋Œ€ํ•œ ๋‹ต๋ณ€์˜ ์ž ์žฌ์  ํšŒ์ƒ์— ๋Œ€ํ•œ ํ•ฉ๋ฆฌ์ ์ธ ๋Œ€๋ฆฌ์ž์ž„์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ์ฃผ์–ด์ง„๋‹ค๋ฉด, ์˜ˆ๋ฅผ ๋“ค์–ด ์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๊ฐ€ _"์ˆ˜ํผ์‹œ์…˜"์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ž…๋‹ˆ๋‹ค. ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š” ๋ฃฐ๋ผ์ž…๋‹ˆ๋‹ค. '์ˆ˜ํผ์‹œ์…˜'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์ด๋ฉด LLM์€ ์ถœ๋ ฅ์— ๋Œ€ํ•ด ๋‚ด๋ถ€์ ์œผ๋กœ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. 1ํ™‰ ํ”„๋กฌํ”„ํŠธ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š”"_์— ๋Œ€ํ•œ ์ถœ๋ ฅ์ด ๋ฌด์—‡์ธ์ง€, ์ž…๋ ฅ์—์„œ ๋‹ต์„ ๋ณต์‚ฌํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ์˜ \(\textsc{CnstScore}\)๋Š” LLM์ด ์„ค๋ช…์  ์–ธ๊ธ‰์ด ๋ˆ„๊ตฌ๋ฅผ ๊ฐ€๋ฆฌํ‚ค๋Š”์ง€์— ๋Œ€ํ•œ ํžŒํŠธ๊ฐ€ ์ฃผ์–ด์ง„ 1ํ™‰ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋‹ต์„ ๋‚ด๋ถ€์ ์œผ๋กœ ์•Œ์•„๋‚ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ, ์˜ˆ๋ฅผ ๋“ค์–ด _"๋ฏฟ์Œ์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ž…๋‹ˆ๋‹ค. '๋ฏฟ์Œ'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"_. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ์ธ์ง€ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡ ๊ฐ€์ง€ Chain-of-Thought (CoT) ์Šคํƒ€์ผ์˜ ํ”„๋กฌํ”„ํŠธ \(\tau^{\prime}\), ์ฆ‰ \(\textsc{CnstScore}(\tau^{\prime},\tau_{\text{1H}})\).
์‹คํ—˜์šฐ๋ฆฌ๋Š” ์ œ์•ˆ๋œ \(\textsc{CnstScore}(\tau_{\text{2H}},\tau_{\text{1H}})\)์˜ ์ •์˜๊ฐ€ ๊ฐ„์ ‘์ ์ธ ์ฆ๊ฑฐ๋ฅผ ํ†ตํ•ด LLM์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๊ฒƒ, ์ฆ‰ \(\tau_{\text{1H}}\)์— ๋Œ€ํ•œ ๋‹ต๋ณ€์˜ ์ž ์žฌ์  ํšŒ์ƒ์— ๋Œ€ํ•œ ํ•ฉ๋ฆฌ์ ์ธ ๋Œ€๋ฆฌ์ž์ž„์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ์ฃผ์–ด์ง„๋‹ค๋ฉด, ์˜ˆ๋ฅผ ๋“ค์–ด ์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๊ฐ€ _"์ˆ˜ํผ์‹œ์…˜"์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ž…๋‹ˆ๋‹ค. ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š” ๋ฃฐ๋ผ์ž…๋‹ˆ๋‹ค. '์ˆ˜ํผ์‹œ์…˜'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์ด๋ฉด LLM์€ ์ถœ๋ ฅ์— ๋Œ€ํ•ด ๋‚ด๋ถ€์ ์œผ๋กœ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. 1ํ™‰ ํ”„๋กฌํ”„ํŠธ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š”"_์— ๋Œ€ํ•œ ์ถœ๋ ฅ์ด ๋ฌด์—‡์ธ์ง€, ์ž…๋ ฅ์—์„œ ๋‹ต์„ ๋ณต์‚ฌํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ์˜ \(\textsc{CnstScore}\)๋Š” LLM์ด ์„ค๋ช…์  ์–ธ๊ธ‰์ด ๋ˆ„๊ตฌ๋ฅผ ๊ฐ€๋ฆฌํ‚ค๋Š”์ง€์— ๋Œ€ํ•œ ํžŒํŠธ๊ฐ€ ์ฃผ์–ด์ง„ 1ํ™‰ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋‹ต์„ ๋‚ด๋ถ€์ ์œผ๋กœ ์•Œ์•„๋‚ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ, ์˜ˆ๋ฅผ ๋“ค์–ด _"๋ฏฟ์Œ์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ž…๋‹ˆ๋‹ค. '๋ฏฟ์Œ'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"_. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ์ธ์ง€ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡ ๊ฐ€์ง€ ์ƒ๊ฐ์˜ ์—ฐ์‡„(CoT) ์Šคํƒ€์ผ์˜ ํ”„๋กฌํ”„ํŠธ \(\tau^{\prime}\), ์ฆ‰ \(\textsc{CnstScore}(\tau^{\prime},\tau_{\text{1H}})\).
**RLHF System.** The RLHF system mainly comprises three key components: a pre-trained LM to be aligned, a reward model learning from human feedback, and a RL algorithm training the LM. Specifically, the _pre-trained LM_ is typically a generative model that is initialized with existing pre-trained LM parameters. For example, OpenAI uses 175B GPT-3 for its first popular RLHF model, InstructGPT [66], and DeepMind uses the 280 billion parameter model Go-pher [64] for its GopherCite model [370]. Further, the _reward model (RM)_ provides (learned) guidance signals that reflect human preferences for the text generated by the LM, usually in the form of a scalar value. The reward model can take on two forms: a fine-tuned LM or a LM trained de novo using human preference data. Existing work typically employs reward models having a parameter scale different from that of the aligned LM [66, 370]. For example, OpenAI uses 6B GPT-3 and DeepMind uses 7B Gopher as the reward model, respectively. Finally, to optimize the pre-trained LM using the signal from the reward model, a specific _RL algorithm_ is designed for large-scale model tuning. Specifically, Proximal Policy Optimization (PPO) [128] is a widely used RL algorithm for alignment in existing work [66, 116, 370].
**RLHF ์‹œ์Šคํ…œ.** RLHF ์‹œ์Šคํ…œ์€ ์ฃผ๋กœ ์ •๋ ฌํ•  pre-trained LM, ์‚ฌ๋žŒ์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ•™์Šตํ•˜๋Š” reward model, LM์„ ํ•™์Šตํ•˜๋Š” RL algorithm์˜ ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, _pre-trained LM_์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ์กด pre-trained LM ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ์ดˆ๊ธฐํ™”๋˜๋Š” ์ƒ์„ฑ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI๋Š” ์ฒซ ๋ฒˆ์งธ ์ธ๊ธฐ RLHF ๋ชจ๋ธ์ธ InstructGPT [66]์— 175B GPT-3๋ฅผ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, DeepMind๋Š” 2800์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์ธ Go-pher [64]๋ฅผ GopherCite ๋ชจ๋ธ [370]์— ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, _reward model (RM)_์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์Šค์นผ๋ผ ๊ฐ’์˜ ํ˜•ํƒœ๋กœ LM์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” (ํ•™์Šต๋œ) ์•ˆ๋‚ด ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. reward model์€ Fine-tuned LM ๋˜๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒˆ๋กœ ํ•™์Šต๋œ LM์˜ ๋‘ ๊ฐ€์ง€ ํ˜•ํƒœ๋ฅผ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ •๋ ฌ๋œ LM๊ณผ ๋‹ค๋ฅธ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๋ฅผ ๊ฐ€์ง„ reward model์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค[66, 370]. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI๋Š” 6B GPT-3๋ฅผ, DeepMind๋Š” 7B Gopher๋ฅผ ๊ฐ๊ฐ reward model๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, reward model์˜ ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ pre-trained LM์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ํŠœ๋‹์„ ์œ„ํ•ด ํŠน์ • _RL algorithm_์ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๊ธฐ์กด ์ž‘์—… [66, 116, 370]์—์„œ ์ •๋ ฌ์„ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” RL algorithm์€ Proximal Policy Optimization (PPO) [128]์ž…๋‹ˆ๋‹ค.
**RLHF ์‹œ์Šคํ…œ.** RLHF ์‹œ์Šคํ…œ์€ ์ฃผ๋กœ ์ •๋ ฌํ•  ์‚ฌ์ „ ํ•™์Šต๋œ LM, ์‚ฌ๋žŒ์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ•™์Šตํ•˜๋Š” ๋ณด์ƒ ๋ชจ๋ธ, LM์„ ํ•™์Šตํ•˜๋Š” RL ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, _์‚ฌ์ „ ํ•™์Šต๋œ LM_์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ์กด ์‚ฌ์ „ ํ•™์Šต๋œ LM ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ์ดˆ๊ธฐํ™”๋˜๋Š” ์ƒ์„ฑ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI๋Š” ์ฒซ ๋ฒˆ์งธ ์ธ๊ธฐ RLHF ๋ชจ๋ธ์ธ InstructGPT [66]์— 175B GPT-3๋ฅผ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, DeepMind๋Š” 2800์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์ธ Go-pher [64]๋ฅผ GopherCite ๋ชจ๋ธ [370]์— ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, _๋ณด์ƒ ๋ชจ๋ธ(RM)_์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์Šค์นผ๋ผ ๊ฐ’์˜ ํ˜•ํƒœ๋กœ LM์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” (ํ•™์Šต๋œ) ์•ˆ๋‚ด ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ƒ ๋ชจ๋ธ์€ ๋ฏธ์„ธ ์กฐ์ •๋œ LM ๋˜๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒˆ๋กœ ํ•™์Šต๋œ LM์˜ ๋‘ ๊ฐ€์ง€ ํ˜•ํƒœ๋ฅผ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ •๋ ฌ๋œ LM๊ณผ ๋‹ค๋ฅธ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๋ฅผ ๊ฐ€์ง„ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค[66, 370]. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI๋Š” 6B GPT-3๋ฅผ, DeepMind๋Š” 7B Gopher๋ฅผ ๊ฐ๊ฐ ๋ณด์ƒ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ณด์ƒ ๋ชจ๋ธ์˜ ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ํŠœ๋‹์„ ์œ„ํ•ด ํŠน์ • _RL ์•Œ๊ณ ๋ฆฌ์ฆ˜_์ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๊ธฐ์กด ์ž‘์—… [66, 116, 370]์—์„œ ์ •๋ ฌ์„ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” RL ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๊ทผ์ ‘ ์ •์ฑ… ์ตœ์ ํ™”(PPO) [128]์ž…๋‹ˆ๋‹ค.
Despite their remarkable capabilities, large language models (LLMs) often produce responses containing factual inaccuracies due to their sole reliance on the parametric knowledge they encapsulate. Retrieval-Augmented Generation (RAG), an ad hoc approach that augments LMs with retrieval of relevant knowledge, decreases such issues. However, indiscriminately retrieving and incorporating a fixed number of retrieved passages, regardless of whether retrieval is necessary, or passages are relevant, diminishes LM versatility or can lead to unhelpful response generation. We introduce a new framework called **Self-Reflective Retrieval-Augmented Generation (Self-RAG)** that enhances an LM's quality and factuality through retrieval and self-reflection. Our framework trains a single arbitrary LM that adaptively retrieves passages on-demand, and generates and reflects on retrieved passages and its own generations using special tokens, called _reflection_ tokens. Generating reflection tokens makes the LM controllable during the inference phase, enabling it to tailor its behavior to diverse task requirements. Experiments show that Self-RAG (7B and 13B parameters) significantly outperforms state-of-the-art LLMs and retrieval-augmented models on a diverse set of tasks.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ๋†€๋ผ์šด ๊ธฐ๋Šฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์ง€์‹์—๋งŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์‹ค์— ๋ถ€์ •ํ™•ํ•œ ๋‚ด์šฉ์ด ํฌํ•จ๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ด€๋ จ ์ง€์‹ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด LLM์„ ๋ณด๊ฐ•ํ•˜๋Š” ์ž„์‹œ๋ฐฉํŽธ ์ ‘๊ทผ ๋ฐฉ์‹์ธ Retrieval-Augmented Generation (RAG)์€ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰์ด ํ•„์š”ํ•œ์ง€ ์—ฌ๋ถ€๋‚˜ ์ง€๋ฌธ์ด ๊ด€๋ จ์„ฑ์ด ์žˆ๋Š”์ง€์— ๊ด€๊ณ„์—†์ด ๊ณ ์ •๋œ ์ˆ˜์˜ ๊ฒ€์ƒ‰๋œ ์ง€๋ฌธ์„ ๋ฌด๋ถ„๋ณ„ํ•˜๊ฒŒ ๊ฒ€์ƒ‰ํ•˜์—ฌ ํ†ตํ•ฉํ•˜๋ฉด LM์˜ ๋‹ค์šฉ๋„์„ฑ์ด ๋–จ์–ด์ง€๊ฑฐ๋‚˜ ์œ ์šฉํ•˜์ง€ ์•Š์€ ์‘๋‹ต ์ƒ์„ฑ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๋ฐ ์ž๊ธฐ ์„ฑ์ฐฐ์„ ํ†ตํ•ด LM์˜ ํ’ˆ์งˆ๊ณผ ์‚ฌ์‹ค์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” **Self-Reflective Retrieval-Augmented Generation (Self-RAG)**์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ž„์˜์˜ ๋‹จ์ผ LM์„ ํ›ˆ๋ จํ•˜์—ฌ ํ•„์š”์— ๋”ฐ๋ผ ์ ์‘์ ์œผ๋กœ ์ง€๋ฌธ์„ ๊ฒ€์ƒ‰ํ•˜๊ณ , ๊ฒ€์ƒ‰๋œ ์ง€๋ฌธ๊ณผ ์ž์ฒด ์ƒ์„ฑ์„ _reflection_ ํ† ํฐ์ด๋ผ๋Š” ํŠน์ˆ˜ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ•˜๊ณ  ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. reflection ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋ฉด ์ถ”๋ก  ๋‹จ๊ณ„์—์„œ LM์„ ์ œ์–ดํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ๋™์ž‘์„ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, Self-RAG(ํŒŒ๋ผ๋ฏธํ„ฐ 7B ๋ฐ 13B)๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—… ์„ธํŠธ์—์„œ ์ตœ์‹  LLM๊ณผ Retrieval-Augmented ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋†€๋ผ์šด ๊ธฐ๋Šฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์ง€์‹์—๋งŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์‹ค์— ๋ถ€์ •ํ™•ํ•œ ๋‚ด์šฉ์ด ํฌํ•จ๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ด€๋ จ ์ง€์‹ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด LLM์„ ๋ณด๊ฐ•ํ•˜๋Š” ์ž„์‹œ๋ฐฉํŽธ ์ ‘๊ทผ ๋ฐฉ์‹์ธ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์€ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰์ด ํ•„์š”ํ•œ์ง€ ์—ฌ๋ถ€๋‚˜ ์ง€๋ฌธ์ด ๊ด€๋ จ์„ฑ์ด ์žˆ๋Š”์ง€์— ๊ด€๊ณ„์—†์ด ๊ณ ์ •๋œ ์ˆ˜์˜ ๊ฒ€์ƒ‰๋œ ์ง€๋ฌธ์„ ๋ฌด๋ถ„๋ณ„ํ•˜๊ฒŒ ๊ฒ€์ƒ‰ํ•˜์—ฌ ํ†ตํ•ฉํ•˜๋ฉด LM์˜ ๋‹ค์šฉ๋„์„ฑ์ด ๋–จ์–ด์ง€๊ฑฐ๋‚˜ ์œ ์šฉํ•˜์ง€ ์•Š์€ ์‘๋‹ต ์ƒ์„ฑ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๋ฐ ์ž๊ธฐ ์„ฑ์ฐฐ์„ ํ†ตํ•ด LM์˜ ํ’ˆ์งˆ๊ณผ ์‚ฌ์‹ค์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” **์ž๊ธฐ ๋ฐ˜์˜ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(Self-RAG)**์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ž„์˜์˜ ๋‹จ์ผ LM์„ ํ›ˆ๋ จํ•˜์—ฌ ํ•„์š”์— ๋”ฐ๋ผ ์ ์‘์ ์œผ๋กœ ์ง€๋ฌธ์„ ๊ฒ€์ƒ‰ํ•˜๊ณ , ๊ฒ€์ƒ‰๋œ ์ง€๋ฌธ๊ณผ ์ž์ฒด ์ƒ์„ฑ์„ _๋ฐ˜์‚ฌ_ ํ† ํฐ์ด๋ผ๋Š” ํŠน์ˆ˜ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ•˜๊ณ  ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜์‚ฌ ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋ฉด ์ถ”๋ก  ๋‹จ๊ณ„์—์„œ LM์„ ์ œ์–ดํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ๋™์ž‘์„ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, Self-RAG(ํŒŒ๋ผ๋ฏธํ„ฐ 7B ๋ฐ 13B)๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—… ์„ธํŠธ์—์„œ ์ตœ์‹  LLM๊ณผ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
However, these methods and datasets have not yet undergone systematic review and analysis. To fill this gap, we propose this work to conduct a comprehensive and detailed analysis of the XoT family. Even though there have been some surveys discussing chain-of-thought, they are limited to specific aspects, such as LLM reasoning with prompts Qiao et al. (2023) and chain-of-thought prompt strategies Yu et al. (2023). In contrast, our survey not only provides a more thorough and comprehensive discussion of the topics they've already covered, but also includes additional topics and discussions, such as XoT construction, XoT structural variants and frontier application, etc. Concretely, in this paper, we first introduce the relevant background and preliminary (SS2). Furthermore,we carefully classify the XoT series of work from multiple perspectives and complete an in-depth analysis (SS4), including XoT construction methods (SS4.1), XoT structure variants (SS4.2) and XoT enhancement methods (SS4.3). Then, we provide practical applications of the XoT in the frontier fields (SS5). In order to inspire the follow-up work of XoT, we offer insights into potential avenues for future research in this area (SS6). Finally, we compare and discuss existing methods (SS7).
๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์•„์ง ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ์™€ ๋ถ„์„์„ ๊ฑฐ์น˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ์ด ์ž‘์—…์„ ์ œ์•ˆํ•˜์—ฌ XoT ์ œํ’ˆ๊ตฐ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์ƒ์„ธํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์‚ฌ๊ณ ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•œ ๋ช‡ ๊ฐ€์ง€ ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ์žˆ์—ˆ์ง€๋งŒ, prompts๋ฅผ ์‚ฌ์šฉํ•œ LLM reasoning Qiao et al. (2023) ๋ฐ chain-of-thought prompt strategies Yu et al. (2023)๊ณผ ๊ฐ™์€ ํŠน์ • ์ธก๋ฉด์— ๊ตญํ•œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์ด๋ฏธ ๋‹ค๋ฃจ์–ด์ง„ ์ฃผ์ œ์— ๋Œ€ํ•ด ๋ณด๋‹ค ์ฒ ์ €ํ•˜๊ณ  ํฌ๊ด„์ ์ธ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ XoT construction, XoT structural variants ๋ฐ frontier application ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๊ฐ€ ์ฃผ์ œ์™€ ๋…ผ์˜๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋จผ์ € ๊ด€๋ จ ๋ฐฐ๊ฒฝ๊ณผ ์˜ˆ๋น„(SS2)๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, XoT construction methods(SS4.1), XoT structure variants(SS4.2), XoT enhancement methods(SS4.3) ๋“ฑ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ XoT ์‹œ๋ฆฌ์ฆˆ ์ž‘์—…์„ ๋ฉด๋ฐ€ํžˆ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์‹ฌ์ธต ๋ถ„์„์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค(SS4). ๊ทธ๋Ÿฐ ๋‹ค์Œ frontier fields์—์„œ XoT์˜ ์‹ค์ œ ์ ์šฉ ์‚ฌ๋ก€๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS5). XoT์˜ ํ›„์† ์ž‘์—…์— ์˜๊ฐ์„ ์ฃผ๊ธฐ ์œ„ํ•ด ์ด ๋ถ„์•ผ์˜ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž ์žฌ์  ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS6). ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ธฐ์กด ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•˜๊ณ  ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค(SS7).
๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์•„์ง ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ์™€ ๋ถ„์„์„ ๊ฑฐ์น˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ์ด ์ž‘์—…์„ ์ œ์•ˆํ•˜์—ฌ XoT ์ œํ’ˆ๊ตฐ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์ƒ์„ธํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์‚ฌ๊ณ ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•œ ๋ช‡ ๊ฐ€์ง€ ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ์žˆ์—ˆ์ง€๋งŒ, ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•œ LLM ์ถ”๋ก  Qiao et al. (2023) ๋ฐ ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ ์ „๋žต Yu et al. (2023)๊ณผ ๊ฐ™์€ ํŠน์ • ์ธก๋ฉด์— ๊ตญํ•œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์ด๋ฏธ ๋‹ค๋ฃจ์–ด์ง„ ์ฃผ์ œ์— ๋Œ€ํ•ด ๋ณด๋‹ค ์ฒ ์ €ํ•˜๊ณ  ํฌ๊ด„์ ์ธ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ XoT ๊ตฌ์„ฑ, XoT ๊ตฌ์กฐ ๋ณ€ํ˜• ๋ฐ ํ”„๋ก ํ‹ฐ์–ด ์ ์šฉ ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๊ฐ€ ์ฃผ์ œ์™€ ๋…ผ์˜๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋จผ์ € ๊ด€๋ จ ๋ฐฐ๊ฒฝ๊ณผ ์˜ˆ๋น„(SS2)๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, XoT ๊ตฌ์„ฑ ๋ฐฉ๋ฒ•(SS4.1), XoT ๊ตฌ์กฐ ๋ณ€ํ˜•(SS4.2), XoT ํ–ฅ์ƒ ๋ฐฉ๋ฒ•(SS4.3) ๋“ฑ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ XoT ์‹œ๋ฆฌ์ฆˆ ์ž‘์—…์„ ๋ฉด๋ฐ€ํžˆ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์‹ฌ์ธต ๋ถ„์„์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค(SS4). ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ”„๋ก ํ‹ฐ์–ด ๋ถ„์•ผ์—์„œ XoT์˜ ์‹ค์ œ ์ ์šฉ ์‚ฌ๋ก€๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS5). XoT์˜ ํ›„์† ์ž‘์—…์— ์˜๊ฐ์„ ์ฃผ๊ธฐ ์œ„ํ•ด ์ด ๋ถ„์•ผ์˜ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž ์žฌ์  ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS6). ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ธฐ์กด ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•˜๊ณ  ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค(SS7).
\begin{table} \begin{tabular}{l|c|c|c|c|c|c} \hline \hline Dataset & \(L_{\text{test-regen}}\) & \(L_{\text{test-ref}}\) & \(L_{\text{train}}\) & \(L_{\text{train-regen}}\) & \(\Delta_{1}\) & \(\Delta_{2}\) \\ \hline GSM8K & 0.52 & 0.50 & 0.11 & 0.33 & 0.02 & 0.19 \\ MATH & 0.59 & 0.58 & 0.23 & 0.39 & 0.01 & 0.20 \\ \hline \hline \end{tabular} \end{table} Table 6: To validate benchmark data leakage during data generation, we compare LM loss on: 1) a training subset, which is a small subset with 256 samples from all synthetic data; 2) a regenerated training subset, where we maintain the original questions from the training subset and use GPT-4 Turbo to rewrite answers; 3) a regenerated test set, where we keep the questions from the evaluation metrics unchanged and use GPT-4 Turbo to rewrite answers; 4) a reference test set, where we use the test set as seed to generate new questions and answers via GPT-4 Turbo. Referring to Skywork, we also report two key metrics: \(\Delta_{1}=L_{\text{test-regen}}-L_{\text{test-ref}},\Delta_{2}=L_{\text{test- regen}}-L_{\text{train-regen}}\), As \(\Delta_{1}\) is close to 0 and \(\Delta_{2}\) is significantly greater than 0 in two benchmarks, we believe that there is no leakage during the process of data synthesis.
\begin{table} \begin{tabular}{l|c|c|c|c|c|c} \hline \hline Dataset & \(L_{\text{test-regen}}\) & \(L_{\text{test-ref}}\) & \(L_{\text{train}}\) & \(L_{\text{train-regen}}\) & \(\Delta_{1}\) & \(\Delta_{2}\) \\ \hline GSM8K & 0.52 & 0.50 & 0.11 & 0.33 & 0.02 & 0.19 \\ MATH & 0.59 & 0.58 & 0.23 & 0.39 & 0.01 & 0.20 \\ \hline \hline \end{tabular} \end{table} ํ‘œ 6: ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ์ค‘ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์œ ์ถœ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ์— ๋Œ€ํ•œ LM loss๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค: 1) ๋ชจ๋“  ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—์„œ 256๊ฐœ์˜ ์ƒ˜ํ”Œ๋กœ ๊ตฌ์„ฑ๋œ ์ž‘์€ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ ํ›ˆ๋ จ ํ•˜์œ„ ์ง‘ํ•ฉ; 2) ์žฌ์ƒ์„ฑ๋œ ํ›ˆ๋ จ ํ•˜์œ„ ์ง‘ํ•ฉ, ์—ฌ๊ธฐ์„œ ํ›ˆ๋ จ ํ•˜์œ„ ์ง‘ํ•ฉ์˜ ์›๋ž˜ ์งˆ๋ฌธ์„ ์œ ์ง€ํ•˜๊ณ  GPT-4 Turbo๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์„ ๋‹ค์‹œ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค; 3) ์žฌ์ƒ์„ฑ๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ, ์—ฌ๊ธฐ์„œ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์˜ ์งˆ๋ฌธ์€ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  GPT-4 Turbo๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์„ ๋‹ค์‹œ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค; 4) ์ฐธ์กฐ ํ…Œ์ŠคํŠธ ์„ธํŠธ, ์—ฌ๊ธฐ์„œ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์‹œ๋“œ๋กœ ์‚ฌ์šฉํ•˜์—ฌ GPT-4 Turbo๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Skywork๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๋ฉ”ํŠธ๋ฆญ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค: \(\Delta_{1}=L_{\text{test-regen}}-L_{\text{test-ref}},\Delta_{2}=L_{\text{test- regen}}-L_{\text{train-regen}}\), ๋‘ ๋ฒค์น˜๋งˆํฌ์—์„œ \(\Delta_{1}\)๊ฐ€ 0์— ๊ฐ€๊น๊ณ  \(\Delta_{2}\)๊ฐ€ 0๋ณด๋‹ค ํ˜„์ €ํžˆ ํฌ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ ๊ณผ์ •์—์„œ ์œ ์ถœ์ด ์—†๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
\begin{table} \begin{tabular}{l|c|c|c|c|c|c} \hline \hline ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ & \(L_{\text{test-regen}}\) & \(L_{\text{test-ref}}\) & \(L_{\text{train}}\) & \(L_{\text{train-regen}}\) & \(\Delta_{1}\) & \(\Delta_{2}\) \\ \hline GSM8K & 0.52 & 0.50 & 0.11 & 0.33 & 0.02 & 0.19 \\ MATH & 0.59 & 0.58 & 0.23 & 0.39 & 0.01 & 0.20 \\ \hline \hline \end{tabular} \end{table} ํ‘œ 6: ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ์ค‘ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์œ ์ถœ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ์— ๋Œ€ํ•œ LM ์†์‹ค์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค: 1) ๋ชจ๋“  ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—์„œ 256๊ฐœ์˜ ์ƒ˜ํ”Œ๋กœ ๊ตฌ์„ฑ๋œ ์ž‘์€ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ ํ›ˆ๋ จ ํ•˜์œ„ ์ง‘ํ•ฉ; 2) ์žฌ์ƒ์„ฑ๋œ ํ›ˆ๋ จ ํ•˜์œ„ ์ง‘ํ•ฉ, ์—ฌ๊ธฐ์„œ ํ›ˆ๋ จ ํ•˜์œ„ ์ง‘ํ•ฉ์˜ ์›๋ž˜ ์งˆ๋ฌธ์„ ์œ ์ง€ํ•˜๊ณ  GPT-4 Turbo๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์„ ๋‹ค์‹œ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค; 3) ์žฌ์ƒ์„ฑ๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ, ์—ฌ๊ธฐ์„œ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์˜ ์งˆ๋ฌธ์€ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  GPT-4 Turbo๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์„ ๋‹ค์‹œ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค; 4) ์ฐธ์กฐ ํ…Œ์ŠคํŠธ ์„ธํŠธ, ์—ฌ๊ธฐ์„œ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์‹œ๋“œ๋กœ ์‚ฌ์šฉํ•˜์—ฌ GPT-4 Turbo๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Skywork๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๋ฉ”ํŠธ๋ฆญ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค: \(\Delta_{1}=L_{\text{test-regen}}-L_{\text{test-ref}},\Delta_{2}=L_{\text{test- regen}}-L_{\text{train-regen}}\), ๋‘ ๋ฒค์น˜๋งˆํฌ์—์„œ \(\Delta_{1}\)๊ฐ€ 0์— ๊ฐ€๊น๊ณ  \(\Delta_{2}\)๊ฐ€ 0๋ณด๋‹ค ํ˜„์ €ํžˆ ํฌ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ ๊ณผ์ •์—์„œ ์œ ์ถœ์ด ์—†๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
**Models.** We train 1B, 3B, and 7B parameters autoregressive decoder-only models, based on configurations and hyperparameters similar to GPT-3 (Brown et al., 2020), diverging mostly on our use of ALiBi (Press et al., 2021). We use FlashAttention (Dao et al., 2022) in a custom codebase. We train internal models on both The Pile and RefinedWeb to control for deviations caused by our pretraining setup-we found The Pile models to perform in-line with others. For small-scale and ablation studies (first half of Section 4.2; Section 4.3), we train models to optimality according to the scaling laws of Hoffmann et al. (2022): on 27B and 60B tokens respectively for our 1B and 3B parameters models. For the main experiments demonstrating our approach (Falcon-RW models in Section 4.2), we train the models to 350GT, in line with popular public models (Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022). Note that we do not compare against the recently introduced LLaMA models (Touvron et al., 2023), as the smallest of them is trained on x2.5 more compute than our largest model, preventing a meaningful comparison from being made dataset-wise. For a more in-depth overview of the models and pretraining datasets with which we compare, see Appendix F.
**๋ชจ๋ธ.** GPT-3(Brown et al., 2020)๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B ๋ฐ 7B ํŒŒ๋ผ๋ฏธํ„ฐ autoregressive decoder-only ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ์ฃผ๋กœ ALiBi(Press et al., 2021)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์—์„œ ์ฐจ์ด๊ฐ€ ๋‚ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ์ง€์ • ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. pretraining setup์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด The Pile๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, The Pile ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(4.2 ์ ˆ์˜ ์ „๋ฐ˜๋ถ€, 4.3 ์ ˆ)์˜ ๊ฒฝ์šฐ, 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด Hoffmann ๋“ฑ(2022)์˜ scaling laws์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(4.2์ ˆ์˜ Falcon-RW ๋ชจ๋ธ)์˜ ๊ฒฝ์šฐ, ์ธ๊ธฐ ์žˆ๋Š” ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022)์— ๋งž์ถฐ 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์€ ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค 2.5๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ๋ฐฉํ•ดํ•˜๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๋ฐ ๋น„๊ต ๋Œ€์ƒ pretraining ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
**๋ชจ๋ธ.** GPT-3(Brown et al., 2020)๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B ๋ฐ 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ์ฃผ๋กœ ALiBi(Press et al., 2021)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์—์„œ ์ฐจ์ด๊ฐ€ ๋‚ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ์ง€์ • ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋ ˆ์ผ๋ง ์Šคํƒ์—…์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๋” ํŒŒ์ผ๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ๋” ํŒŒ์ผ ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(4.2 ์ ˆ์˜ ์ „๋ฐ˜๋ถ€, 4.3 ์ ˆ)์˜ ๊ฒฝ์šฐ, 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด Hoffmann ๋“ฑ(2022)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(4.2์ ˆ์˜ Falcon-RW ๋ชจ๋ธ)์˜ ๊ฒฝ์šฐ, ์ธ๊ธฐ ์žˆ๋Š” ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022)์— ๋งž์ถฐ 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์€ ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค 2.5๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ๋ฐฉํ•ดํ•˜๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๋ฐ ๋น„๊ต ๋Œ€์ƒ ์ „์ฒ˜๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
The GPT family is one of the most common examples of AutoRegressive language models. It was first proposed by Radford et al. (2018), who identified a basic architecture of unsupervised pre-training followed by fine-tuning. Radford et al. (2019) later proposed zero-shot learning based on GPT. Later, Brown et al. (2020) proposed GPT-3 using an approach similar to Radford et al. (2019), which involved scaling up and abandoning fine-tuning. They also utilized alternately dense and locally banded sparse attentional patterns in the transformer layer, similar to the sparse transformer Child et al. (2019). There are also several related studies, such as GPT-NEO(Black et al., 2022) and ChatGPT, which use Reinforcement Learning from Human Feedback (RLHF). RLHF has significantly enhanced the accuracy of GPT models. Although ChatGPT is not open source, many researchers Jiang et al. (2023) (Khattab et al., 2022) still use its API for generative tasks in RALM. Recently, report on GPT-4(Achiam et al., 2023) have appeared, with the main focus on building a predictable and scalable deep learning stack dedicated to improving GPT-4's safety and alignment. Many researchers Asai et al. (2023) (Luo et al., 2023) have recently used GPT-4 to generate prompts for RALM.
GPT ์ œํ’ˆ๊ตฐ์€ AutoRegressive ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์˜ˆ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. Radford ๋“ฑ(2018)์ด ์ฒ˜์Œ ์ œ์•ˆํ–ˆ์œผ๋ฉฐ, ์ด๋“ค์€ ๋ฏธ์ง€๋„ ์‚ฌ์ „ ํ•™์Šต์— ์ด์–ด Fine-tuning์„ ๊ฑฐ์น˜๋Š” ๊ธฐ๋ณธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. Radford ๋“ฑ(2019)์€ ์ดํ›„ GPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ zero-shot ํ•™์Šต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ Brown ๋“ฑ(2020)์€ Radford ๋“ฑ(2019)๊ณผ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ GPT-3๋ฅผ ์ œ์•ˆํ–ˆ๋Š”๋ฐ, ์ด๋Š” ๊ทœ๋ชจ๋ฅผ ํ™•์žฅํ•˜๊ณ  Fine-tuning์„ ํฌ๊ธฐํ•˜๋Š” ๋ฐฉ์‹์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ sparse Transformer Child ๋“ฑ(2019)๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ Transformer ๋ ˆ์ด์–ด์— ๊ตํ˜ธ์ ์œผ๋กœ ์กฐ๋ฐ€ํ•˜๊ณ  ๊ตญ๋ถ€์ ์œผ๋กœ ๋ฐด๋“œ๊ฐ€ ์žˆ๋Š” sparse attention ํŒจํ„ด์„ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-NEO(Black et al., 2022) ๋ฐ ChatGPT์™€ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ์—ฐ๊ตฌ๋„ ์žˆ์œผ๋ฉฐ, ์ด ์—ฐ๊ตฌ๋“ค์€ Reinforcement Learning from Human Feedback (RLHF)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. RLHF๋Š” GPT ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ChatGPT๋Š” ์˜คํ”ˆ ์†Œ์Šค๊ฐ€ ์•„๋‹ˆ์ง€๋งŒ, ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด Jiang et al. (2023) (Khattab et al., 2022)์€ ์—ฌ์ „ํžˆ RALM์˜ ์ƒ์„ฑ ์ž‘์—…์— ChatGPT์˜ API๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” GPT-4(Achiam et al., 2023)์— ๋Œ€ํ•œ ๋ณด๊ณ ์„œ๊ฐ€ ๋“ฑ์žฅํ–ˆ์œผ๋ฉฐ, ์ฃผ์š” ์ดˆ์ ์€ GPT-4์˜ ์•ˆ์ „์„ฑ๊ณผ ์ •๋ ฌ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์˜ˆ์ธก ๊ฐ€๋Šฅํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋”ฅ ๋Ÿฌ๋‹ ์Šคํƒ ๊ตฌ์ถ•์— ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด Asai ๋“ฑ(2023) (Luo ๋“ฑ, 2023)์€ ์ตœ๊ทผ RALM์— ๋Œ€ํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
GPT ์ œํ’ˆ๊ตฐ์€ ์ž๋™ ํšŒ๊ท€ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์˜ˆ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. Radford ๋“ฑ(2018)์ด ์ฒ˜์Œ ์ œ์•ˆํ–ˆ์œผ๋ฉฐ, ์ด๋“ค์€ ๋ฏธ์ง€๋„ ์‚ฌ์ „ ํ•™์Šต์— ์ด์–ด ๋ฏธ์„ธ ์กฐ์ •์„ ๊ฑฐ์น˜๋Š” ๊ธฐ๋ณธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. Radford ๋“ฑ(2019)์€ ์ดํ›„ GPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ œ๋กœ ์ƒท ํ•™์Šต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ Brown ๋“ฑ(2020)์€ Radford ๋“ฑ(2019)๊ณผ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ GPT-3๋ฅผ ์ œ์•ˆํ–ˆ๋Š”๋ฐ, ์ด๋Š” ๊ทœ๋ชจ๋ฅผ ํ™•์žฅํ•˜๊ณ  ๋ฏธ์„ธ ์กฐ์ •์„ ํฌ๊ธฐํ•˜๋Š” ๋ฐฉ์‹์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ŠคํŒŒ์Šค ํŠธ๋žœ์Šคํฌ๋จธ Child ๋“ฑ(2019)๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์— ๊ตํ˜ธ์ ์œผ๋กœ ์กฐ๋ฐ€ํ•˜๊ณ  ๊ตญ๋ถ€์ ์œผ๋กœ ๋ฐด๋“œ๊ฐ€ ์žˆ๋Š” ์ŠคํŒŒ์Šค ์ฃผ์˜ ํŒจํ„ด์„ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-NEO(Black et al., 2022) ๋ฐ ChatGPT์™€ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ์—ฐ๊ตฌ๋„ ์žˆ์œผ๋ฉฐ, ์ด ์—ฐ๊ตฌ๋“ค์€ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ(RLHF)์„ ํ†ตํ•œ ๊ฐ•ํ™” ํ•™์Šต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. RLHF๋Š” GPT ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ChatGPT๋Š” ์˜คํ”ˆ ์†Œ์Šค๊ฐ€ ์•„๋‹ˆ์ง€๋งŒ, ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด Jiang et al. (2023) (Khattab et al., 2022)์€ ์—ฌ์ „ํžˆ RALM์˜ ์ƒ์„ฑ ์ž‘์—…์— ChatGPT์˜ API๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” GPT-4(Achiam et al., 2023)์— ๋Œ€ํ•œ ๋ณด๊ณ ์„œ๊ฐ€ ๋“ฑ์žฅํ–ˆ์œผ๋ฉฐ, ์ฃผ์š” ์ดˆ์ ์€ GPT-4์˜ ์•ˆ์ „์„ฑ๊ณผ ์ •๋ ฌ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์˜ˆ์ธก ๊ฐ€๋Šฅํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋”ฅ ๋Ÿฌ๋‹ ์Šคํƒ ๊ตฌ์ถ•์— ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด Asai ๋“ฑ(2023) (Luo ๋“ฑ, 2023)์€ ์ตœ๊ทผ RALM์— ๋Œ€ํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
The KL-constrained reward maximization objective used in typical RLHF algorithms balances exploitation of reward while restricting the policy from deviating far from the reference policy. Therefore, when comparing algorithms, we must take into account both reward achieved as well as the KL discrepancy; achieving slightly higher reward but with much higher KL is not necessarily desirable. Figure 2 shows the reward-KL frontier for various algorithms in the sentiment setting. We execute multiple training runs for each algorithm, using a different hyperparameter for policy conservativeness in each run (target \(\text{KL}\in\{3,6,9,12\}\) for PPO, \(\beta\in\{0.05,0.1,1,5\}\), \(\alpha\in\{0.05,0.1,0.5,1\}\) for unlikelihood, random seeds for preferred-FT). This sweep includes 22 runs in total. After each 100 training steps until convergence, we evaluate each policy on a set of test prompts, computing the average reward under the true reward function as well as the average sequence-level KL3 with the reference policy \(\text{KL}\left(\pi\mid\mid\pi_{\text{ref}}\right)\). We find that DPO produces by far the most efficient frontier, achieving the highest reward while still achieving low KL. This result is particularly notable for multiple reasons.
์ผ๋ฐ˜์ ์ธ RLHF ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์‚ฌ์šฉ๋˜๋Š” KL-constrained reward maximization objective๋Š” ๋ณด์ƒ์˜ ์ฐฉ์ทจ๋ฅผ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ ์ •์ฑ…์ด ๊ธฐ์ค€ ์ •์ฑ…์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๋„๋ก ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋น„๊ตํ•  ๋•Œ๋Š” ๋‹ฌ์„ฑํ•œ ๋ณด์ƒ๊ณผ KL discrepancy๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•ด์•ผ ํ•˜๋ฉฐ, ์•ฝ๊ฐ„ ๋” ๋†’์€ ๋ณด์ƒ์„ ๋‹ฌ์„ฑํ•˜๋˜ KL์ด ํ›จ์”ฌ ๋” ๋†’์€ ๊ฒƒ์€ ๋ฐ˜๋“œ์‹œ ๋ฐ”๋žŒ์งํ•œ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ๊ฐ์„ฑ ์„ค์ •์—์„œ ๋‹ค์–‘ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ reward-KL frontier๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์ค‘ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค(PPO์˜ ๊ฒฝ์šฐ \(\text{KL}\in\{3,6,9,12\}\), \(\beta\in\{0.05,0.1,1,5\}\), \(\alpha\in\{0.05,0.1,0.5,1\}\), ์„ ํ˜ธ๋„-FT์˜ ๊ฒฝ์šฐ ๋žœ๋ค ์‹œ๋“œ). ์ด ์Šค์œ•์—๋Š” ์ด 22๊ฐœ์˜ ์‹คํ–‰์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ˆ˜๋ ด๊นŒ์ง€ ๊ฐ 100๊ฐœ์˜ ํ›ˆ๋ จ ๋‹จ๊ณ„๋งˆ๋‹ค ๊ฐ ์ •์ฑ…์„ ํ…Œ์ŠคํŠธ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•˜๊ณ , ์‹ค์ œ ๋ณด์ƒ ํ•จ์ˆ˜์—์„œ ํ‰๊ท  ๋ณด์ƒ๊ณผ ์ฐธ์กฐ ์ •์ฑ… \(\text{KL}\left(\pi\mid\mid\pi_{\text{ref}}\right)\)์˜ ํ‰๊ท  ์‹œํ€€์Šค ์ˆ˜์ค€ KL3์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. DPO๋Š” ๋‚ฎ์€ KL์„ ๋‹ฌ์„ฑํ•˜๋ฉด์„œ ๊ฐ€์žฅ ๋†’์€ ๋ณด์ƒ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฐ€์žฅ ํšจ์œจ์ ์ธ ํ”„๋ก ํ‹ฐ์–ด๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ด์œ ๋กœ ํŠนํžˆ ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค.
์ผ๋ฐ˜์ ์ธ RLHF ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์‚ฌ์šฉ๋˜๋Š” KL ์ œํ•œ ๋ณด์ƒ ์ตœ๋Œ€ํ™” ๋ชฉ์ ์€ ๋ณด์ƒ์˜ ์ฐฉ์ทจ๋ฅผ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ ์ •์ฑ…์ด ๊ธฐ์ค€ ์ •์ฑ…์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๋„๋ก ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋น„๊ตํ•  ๋•Œ๋Š” ๋‹ฌ์„ฑํ•œ ๋ณด์ƒ๊ณผ KL ๋ถˆ์ผ์น˜๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•ด์•ผ ํ•˜๋ฉฐ, ์•ฝ๊ฐ„ ๋” ๋†’์€ ๋ณด์ƒ์„ ๋‹ฌ์„ฑํ•˜๋˜ KL์ด ํ›จ์”ฌ ๋” ๋†’์€ ๊ฒƒ์€ ๋ฐ˜๋“œ์‹œ ๋ฐ”๋žŒ์งํ•œ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ๊ฐ์„ฑ ์„ค์ •์—์„œ ๋‹ค์–‘ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ ๋ณด์ƒ-KL ํ”„๋ก ํ‹ฐ์–ด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์ค‘ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค(PPO์˜ ๊ฒฝ์šฐ \text{KL}\in\{3,6,9,12\}, \beta\in\{0.05,0.1,1,5\}, \alpha\in\{0.05,0.1,0.5,1\}, ์„ ํ˜ธ๋„-FT์˜ ๊ฒฝ์šฐ ๋žœ๋ค ์‹œ๋“œ). ์ด ์Šค์œ•์—๋Š” ์ด 22๊ฐœ์˜ ์‹คํ–‰์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ˆ˜๋ ด๊นŒ์ง€ ๊ฐ 100๊ฐœ์˜ ํ›ˆ๋ จ ๋‹จ๊ณ„๋งˆ๋‹ค ๊ฐ ์ •์ฑ…์„ ํ…Œ์ŠคํŠธ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•˜๊ณ , ์‹ค์ œ ๋ณด์ƒ ํ•จ์ˆ˜์—์„œ ํ‰๊ท  ๋ณด์ƒ๊ณผ ์ฐธ์กฐ ์ •์ฑ… \text{KL}\left(\pi\mid\mid\pi_{\text{ref}}\right)\)์˜ ํ‰๊ท  ์‹œํ€€์Šค ์ˆ˜์ค€ KL3์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. DPO๋Š” ๋‚ฎ์€ KL์„ ๋‹ฌ์„ฑํ•˜๋ฉด์„œ ๊ฐ€์žฅ ๋†’์€ ๋ณด์ƒ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฐ€์žฅ ํšจ์œจ์ ์ธ ํ”„๋ก ํ‹ฐ์–ด๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ด์œ ๋กœ ํŠนํžˆ ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค.
Over recent years, an increasing amount of compute and data has been poured into training large language models (LLMs), usually by doing one-pass learning on as many tokens as possible randomly selected from large-scale web corpora. While training on ever-larger portions of the internet leads to consistent performance improvements, the size of these improvements diminishes with scale, and there has been little work exploring the effect of data selection on pre-training and downstream performance beyond simple de-duplication methods such as MinHash. Here, we show that careful data selection (on top of de-duplicated data) via pre-trained model embeddings can speed up training (20% efficiency gains) and improves average downstream accuracy on 16 NLP tasks (up to 2%) at the 6.7B model scale. Furthermore, we show that repeating data intelligently consistently _outperforms_ baseline training (while repeating random data performs worse than baseline training). Our results indicate that clever data selection can significantly improve LLM pre-training, calls into question the common practice of training for a single epoch on as much data as possible, and demonstrates a path to keep improving our models past the limits of randomly sampling web data.
์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐ๋Š” ์›น ์ฝ”ํผ์Šค์—์„œ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ๋งŽ์€ ํ† ํฐ์— ๋Œ€ํ•ด ํ•œ ๋ฒˆ ํ•™์Šตํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๊ณ„์† ๋งŽ์€ compute์™€ data๊ฐ€ ํˆฌ์ž…๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ธํ„ฐ๋„ท์˜ ์ ์  ๋” ๋งŽ์€ ๋ถ€๋ถ„์„ ํ•™์Šตํ•˜๋ฉด ์ง€์†์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ–ฅ์ƒ์˜ ๊ทœ๋ชจ๋Š” ์ ์  ์ค„์–ด๋“ค๊ณ  ์žˆ์œผ๋ฉฐ, MinHash์™€ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์„ ๋„˜์–ด์„  data selection์ด pre-training ๋ฐ downstream ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” pre-trained model embeddings์„ ํ†ตํ•ด ์‹ ์ค‘ํ•œ data selection(์ค‘๋ณต ์ œ๊ฑฐ๋œ data ์œ„์—)์„ ํ†ตํ•ด ํ•™์Šต ์†๋„๋ฅผ ๋†’์ด๊ณ (ํšจ์œจ์„ฑ 20% ํ–ฅ์ƒ), 16๊ฐœ์˜ NLP ์ž‘์—…์—์„œ downstream ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(์ตœ๋Œ€ 2%, ๋ชจ๋ธ ๊ทœ๋ชจ 6.7B). ๋˜ํ•œ, data๋ฅผ ์ง€๋Šฅ์ ์œผ๋กœ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด baseline ํ›ˆ๋ จ์„ ์ง€์†์ ์œผ๋กœ _๋Šฅ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(๋ฌด์ž‘์œ„ data๋ฅผ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์€ baseline ํ›ˆ๋ จ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง). ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์˜๋ฆฌํ•œ data selection์€ LLM pre-training์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ€๋Šฅํ•œ ํ•œ ๋งŽ์€ data์— ๋Œ€ํ•ด ํ•œ ๋ฒˆ๋งŒ ํ•™์Šตํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰์— ์˜๋ฌธ์„ ์ œ๊ธฐํ•˜๊ณ , ์›น data๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ํ•œ๊ณ„๋ฅผ ๋„˜์–ด ๋ชจ๋ธ์„ ๊ณ„์† ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ๋กœ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐ๋Š” ์›น ์ฝ”ํผ์Šค์—์„œ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ๋งŽ์€ ํ† ํฐ์— ๋Œ€ํ•ด ํ•œ ๋ฒˆ ํ•™์Šตํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๊ณ„์† ๋งŽ์€ ์ปดํ“จํŒ…๊ณผ ๋ฐ์ดํ„ฐ๊ฐ€ ํˆฌ์ž…๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ธํ„ฐ๋„ท์˜ ์ ์  ๋” ๋งŽ์€ ๋ถ€๋ถ„์„ ํ•™์Šตํ•˜๋ฉด ์ง€์†์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ–ฅ์ƒ์˜ ๊ทœ๋ชจ๋Š” ์ ์  ์ค„์–ด๋“ค๊ณ  ์žˆ์œผ๋ฉฐ, MinHash์™€ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์„ ๋„˜์–ด์„  ๋ฐ์ดํ„ฐ ์„ ํƒ์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ ์ž„๋ฒ ๋”ฉ์„ ํ†ตํ•ด ์‹ ์ค‘ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ(์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ ์œ„์—)์„ ํ†ตํ•ด ํ•™์Šต ์†๋„๋ฅผ ๋†’์ด๊ณ (ํšจ์œจ์„ฑ 20% ํ–ฅ์ƒ), 16๊ฐœ์˜ NLP ์ž‘์—…์—์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(์ตœ๋Œ€ 2%, ๋ชจ๋ธ ๊ทœ๋ชจ 6.7B). ๋˜ํ•œ, ๋ฐ์ดํ„ฐ๋ฅผ ์ง€๋Šฅ์ ์œผ๋กœ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด ๊ธฐ๋ณธ์„  ํ›ˆ๋ จ์„ ์ง€์†์ ์œผ๋กœ _๋Šฅ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(๋ฌด์ž‘์œ„ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์€ ๊ธฐ๋ณธ์„  ํ›ˆ๋ จ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง). ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์˜๋ฆฌํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์€ LLM ์‚ฌ์ „ ํ•™์Šต์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ€๋Šฅํ•œ ํ•œ ๋งŽ์€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•œ ๋ฒˆ๋งŒ ํ•™์Šตํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰์— ์˜๋ฌธ์„ ์ œ๊ธฐํ•˜๊ณ , ์›น ๋ฐ์ดํ„ฐ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ํ•œ๊ณ„๋ฅผ ๋„˜์–ด ๋ชจ๋ธ์„ ๊ณ„์† ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ๋กœ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
OLMo would not have been possible without the support of many individuals and institutions. The experimental components of this work were made possible through a partnership with AMD and CSC, enabling use of the LUMI supercomputer, and Kemper Institute at Harvard University. We thank Jonathan Frankle and the team at MosaicML (now Databricks) for sharing their experiences with FSDP, and building the code base that OLMo is based on. We thank our teammates Taira Anderson, Michelle Benedict, Jon Borchardt, Evie Cheng, Arnavi Chheda, Johann Dahm, Matt Latzke, Kelsey MacMillan, Aaron Sarnat, Carissa Schoenick, Sam Skjonsberg, Michael Schmitz, Michael Wilson, Caitlin Wittlif, and the entire IT team, for their help with the website, design, internal and external communications, budgeting, and other activities that supported smooth progress on this project. Finally, we also express gratitude for the helpful discussions and feedback from our teammates at AI2 and close collaborators, including Prithviraj (Raj) Ammanabrolu, Peter Clark, Nicole DeCario, Doug Downey, Ali Farhadi, Ian Ferreira, Vaino Hatanpaa, Sham M. Kakade, Julien Launay, Sydney Levine, Pekka Manninen, Franzi Roessner, Maarten Sap, Ludwig Schmidt, Yulia Tsvetkov, and Daniel S. Weld.
OLMo๋Š” ๋งŽ์€ ๊ฐœ์ธ๊ณผ ๊ธฐ๊ด€์˜ ์ง€์› ์—†์ด๋Š” ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์‹คํ—˜์  ๊ตฌ์„ฑ ์š”์†Œ๋Š” AMD ๋ฐ CSC์™€์˜ ํŒŒํŠธ๋„ˆ์‹ญ์„ ํ†ตํ•ด LUMI ์Šˆํผ์ปดํ“จํ„ฐ์™€ Harvard University์˜ Kemper Institute๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค. Jonathan Frankle๊ณผ MosaicML(ํ˜„ Databricks) ํŒ€์—๊ฒŒ FSDP์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๊ณต์œ ํ•˜๊ณ  OLMo์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ์ฝ”๋“œ๋ฒ ์ด์Šค๋ฅผ ๊ตฌ์ถ•ํ•ด์ค€ ๊ฒƒ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์›น์‚ฌ์ดํŠธ, ๋””์ž์ธ, ๋Œ€๋‚ด์™ธ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜, ์˜ˆ์‚ฐ ์ฑ…์ • ๋ฐ ๊ธฐํƒ€ ํ™œ๋™์„ ์ง€์›ํ•˜์—ฌ ํ”„๋กœ์ ํŠธ์˜ ์›ํ™œํ•œ ์ง„ํ–‰์„ ๋„์™€์ค€ ๋™๋ฃŒ์ธ Taira Anderson, Michelle Benedict, Jon Borchardt, Evie Cheng, Arnavi Chheda, Johann Dahm, Matt Latzke, Kelsey MacMillan, Aaron Sarnat, Carissa Schoenick, Sam Skjonsberg, Michael Schmitz, Michael Wilson, Caitlin Wittlif, ๊ทธ๋ฆฌ๊ณ  ์ „์ฒด IT ํŒ€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ AI2์˜ ํŒ€์›๋“ค๊ณผ Prithviraj (Raj) Ammanabrolu, Peter Clark, Nicole DeCario, Doug Downey, Ali Farhadi, Ian Ferreira, Vaino Hatanpaa, Sham M. Kakade, Julien Launay, Sydney Levine, Pekka Manninen, Franzi Roessner, Maarten Sap, Ludwig Schmidt, Yulia Tsvetkov, ๊ทธ๋ฆฌ๊ณ  Daniel S. Weld๋ฅผ ํฌํ•จํ•œ ๊ธด๋ฐ€ํ•œ ํ˜‘๋ ฅ์ž๋“ค์˜ ๋„์›€์ด ๋œ ํ† ๋ก ๊ณผ ํ”ผ๋“œ๋ฐฑ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค.
OLMo๋Š” ๋งŽ์€ ๊ฐœ์ธ๊ณผ ๊ธฐ๊ด€์˜ ์ง€์› ์—†์ด๋Š” ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์‹คํ—˜์  ๊ตฌ์„ฑ ์š”์†Œ๋Š” AMD ๋ฐ CSC์™€์˜ ํŒŒํŠธ๋„ˆ์‹ญ์„ ํ†ตํ•ด ๋ฃธ๋ฏธ ์Šˆํผ์ปดํ“จํ„ฐ์™€ ํ•˜๋ฒ„๋“œ ๋Œ€ํ•™๊ต์˜ ์ผํผ ์—ฐ๊ตฌ์†Œ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์กฐ๋‚˜๋‹จ ํ”„๋žญํด๊ณผ ๋ชจ์ž์ดํฌML(ํ˜„ ๋‹คํƒ€๋ธŒ๋ฆญ์Šค) ํŒ€์—๊ฒŒ FSDP์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๊ณต์œ ํ•˜๊ณ  OLMo์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ์ฝ”๋“œ๋ฒ ์ด์Šค๋ฅผ ๊ตฌ์ถ•ํ•ด์ค€ ๊ฒƒ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์›น์‚ฌ์ดํŠธ, ๋””์ž์ธ, ๋Œ€๋‚ด์™ธ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜, ์˜ˆ์‚ฐ ์ฑ…์ • ๋ฐ ๊ธฐํƒ€ ํ™œ๋™์„ ์ง€์›ํ•˜์—ฌ ํ”„๋กœ์ ํŠธ์˜ ์›ํ™œํ•œ ์ง„ํ–‰์„ ๋„์™€์ค€ ๋™๋ฃŒ์ธ ํƒ€์ด๋ผ ์•ค๋”์Šจ, ๋ฏธ์…ธ ๋ฒ ๋„ค๋”•ํŠธ, ์กด ๋ณด์ฐจ๋“œ, ์—๋น„ ์ฒธ, ์•„๋ฅด๋‚˜๋น„ ์ฒด๋‹ค, ์š”ํ•œ ๋‹ด, ๋งคํŠธ ๋ผ์ธ ์ผ€, ์ผˆ์‹œ ๋งฅ๋ฐ€๋ž€, ์•„๋ก  ์‚ฌ๋ฅด๋‚˜ํŠธ, ์นด๋ฆฌ์‚ฌ ์‡ผ๋‹‰, ์ƒ˜ ์Šคํ‚ค์˜จ์Šค๋ฒ„๊ทธ, ๋งˆ์ดํด ์Šˆ๋ฏธ์ธ , ๋งˆ์ดํด ์œŒ์Šจ, ์ผ€์ผ๋ฆฐ ์œ—๋ฆฌํ”„, ๊ทธ๋ฆฌ๊ณ  ์ „์ฒด IT ํŒ€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ AI2์˜ ํŒ€์›๋“ค๊ณผ ํ”„๋ฆฌํ‹ฐ๋น„๋ผ์ฆˆ(Raj) ์•”๋งˆ๋‚˜๋ธŒ๋กœ๋ฃจ, ํ”ผํ„ฐ ํด๋ผํฌ, ๋‹ˆ์ฝœ ๋ฐ์นด๋ฆฌ์˜ค, ๋”๊ทธ ๋‹ค์šฐ๋‹ˆ, ์•Œ๋ฆฌ ํŒŒ๋ฅดํ•˜๋””, ์ด์•ˆ ํŽ˜๋ ˆ์ด๋ผ, ๋ฐ”์ด๋…ธ ํ•˜ํƒ„ํŒŒ, ์ƒด M. ์นด์นด๋ฐ, ์ค„๋ฆฌ์•™ ๋กœ๋„ค, ์‹œ๋“œ๋‹ˆ ๋ ˆ๋นˆ, ํŽ˜์นด ๋งˆ๋‹ˆ๋„จ, ํ”„๋ž€์ง€ ๋กœ์—๋ฅด๋„ˆ, ๋งˆ๋ฅดํ… ์žฝ๋ฒ„๊ทธ, ๋ฃจ๋“œ๋น„ํžˆ ์Šˆ๋ฏธํŠธ, ์œจ๋ฆฌ์•„ ์ธ ๋ฒ ํŠธ์ฝ”ํ”„, ๋‹ค๋‹ˆ์—˜ S. ์›ฐ๋“œ๋ฅผ ํฌํ•จํ•œ ๊ธด๋ฐ€ํ•œ ํ˜‘๋ ฅ์ž๋“ค์˜ ๋„์›€์ด ๋œ ํ† ๋ก ๊ณผ ํ”ผ๋“œ๋ฐฑ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค.
Recently, the development of open-source large language models (LLMs) has advanced rapidly. Nevertheless, due to data constraints, the capabilities of most open-source LLMs are primarily focused on English. To address this issue, we introduce the concept of _chat vector_ to equip pre-trained language models with instruction following and human value alignment via simple model arithmetic. The chat vector is derived by subtracting the weights of a pre-trained base model (e.g. LLaMA2) from those of its corresponding chat model (e.g. LLaMA2-chat). By simply adding the chat vector to a continual pre-trained model's weights, we can endow the model with chat capabilities in new languages without the need for further training. Our empirical studies demonstrate the superior efficacy of the chat vector from three different aspects: instruction following, toxicity mitigation, and multi-turn dialogue. Moreover, to showcase the adaptability of our approach, we extend our experiments to encompass various languages, base models, and chat vectors. The results underscore the chat vector's simplicity, effectiveness, and wide applicability, making it a compelling solution for efficiently enabling conversational capabilities in pre-trained language models.
์ตœ๊ทผ ์˜คํ”ˆ ์†Œ์Šค Large Language Models (LLMs)์˜ ๊ฐœ๋ฐœ์ด ๋น ๋ฅด๊ฒŒ ์ง„์ „๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ฐ์ดํ„ฐ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๋Œ€๋ถ€๋ถ„์˜ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ๊ธฐ๋Šฅ์€ ์ฃผ๋กœ ์˜์–ด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด _chat vector_์˜ ๊ฐœ๋…์„ ๋„์ž…ํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์— ๊ฐ„๋‹จํ•œ ๋ชจ๋ธ ์—ฐ์‚ฐ์„ ํ†ตํ•ด Instruction following ๋ฐ ์ธ๊ฐ„์˜ ๊ฐ€์น˜ ์ •๋ ฌ ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. chat vector๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๊ธฐ๋ณธ ๋ชจ๋ธ(์˜ˆ: LLaMA2)์˜ ๊ฐ€์ค‘์น˜์—์„œ ํ•ด๋‹น chat ๋ชจ๋ธ(์˜ˆ: LLaMA2-chat)์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋บ€ ๊ฐ’์œผ๋กœ๋ถ€ํ„ฐ ๋„์ถœ๋ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜์— chat vector๋ฅผ ๋”ํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ์ถ”๊ฐ€ ํ•™์Šต ์—†์ด๋„ ์ƒˆ๋กœ์šด ์–ธ์–ด๋กœ ๋ชจ๋ธ์— chat ๊ธฐ๋Šฅ์„ ๋ถ€์—ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด Instruction following, ๋…์„ฑ ์™„ํ™”, ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™”์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ chat vector์˜ ์šฐ์ˆ˜ํ•œ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ ์‘๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์–ธ์–ด, ๊ธฐ๋ณธ ๋ชจ๋ธ ๋ฐ chat vectors๋ฅผ ํฌ๊ด„ํ•˜๋Š” ์‹คํ—˜์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ chat vector์˜ ๋‹จ์ˆœ์„ฑ, ํšจ๊ณผ์„ฑ ๋ฐ ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋Œ€ํ™” ๊ธฐ๋Šฅ์„ ํšจ์œจ์ ์œผ๋กœ ํ™œ์„ฑํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ์†”๋ฃจ์…˜์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ตœ๊ทผ ์˜คํ”ˆ ์†Œ์Šค ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ๊ฐœ๋ฐœ์ด ๋น ๋ฅด๊ฒŒ ์ง„์ „๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ฐ์ดํ„ฐ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๋Œ€๋ถ€๋ถ„์˜ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ๊ธฐ๋Šฅ์€ ์ฃผ๋กœ ์˜์–ด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด _์ฑ„ํŒ… ๋ฒกํ„ฐ_์˜ ๊ฐœ๋…์„ ๋„์ž…ํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์— ๊ฐ„๋‹จํ•œ ๋ชจ๋ธ ์—ฐ์‚ฐ์„ ํ†ตํ•ด ์ง€์‹œ์–ด ์ถ”์ข… ๋ฐ ์ธ๊ฐ„์˜ ๊ฐ€์น˜ ์ •๋ ฌ ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ฑ„ํŒ… ๋ฒกํ„ฐ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๊ธฐ๋ณธ ๋ชจ๋ธ(์˜ˆ: LLaMA2)์˜ ๊ฐ€์ค‘์น˜์—์„œ ํ•ด๋‹น ์ฑ„ํŒ… ๋ชจ๋ธ(์˜ˆ: LLaMA2-chat)์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋บ€ ๊ฐ’์œผ๋กœ๋ถ€ํ„ฐ ๋„์ถœ๋ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜์— ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ๋”ํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ์ถ”๊ฐ€ ํ•™์Šต ์—†์ด๋„ ์ƒˆ๋กœ์šด ์–ธ์–ด๋กœ ๋ชจ๋ธ์— ์ฑ„ํŒ… ๊ธฐ๋Šฅ์„ ๋ถ€์—ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด ์ง€์‹œ์–ด ์ถ”์ข…, ๋…์„ฑ ์™„ํ™”, ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™”์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ฑ„ํŒ… ๋ฒกํ„ฐ์˜ ์šฐ์ˆ˜ํ•œ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ ์‘๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์–ธ์–ด, ๊ธฐ๋ณธ ๋ชจ๋ธ ๋ฐ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํฌ๊ด„ํ•˜๋Š” ์‹คํ—˜์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ์ฑ„ํŒ… ๋ฒกํ„ฐ์˜ ๋‹จ์ˆœ์„ฑ, ํšจ๊ณผ์„ฑ ๋ฐ ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋Œ€ํ™” ๊ธฐ๋Šฅ์„ ํšจ์œจ์ ์œผ๋กœ ํ™œ์„ฑํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ์†”๋ฃจ์…˜์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**Data Generation for Instruction Tuning.** To address the high cost of human annotation for high-quality instruction-response pairs, several studies advocate for automating the data generation process Schick and Schutze (2021); Liu et al. (2022); Meng et al. (2023). Leveraging the in-context learning Brown et al. (2020) ability of LLMs, Wang et al. (2020), and et al. (2022); Honovich et al. (2022) prompt LLMs with seed instructions to generate synthetic ones. These are then fed to stronger LLMs, _e.g._, Chat-GPT, to generate responses for training the target (often smaller) LLM (Taori et al., 2023). As a representative work, WizardLM (Xu et al., 2023), designs a fixed set of human-crafted operations to increase complexity of instructions and control difficulty of generated data. Zhao et al. (2023); Zhou et al. (2023) further confirm the importance of instruction complexity for LLM alignment through empirical studies. Different from these works that rely on pre-defined rules without considering the downstream tasks, CodecLM enables automatically tailoring instructions for different downstream tasks and target LLMs. We also introduce Self-Rubrics and Contrastive Filtering to further identify the most effective instruction-response pairs.
**Instruction Tuning์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ**. ๊ณ ํ’ˆ์งˆ Instruction-์‘๋‹ต ์Œ์— ๋Œ€ํ•œ ๋†’์€ ์ธ์  ์ฃผ์„ ๋น„์šฉ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡๋ช‡ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•  ๊ฒƒ์„ ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค(Schick and Schutze, 2021); Liu et al. (2022); Meng et al. (2023). LLM์˜ In-Context Learning Brown et al. (2020) ๋Šฅ๋ ฅ์„ ํ™œ์šฉํ•˜์—ฌ Wang et al. (2020), et al. (2022); Honovich et al. (2022)์€ ์‹œ๋“œ Instruction์œผ๋กœ LLM์„ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ํ•ฉ์„ฑ Instruction์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ Instruction์€ ๊ฐ•๋ ฅํ•œ LLM(์˜ˆ: Chat-GPT)์— ์ž…๋ ฅ๋˜์–ด ๋Œ€์ƒ(์†Œํ˜•) LLM์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(Taori et al., 2023). ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ๋กœ, WizardLM(Xu et al., 2023)์€ Instruction์˜ ๋ณต์žก์„ฑ์„ ๋†’์ด๊ณ  ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๋‚œ์ด๋„๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ์ด ์ง์ ‘ ๋งŒ๋“  ์ž‘์—… ์ง‘ํ•ฉ์„ ๊ณ ์ •์ ์œผ๋กœ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. Zhao et al. (2023); Zhou et al. (2023)์€ ๊ฒฝํ—˜์  ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด LLM ์ •๋ ฌ์— ์žˆ์–ด Instruction ๋ณต์žก์„ฑ์˜ ์ค‘์š”์„ฑ์„ ์ถ”๊ฐ€๋กœ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ์‚ฌ์ „ ์ •์˜๋œ ๊ทœ์น™์— ์˜์กดํ•˜๋Š” ์ด๋Ÿฌํ•œ ์ž‘์—…๊ณผ๋Š” ๋‹ฌ๋ฆฌ, CodecLM์€ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…๊ณผ ๋Œ€์ƒ LLM์— ๋งž๊ฒŒ Instruction์„ ์ž๋™์œผ๋กœ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ€์žฅ ํšจ๊ณผ์ ์ธ Instruction-์‘๋‹ต ์Œ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด Self-Rubrics์™€ Contrastive Filtering์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
**๋ช…๋ น ํŠœ๋‹์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ**. ๊ณ ํ’ˆ์งˆ ๋ช…๋ น์–ด-์‘๋‹ต ์Œ์— ๋Œ€ํ•œ ๋†’์€ ์ธ์  ์ฃผ์„ ๋น„์šฉ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡๋ช‡ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•  ๊ฒƒ์„ ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค(Schick and Schutze, 2021); Liu et al. (2022); Meng et al. (2023). LLM์˜ ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต Brown et al. (2020) ๋Šฅ๋ ฅ์„ ํ™œ์šฉํ•˜์—ฌ Wang et al. (2020), et al. (2022); Honovich et al. (2022)์€ ์‹œ๋“œ ๋ช…๋ น์–ด๋กœ LLM์„ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ํ•ฉ์„ฑ ๋ช…๋ น์–ด๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ช…๋ น์–ด๋Š” ๊ฐ•๋ ฅํ•œ LLM(์˜ˆ: Chat-GPT)์— ์ž…๋ ฅ๋˜์–ด ๋Œ€์ƒ(์†Œํ˜•) LLM์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(Taori et al., 2023). ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ๋กœ, WizardLM(Xu et al., 2023)์€ ๋ช…๋ น์–ด์˜ ๋ณต์žก์„ฑ์„ ๋†’์ด๊ณ  ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๋‚œ์ด๋„๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ์ด ์ง์ ‘ ๋งŒ๋“  ์ž‘์—… ์ง‘ํ•ฉ์„ ๊ณ ์ •์ ์œผ๋กœ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. Zhao et al. (2023); Zhou et al. (2023)์€ ๊ฒฝํ—˜์  ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด LLM ์ •๋ ฌ์— ์žˆ์–ด ๋ช…๋ น์–ด ๋ณต์žก์„ฑ์˜ ์ค‘์š”์„ฑ์„ ์ถ”๊ฐ€๋กœ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ์‚ฌ์ „ ์ •์˜๋œ ๊ทœ์น™์— ์˜์กดํ•˜๋Š” ์ด๋Ÿฌํ•œ ์ž‘์—…๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์ฝ”๋ฑLM์€ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…๊ณผ ๋Œ€์ƒ LLM์— ๋งž๊ฒŒ ๋ช…๋ น์–ด๋ฅผ ์ž๋™์œผ๋กœ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ€์žฅ ํšจ๊ณผ์ ์ธ ๋ช…๋ น์–ด-์‘๋‹ต์–ด ์Œ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ์…€ํ”„ ๋ฃจ๋ธŒ๋ฆญ์Šค์™€ ๋Œ€์กฐ ํ•„ํ„ฐ๋ง์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
Outside of the context of language, learning policies from preferences has been studied in both bandit and reinforcement learning settings, and several approaches have been proposed. Contextual bandit learning using preferences or rankings of actions, rather than rewards, is known as a contextual dueling bandit (CDB; [48; 14]). In the absence of absolute rewards, theoretical analysis of CDBs substitutes the notion of an optimal policy with a _von Neumann winner_, a policy whose expected win rate against _any_ other policy is at least 50% [14]. However, in the CDB setting, preference labels are given online, while in learning from human preferences, we typically learn from a fixed batch of offline preference-annotated action pairs [47]. Similarly, _preference-based RL_ (PbRL) learns from binary preferences generated by an _unknown_'scoring' function rather than rewards [9; 35]. Various algorithms for PbRL exist, including methods that can reuse off-policy preference data, but generally involve first explicitly estimating the latent scoring function (i.e. the reward model) and subsequently optimizing it [16; 9; 12; 34; 19]. We instead present a single stage policy learning approach that directly optimizes a policy to satisfy preferences.
์–ธ์–ด์˜ ๋งฅ๋ฝ์—์„œ ๋ฒ—์–ด๋‚˜ ์„ ํ˜ธ๋„์—์„œ ์ •์ฑ…์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ bandit ๋ฐ Reinforcement Learning ํ™˜๊ฒฝ์—์„œ ์—ฐ๊ตฌ๋˜์–ด ์™”์œผ๋ฉฐ, ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ณด์ƒ์ด ์•„๋‹Œ ํ–‰๋™์˜ ์„ ํ˜ธ๋„ ๋˜๋Š” ์ˆœ์œ„๋ฅผ ์‚ฌ์šฉํ•˜๋Š” contextual bandit ํ•™์Šต์„ contextual dueling bandit (CDB; [48; 14])๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ ˆ๋Œ€์ ์ธ ๋ณด์ƒ์ด ์—†๋Š” ๊ฒฝ์šฐ, CDB์— ๋Œ€ํ•œ ์ด๋ก ์  ๋ถ„์„์€ ์ตœ์ ์˜ ์ •์ฑ… ๊ฐœ๋…์„ _von Neumann winner_๋กœ ๋Œ€์ฒดํ•˜๋ฉฐ, ์ด๋Š” _๋‹ค๋ฅธ ์ •์ฑ…์— ๋Œ€ํ•œ ๊ธฐ๋Œ€ ์Šน๋ฅ ์ด ์ตœ์†Œ 50%์ธ ์ •์ฑ…์ž…๋‹ˆ๋‹ค[14]. ๊ทธ๋Ÿฌ๋‚˜ CDB ์„ค์ •์—์„œ๋Š” ์„ ํ˜ธ๋„ ๋ ˆ์ด๋ธ”์ด ์˜จ๋ผ์ธ์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๋ฐ˜๋ฉด, ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜คํ”„๋ผ์ธ ์„ ํ˜ธ๋„๊ฐ€ ์ฃผ์–ด์ง„ ๊ณ ์ •๋œ ์ผ๊ด„ ์ž‘์—…์—์„œ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[47]. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, _preference-based RL_ (PbRL)์€ ๋ณด์ƒ์ด ์•„๋‹Œ _์•Œ ์ˆ˜ ์—†๋Š”_'์ ์ˆ˜' ํ•จ์ˆ˜์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ด์ง„ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[9; 35]. off-policy ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ๋‹ค์–‘ํ•œ PbRL ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์กด์žฌํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ž ์žฌ ์ ์ˆ˜ ํ•จ์ˆ˜(์ฆ‰, ๋ณด์ƒ ๋ชจ๋ธ)๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์ถ”์ •ํ•œ ๋‹ค์Œ ์ด๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค[16; 9; 12; 34; 19]. ๋Œ€์‹  ์„ ํ˜ธ๋„๋ฅผ ๋งŒ์กฑํ•˜๋„๋ก ์ •์ฑ…์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋‹จ์ผ ๋‹จ๊ณ„ ์ •์ฑ… ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
์–ธ์–ด์˜ ๋งฅ๋ฝ์—์„œ ๋ฒ—์–ด๋‚˜ ์„ ํ˜ธ๋„์—์„œ ์ •์ฑ…์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋ฐด๋””ํŠธ ๋ฐ ๊ฐ•ํ™” ํ•™์Šต ํ™˜๊ฒฝ์—์„œ ์—ฐ๊ตฌ๋˜์–ด ์™”์œผ๋ฉฐ, ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ณด์ƒ์ด ์•„๋‹Œ ํ–‰๋™์˜ ์„ ํ˜ธ๋„ ๋˜๋Š” ์ˆœ์œ„๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ปจํ…์ŠคํŠธ ๋ฐด๋””ํŠธ ํ•™์Šต์„ ์ปจํ…์ŠคํŠธ ๋“€์–ผ๋ง ๋ฐด๋””ํŠธ(CDB; [48; 14])๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ ˆ๋Œ€์ ์ธ ๋ณด์ƒ์ด ์—†๋Š” ๊ฒฝ์šฐ, CDB์— ๋Œ€ํ•œ ์ด๋ก ์  ๋ถ„์„์€ ์ตœ์ ์˜ ์ •์ฑ… ๊ฐœ๋…์„ _ํฐ ๋…ธ์ด๋งŒ ์Šน์ž_๋กœ ๋Œ€์ฒดํ•˜๋ฉฐ, ์ด๋Š” _๋‹ค๋ฅธ ์ •์ฑ…์— ๋Œ€ํ•œ ๊ธฐ๋Œ€ ์Šน๋ฅ ์ด ์ตœ์†Œ 50%์ธ ์ •์ฑ…์ž…๋‹ˆ๋‹ค[14]. ๊ทธ๋Ÿฌ๋‚˜ CDB ์„ค์ •์—์„œ๋Š” ์„ ํ˜ธ๋„ ๋ ˆ์ด๋ธ”์ด ์˜จ๋ผ์ธ์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๋ฐ˜๋ฉด, ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜คํ”„๋ผ์ธ ์„ ํ˜ธ๋„๊ฐ€ ์ฃผ์–ด์ง„ ๊ณ ์ •๋œ ์ผ๊ด„ ์ž‘์—…์—์„œ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[47]. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, _์„ ํ˜ธ๋„ ๊ธฐ๋ฐ˜ RL_(PbRL)์€ ๋ณด์ƒ์ด ์•„๋‹Œ _์•Œ ์ˆ˜ ์—†๋Š”_'์ ์ˆ˜' ํ•จ์ˆ˜์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ด์ง„ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[9; 35]. ์˜คํ”„ํด๋ฆฌ์‹œ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ๋‹ค์–‘ํ•œ PbRL ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์กด์žฌํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ž ์žฌ ์ ์ˆ˜ ํ•จ์ˆ˜(์ฆ‰, ๋ณด์ƒ ๋ชจ๋ธ)๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์ถ”์ •ํ•œ ๋‹ค์Œ ์ด๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค[16; 9; 12; 34; 19]. ๋Œ€์‹  ์„ ํ˜ธ๋„๋ฅผ ๋งŒ์กฑํ•˜๋„๋ก ์ •์ฑ…์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋‹จ์ผ ๋‹จ๊ณ„ ์ •์ฑ… ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
Multimodal-CoT Zhang et al. (2023) firstly fine-tunes multi-modal models to generate chain-of-thoughts and then reasons over the rationales to obtain final answers. However, it suffers from the limitation of the linearity of the reasoning process and has difficulties in interacting between different modalities. To alleviate the challenges encountered by Multimodal-CoT, Yao et al. (2023) proposes Graph-of-Thought (GoT), which models the thought processes as a graph. It parses the reasoning chains into a thought graph, which enables a more realistic representation of thought processes by capturing non-sequential information interactions. This measure breaks the limitations of linear structure through graphical structures and further improves performance. Furthermore, Yao et al. (2023) proposes Hypergraph-of-Thought (HoT), replacing thought graphs with hypergraphs, which enables models with better ability of high-order multi-hop reasoning and multi-modal comparative judgment. Meanwhile, some work takes an approach based on knowledge distillation. T-SciQ Wang et al. (2023) generates high-quality CoT rationales from LLMs as fine-tuning signals and introduces a novel data mixing strategy to produce effective samples for different questions.
๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ-CoT Zhang et al. (2023)์€ ๋จผ์ € ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์„ Fine-tuningํ•˜์—ฌ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ๊ทธ ๊ทผ๊ฑฐ๋ฅผ ํ†ตํ•ด ์ถ”๋ก ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ถ”๋ก  ๊ณผ์ •์˜ ์„ ํ˜•์„ฑ ํ•œ๊ณ„๋ฅผ ๊ฒช๊ณ  ์žˆ์œผ๋ฉฐ, ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐฉ์‹ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ-CoT๊ฐ€ ๊ฒช๋Š” ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด Yao et al. (2023)์€ ์‚ฌ๊ณ  ๊ณผ์ •์„ ๊ทธ๋ž˜ํ”„๋กœ ๋ชจ๋ธ๋งํ•˜๋Š” GoT(Graph-of-Thought)๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก  ์‚ฌ์Šฌ์„ ์‚ฌ๊ณ  ๊ทธ๋ž˜ํ”„๋กœ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜์—ฌ ๋น„์ˆœ์ฐจ์  ์ •๋ณด ์ƒํ˜ธ ์ž‘์šฉ์„ ํฌ์ฐฉํ•˜์—ฌ ์‚ฌ๊ณ  ๊ณผ์ •์„ ๋ณด๋‹ค ํ˜„์‹ค์ ์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ธก์ •์€ ๊ทธ๋ž˜ํ”ฝ ๊ตฌ์กฐ๋ฅผ ํ†ตํ•ด ์„ ํ˜• ๊ตฌ์กฐ์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ณ  ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ, Yao et al. (2023)์€ ์‚ฌ๊ณ  ๊ทธ๋ž˜ํ”„๋ฅผ ํ•˜์ดํผ๊ทธ๋ž˜ํ”„๋กœ ๋Œ€์ฒดํ•˜๋Š” Hypergraph-of-Thought (HoT)๋ฅผ ์ œ์•ˆํ•˜์—ฌ ๊ณ ์ฐจ ๋ฉ€ํ‹ฐํ™‰ ์ถ”๋ก ๊ณผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋น„๊ต ํŒ๋‹จ ๋Šฅ๋ ฅ์ด ๋”์šฑ ํ–ฅ์ƒ๋œ ๋ชจ๋ธ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง€์‹ ์ฆ๋ฅ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. T-SciQ Wang et al. (2023)์€ LLM์—์„œ ๊ณ ํ’ˆ์งˆ CoT ๊ทผ๊ฑฐ๋ฅผ Fine-tuning ์‹ ํ˜ธ๋กœ ์ƒ์„ฑํ•˜๊ณ  ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ํšจ๊ณผ์ ์ธ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ-CoT Zhang et al. (2023)์€ ๋จผ์ € ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ๊ทธ ๊ทผ๊ฑฐ๋ฅผ ํ†ตํ•ด ์ถ”๋ก ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ถ”๋ก  ๊ณผ์ •์˜ ์„ ํ˜•์„ฑ ํ•œ๊ณ„๋ฅผ ๊ฒช๊ณ  ์žˆ์œผ๋ฉฐ, ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐฉ์‹ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ-CoT๊ฐ€ ๊ฒช๋Š” ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด Yao et al. (2023)์€ ์‚ฌ๊ณ  ๊ณผ์ •์„ ๊ทธ๋ž˜ํ”„๋กœ ๋ชจ๋ธ๋งํ•˜๋Š” GoT(Graph-of-Thought)๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก  ์‚ฌ์Šฌ์„ ์‚ฌ๊ณ  ๊ทธ๋ž˜ํ”„๋กœ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜์—ฌ ๋น„์ˆœ์ฐจ์  ์ •๋ณด ์ƒํ˜ธ ์ž‘์šฉ์„ ํฌ์ฐฉํ•˜์—ฌ ์‚ฌ๊ณ  ๊ณผ์ •์„ ๋ณด๋‹ค ํ˜„์‹ค์ ์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ธก์ •์€ ๊ทธ๋ž˜ํ”ฝ ๊ตฌ์กฐ๋ฅผ ํ†ตํ•ด ์„ ํ˜• ๊ตฌ์กฐ์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ณ  ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ, Yao et al. (2023)์€ ์‚ฌ๊ณ  ๊ทธ๋ž˜ํ”„๋ฅผ ํ•˜์ดํผ๊ทธ๋ž˜ํ”„๋กœ ๋Œ€์ฒดํ•˜๋Š” ํ•˜์ดํผ๊ทธ๋ž˜ํ”„ ์˜ค๋ธŒ ์‚ฌ๊ณ (HoT)๋ฅผ ์ œ์•ˆํ•˜์—ฌ ๊ณ ์ฐจ ๋ฉ€ํ‹ฐํ™‰ ์ถ”๋ก ๊ณผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋น„๊ต ํŒ๋‹จ ๋Šฅ๋ ฅ์ด ๋”์šฑ ํ–ฅ์ƒ๋œ ๋ชจ๋ธ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง€์‹ ์ฆ๋ฅ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. T-SciQ Wang et al. (2023)์€ LLM์—์„œ ๊ณ ํ’ˆ์งˆ CoT ๊ทผ๊ฑฐ๋ฅผ ๋ฏธ์„ธ ์กฐ์ • ์‹ ํ˜ธ๋กœ ์ƒ์„ฑํ•˜๊ณ  ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ํšจ๊ณผ์ ์ธ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Task learning._ In the second way, LLMs learn new tasks unseen in the pre-training stage only through demonstrations. Specially, task learning is analyzed mainly from the perspective of gradient descent and considered as implicit fine-tuning [496, 65]. Then, ICL can be explained as follows: by means of forward computation, LLMs generate meta-gradients with respect to demonstrations and implicitly perform gradient descent via the attention mechanism. Experiments also show that certain attention heads in LLMs are capable of performing task-agnostic atomic operations (_e.g.,_ copying and prefix matching), which are closely related to the ICL ability [497]. Furthermore, some studies abstract ICL as an algorithm learning process [498]. For example, the authors in [498] find that LLMs essentially encode implicit models through their parameters during pre-training. With the examples provided in ICL, LLMs can implement learning algorithms such as gradient descent or directly compute the closed-form solution to update these models during forward computation. Under this explanation framework, it has been shown that LLMs can effectively learn simple linear functions and even some complex functions like decision trees with ICL [498].
\(\bullet\)_Task ํ•™์Šต._ ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๋ณด์—ฌ์ฃผ๊ธฐ๋งŒ ํ†ตํ•ด ๋ณด์ด์ง€ ์•Š๋Š” ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, task ํ•™์Šต์€ ์ฃผ๋กœ gradient descent์˜ ๊ด€์ ์—์„œ ๋ถ„์„๋˜๋ฉฐ ์•”์‹œ์  Fine-tuning์œผ๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค[496, 65]. ๊ทธ๋Ÿฐ ๋‹ค์Œ, In-Context Learning (ICL)์€ ํฌ์›Œ๋“œ ๊ณ„์‚ฐ์„ ํ†ตํ•ด LLM์ด ๋ฐ๋ชจ์— ๋Œ€ํ•œ ๋ฉ”ํƒ€ ๊ทธ๋ผ๋ฐ์ด์…˜์„ ์ƒ์„ฑํ•˜๊ณ  attention ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ด ์•”์‹œ์ ์œผ๋กœ gradient descent๋ฅผ ์ˆ˜ํ–‰ํ•œ๋‹ค๋Š” ์‹์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, LLM์˜ ํŠน์ • attention ํ—ค๋“œ๋Š” ๋ณต์‚ฌ ๋ฐ ์ ‘๋‘์‚ฌ ๋งค์นญ๊ณผ ๊ฐ™์€ ์ž‘์—…์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ์›์ž ์—ฐ์‚ฐ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ICL ๋Šฅ๋ ฅ๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค[497]. ๋˜ํ•œ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ICL์„ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ํ•™์Šต ํ”„๋กœ์„ธ์Šค๋กœ ์ถ”์ƒํ™”ํ•ฉ๋‹ˆ๋‹ค[498]. ์˜ˆ๋ฅผ ๋“ค์–ด, [498]์˜ ์ €์ž๋“ค์€ LLM์ด ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ†ตํ•ด ์•”์‹œ์  ๋ชจ๋ธ์„ ์ธ์ฝ”๋”ฉํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์— ์ œ๊ณต๋œ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด LLM์€ gradient descent์™€ ๊ฐ™์€ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๊ฑฐ๋‚˜ ํฌ์›Œ๋“œ ๊ณ„์‚ฐ ์ค‘์— ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์—…๋ฐ์ดํŠธํ•˜๊ธฐ ์œ„ํ•ด ํ์‡„ํ˜• ์†”๋ฃจ์…˜์„ ์ง์ ‘ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค๋ช… ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ LLM์€ ICL์„ ํ†ตํ•ด ๋‹จ์ˆœํ•œ ์„ ํ˜• ํ•จ์ˆ˜๋Š” ๋ฌผ๋ก  ์˜์‚ฌ ๊ฒฐ์ • ํŠธ๋ฆฌ์™€ ๊ฐ™์€ ์ผ๋ถ€ ๋ณต์žกํ•œ ํ•จ์ˆ˜๋„ ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[498].
\(\bullet\)_Task ํ•™์Šต._ ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๋ณด์—ฌ์ฃผ๊ธฐ๋งŒ ํ†ตํ•ด ๋ณด์ด์ง€ ์•Š๋Š” ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ์ž‘์—… ํ•™์Šต์€ ์ฃผ๋กœ ๊ฒฝ์‚ฌ ํ•˜๊ฐ•์˜ ๊ด€์ ์—์„œ ๋ถ„์„๋˜๋ฉฐ ์•”์‹œ์  ๋ฏธ์„ธ ์กฐ์ •์œผ๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค[496, 65]. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ICL์€ ํฌ์›Œ๋“œ ๊ณ„์‚ฐ์„ ํ†ตํ•ด LLM์ด ๋ฐ๋ชจ์— ๋Œ€ํ•œ ๋ฉ”ํƒ€ ๊ทธ๋ผ๋ฐ์ด์…˜์„ ์ƒ์„ฑํ•˜๊ณ  ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ด ์•”์‹œ์ ์œผ๋กœ ๊ทธ๋ผ๋ฐ์ด์…˜ ํ•˜๊ฐ•์„ ์ˆ˜ํ–‰ํ•œ๋‹ค๋Š” ์‹์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, LLM์˜ ํŠน์ • ์ฃผ์˜ ํ—ค๋“œ๋Š” ๋ณต์‚ฌ ๋ฐ ์ ‘๋‘์‚ฌ ๋งค์นญ๊ณผ ๊ฐ™์€ ์ž‘์—…์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ์›์ž ์—ฐ์‚ฐ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ICL ๋Šฅ๋ ฅ๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค[497]. ๋˜ํ•œ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ICL์„ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ํ•™์Šต ํ”„๋กœ์„ธ์Šค๋กœ ์ถ”์ƒํ™”ํ•ฉ๋‹ˆ๋‹ค[498]. ์˜ˆ๋ฅผ ๋“ค์–ด, [498]์˜ ์ €์ž๋“ค์€ LLM์ด ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ†ตํ•ด ์•”์‹œ์  ๋ชจ๋ธ์„ ์ธ์ฝ”๋”ฉํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์— ์ œ๊ณต๋œ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด LLM์€ ๊ฒฝ์‚ฌ ํ•˜๊ฐ•๊ณผ ๊ฐ™์€ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๊ฑฐ๋‚˜ ํฌ์›Œ๋“œ ๊ณ„์‚ฐ ์ค‘์— ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์—…๋ฐ์ดํŠธํ•˜๊ธฐ ์œ„ํ•ด ํ์‡„ํ˜• ์†”๋ฃจ์…˜์„ ์ง์ ‘ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค๋ช… ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ LLM์€ ICL์„ ํ†ตํ•ด ๋‹จ์ˆœํ•œ ์„ ํ˜• ํ•จ์ˆ˜๋Š” ๋ฌผ๋ก  ์˜์‚ฌ ๊ฒฐ์ • ํŠธ๋ฆฌ์™€ ๊ฐ™์€ ์ผ๋ถ€ ๋ณต์žกํ•œ ํ•จ์ˆ˜๋„ ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[498].
1. **Analyzing patient information:** Given the current conversation history, AMIE was instructed to 1) summarize the positive and negative symptoms of the patient as well as any relevant medical/family/social history and demographic information, 2) produce a current differential diagnosis, 3) note missing information needed for a more accurate diagnosis and 4) assess confidence in the current differential and highlight its urgency. 2. **Formulating response and action:** Building upon the conversation history and the output of step 1, AMIE performed the following: 1) Generate a response to the patient's last message and formulate further questions to acquire missing information and refine the differential diagnosis. 2) If necessary, recommend immediate action, such as an emergency room visit. If confident in the diagnosis based on available information, present the differential. 3. **Refining the response:** AMIE revises its previous output to meet specific criteria based on the conversation history and outputs from earlier steps. The criteria are primarily related to factuality and formatting of the response (e.g., avoid factual inaccuracies on patient facts and unnecessary repetition, show empathy, and display in a clear format).
1. **ํ™˜์ž ์ •๋ณด ๋ถ„์„:** ํ˜„์žฌ ๋Œ€ํ™” ๊ธฐ๋ก์„ ๊ณ ๋ คํ•  ๋•Œ, AMIE๋Š” 1) ํ™˜์ž์˜ ๊ธ์ •์  ๋ฐ ๋ถ€์ •์  ์ฆ์ƒ๊ณผ ๊ด€๋ จ ์˜๋ฃŒ/๊ฐ€์กฑ/์‚ฌํšŒ์  ๋ณ‘๋ ฅ ๋ฐ ์ธ๊ตฌํ†ต๊ณ„ํ•™์  ์ •๋ณด๋ฅผ ์š”์•ฝํ•˜๊ณ , 2) ํ˜„์žฌ differential diagnosis๋ฅผ ์ƒ์„ฑํ•˜๊ณ , 3) ๋ณด๋‹ค ์ •ํ™•ํ•œ ์ง„๋‹จ์— ํ•„์š”ํ•œ ๋ˆ„๋ฝ๋œ ์ •๋ณด๋ฅผ ๊ธฐ๋กํ•˜๊ณ , 4) ํ˜„์žฌ differential diagnosis์— ๋Œ€ํ•œ ์ž์‹ ๊ฐ์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ทธ ๊ธด๊ธ‰์„ฑ์„ ๊ฐ•์กฐํ•˜๋„๋ก ์ง€์‹œ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. 2. **์‘๋‹ต ๋ฐ ์กฐ์น˜ ๊ณต์‹ํ™”ํ•˜๊ธฐ:** ๋Œ€ํ™” ๊ธฐ๋ก๊ณผ 1๋‹จ๊ณ„์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ AMIE๋Š” ๋‹ค์Œ์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค: 1) ํ™˜์ž์˜ ๋งˆ์ง€๋ง‰ ๋ฉ”์‹œ์ง€์— ๋Œ€ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ณ  ๋ˆ„๋ฝ๋œ ์ •๋ณด๋ฅผ ํ™•๋ณดํ•˜๊ณ  differential diagnosis๋ฅผ ๊ตฌ์ฒดํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ์งˆ๋ฌธ์„ ๊ณต์‹ํ™”ํ•ฉ๋‹ˆ๋‹ค. 2) ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์‘๊ธ‰์‹ค ๋ฐฉ๋ฌธ๊ณผ ๊ฐ™์€ ์ฆ‰๊ฐ์ ์ธ ์กฐ์น˜๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ง„๋‹จ์— ์ž์‹ ์ด ์žˆ๋‹ค๋ฉด differential diagnosis๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. 3. **์‘๋‹ต์„ ๊ตฌ์ฒดํ™”ํ•ฉ๋‹ˆ๋‹ค:** AMIE๋Š” ๋Œ€ํ™” ๊ธฐ๋ก๊ณผ ์ด์ „ ๋‹จ๊ณ„์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ํŠน์ • ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋„๋ก ์ด์ „ ๊ฒฐ๊ณผ๋ฌผ์„ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์€ ์ฃผ๋กœ ์‘๋‹ต์˜ ์‚ฌ์‹ค์„ฑ ๋ฐ ํ˜•์‹๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ™˜์ž ์‚ฌ์‹ค์— ๋Œ€ํ•œ ์‚ฌ์‹ค ๋ถ€์ •ํ™• ๋ฐ ๋ถˆํ•„์š”ํ•œ ๋ฐ˜๋ณต ํ”ผํ•˜๊ธฐ, ๊ณต๊ฐ ํ‘œ์‹œ, ๋ช…ํ™•ํ•œ ํ˜•์‹์œผ๋กœ ํ‘œ์‹œ).
1. **ํ™˜์ž ์ •๋ณด ๋ถ„์„:** ํ˜„์žฌ ๋Œ€ํ™” ๊ธฐ๋ก์„ ๊ณ ๋ คํ•  ๋•Œ, AMIE๋Š” 1) ํ™˜์ž์˜ ๊ธ์ •์  ๋ฐ ๋ถ€์ •์  ์ฆ์ƒ๊ณผ ๊ด€๋ จ ์˜๋ฃŒ/๊ฐ€์กฑ/์‚ฌํšŒ์  ๋ณ‘๋ ฅ ๋ฐ ์ธ๊ตฌํ†ต๊ณ„ํ•™์  ์ •๋ณด๋ฅผ ์š”์•ฝํ•˜๊ณ , 2) ํ˜„์žฌ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ƒ์„ฑํ•˜๊ณ , 3) ๋ณด๋‹ค ์ •ํ™•ํ•œ ์ง„๋‹จ์— ํ•„์š”ํ•œ ๋ˆ„๋ฝ๋œ ์ •๋ณด๋ฅผ ๊ธฐ๋กํ•˜๊ณ , 4) ํ˜„์žฌ ๊ฐ๋ณ„ ์ง„๋‹จ์— ๋Œ€ํ•œ ์ž์‹ ๊ฐ์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ทธ ๊ธด๊ธ‰์„ฑ์„ ๊ฐ•์กฐํ•˜๋„๋ก ์ง€์‹œ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. 2. **์‘๋‹ต ๋ฐ ์กฐ์น˜ ๊ณต์‹ํ™”ํ•˜๊ธฐ:** ๋Œ€ํ™” ๊ธฐ๋ก๊ณผ 1๋‹จ๊ณ„์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ AMIE๋Š” ๋‹ค์Œ์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค: 1) ํ™˜์ž์˜ ๋งˆ์ง€๋ง‰ ๋ฉ”์‹œ์ง€์— ๋Œ€ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ณ  ๋ˆ„๋ฝ๋œ ์ •๋ณด๋ฅผ ํ™•๋ณดํ•˜๊ณ  ๊ฐ๋ณ„ ์ง„๋‹จ์„ ๊ตฌ์ฒดํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ์งˆ๋ฌธ์„ ๊ณต์‹ํ™”ํ•ฉ๋‹ˆ๋‹ค. 2) ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์‘๊ธ‰์‹ค ๋ฐฉ๋ฌธ๊ณผ ๊ฐ™์€ ์ฆ‰๊ฐ์ ์ธ ์กฐ์น˜๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ง„๋‹จ์— ์ž์‹ ์ด ์žˆ๋‹ค๋ฉด ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. 3. **์‘๋‹ต์„ ๊ตฌ์ฒดํ™”ํ•ฉ๋‹ˆ๋‹ค:** AMIE๋Š” ๋Œ€ํ™” ๊ธฐ๋ก๊ณผ ์ด์ „ ๋‹จ๊ณ„์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ํŠน์ • ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋„๋ก ์ด์ „ ๊ฒฐ๊ณผ๋ฌผ์„ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์€ ์ฃผ๋กœ ์‘๋‹ต์˜ ์‚ฌ์‹ค์„ฑ ๋ฐ ํ˜•์‹๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ™˜์ž ์‚ฌ์‹ค์— ๋Œ€ํ•œ ์‚ฌ์‹ค ๋ถ€์ •ํ™• ๋ฐ ๋ถˆํ•„์š”ํ•œ ๋ฐ˜๋ณต ํ”ผํ•˜๊ธฐ, ๊ณต๊ฐ ํ‘œ์‹œ, ๋ช…ํ™•ํ•œ ํ˜•์‹์œผ๋กœ ํ‘œ์‹œ).
\begin{table} \begin{tabular}{|c|c|c|c|c|c|c|} \hline Parameter & \(\alpha_{N}\) & \(\alpha_{S}\) & \(\alpha_{B}\) & \(N_{c}\) & \(S_{c}\) & \(B_{*}\) \\ \hline \hline C4 Value & \(0.076\) & \(0.67\) & \(0.205\) & \(1.5\times 10^{14}\) & \(2.6\times 10^{3}\) & \(1.7\times 10^{8}\) \\ \hline **[dataset as the test set and performed a deduplication process to ensure the removal of any overlapping text from the training set. In total, we trained only 10 small models, each with a maximum of 60M parameters, to estimate the constant terms of the scaling-law formulas. The estimated constant terms are presented in Table 1. The actual and predicted loss trajectories of a 2B model (30 times larger than the small model used to estimate the constant terms) using the estimated formulas are depicted on the left of Figure 1. It can be observed that the predicted loss trajectory closely aligns with the actual loss trajectory. Our estimated constant terms are also not far from those estimated in [10] despite using different setups, possibly due to the similar distributions of C4 and WebText, both of which consist of crawled website text. This reinforces the assertion in [11] that the power term in scaling laws primarily relies on the data manifold.
\begin{table} \begin{tabular}{|c|c|c|c|c|c|c|} \hline Parameter & \(\alpha_{N}\) & \(\alpha_{S}\) & \(\alpha_{B}\) & \(N_{c}\) & \(S_{c}\) & \(B_{*}\) \\ \hline \hline C4 Value & \(0.076\) & \(0.67\) & \(0.205\) & \(1.5\times 10^{14}\) & \(2.6\times 10^{3}\) & \(1.7\times 10^{8}\) \\ \hline **[๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ์ค‘๋ณต๋œ ํ…์ŠคํŠธ๊ฐ€ ํ•™์Šต ์„ธํŠธ์—์„œ ์ œ๊ฑฐ๋˜๋„๋ก ์ค‘๋ณต ์ œ๊ฑฐ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด 10๊ฐœ์˜ ์†Œํ˜• ๋ชจ๋ธ๋งŒ ํ•™์Šตํ–ˆ์œผ๋ฉฐ, ๊ฐ ๋ชจ๋ธ์˜ ์ตœ๋Œ€ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” 60M์— ๋ถˆ๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค. Scaling-law ๊ณต์‹์˜ ์ƒ์ˆ˜ ํ•ญ์„ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”์ •๋œ ์ƒ์ˆ˜ ํ•ญ์€ ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ถ”์ •๋œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ 20์–ต ๋ชจ๋ธ(์ƒ์ˆ˜ ํ•ญ ์ถ”์ •์— ์‚ฌ์šฉ๋œ ์†Œํ˜• ๋ชจ๋ธ๋ณด๋‹ค 30๋ฐฐ ํฐ ๋ชจ๋ธ)์˜ ์‹ค์ œ Loss ๊ถค์ ๊ณผ ์˜ˆ์ธก๋œ Loss ๊ถค์ ์ด ๊ทธ๋ฆผ 1์˜ ์™ผ์ชฝ์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์ธก๋œ Loss ๊ถค์ ์ด ์‹ค์ œ Loss ๊ถค์ ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํฌ๋กค๋ง๋œ ์›น์‚ฌ์ดํŠธ ํ…์ŠคํŠธ๋กœ ๊ตฌ์„ฑ๋œ C4์™€ WebText์˜ ์œ ์‚ฌํ•œ ๋ถ„ํฌ๋กœ ์ธํ•ด ์ถ”์ •๋œ ์ƒ์ˆ˜ ํ•ญ์ด [10]์—์„œ ์ถ”์ •๋œ ๊ฒƒ๊ณผ ํฌ๊ฒŒ ๋‹ค๋ฅด์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Š” Scaling laws์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ํ•ญ์ด ์ฃผ๋กœ ๋ฐ์ดํ„ฐ ๋งค๋‹ˆํด๋“œ์— ์˜์กดํ•œ๋‹ค๋Š” [11]์˜ ์ฃผ์žฅ์„ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค.
\begin{table} \begin{tabular}{|c|c|c|c|c|c|c|} \hline Parameter & \(\alpha_{N}\) & \(\alpha_{S}\) & \(\alpha_{B}\) & \(N_{c}\) & \(S_{c}\) & \(B_{*}\) \\ \hline \hline C4 Value & \(0.076\) & \(0.67\) & \(0.205\) & \(1.5\times 10^{14}\) & \(2.6\times 10^{3}\) & \(1.7\times 10^{8}\) \\ \hline **[๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ์ค‘๋ณต๋œ ํ…์ŠคํŠธ๊ฐ€ ํ•™์Šต ์„ธํŠธ์—์„œ ์ œ๊ฑฐ๋˜๋„๋ก ์ค‘๋ณต ์ œ๊ฑฐ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด 10๊ฐœ์˜ ์†Œํ˜• ๋ชจ๋ธ๋งŒ ํ•™์Šตํ–ˆ์œผ๋ฉฐ, ๊ฐ ๋ชจ๋ธ์˜ ์ตœ๋Œ€ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” 60M์— ๋ถˆ๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ ๊ณต์‹์˜ ์ƒ์ˆ˜ ํ•ญ์„ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”์ •๋œ ์ƒ์ˆ˜ ํ•ญ์€ ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ถ”์ •๋œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ 20์–ต ๋ชจ๋ธ(์ƒ์ˆ˜ ํ•ญ ์ถ”์ •์— ์‚ฌ์šฉ๋œ ์†Œํ˜• ๋ชจ๋ธ๋ณด๋‹ค 30๋ฐฐ ํฐ ๋ชจ๋ธ)์˜ ์‹ค์ œ ์†์‹ค ๊ถค์ ๊ณผ ์˜ˆ์ธก๋œ ์†์‹ค ๊ถค์ ์ด ๊ทธ๋ฆผ 1์˜ ์™ผ์ชฝ์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์ธก๋œ ์†์‹ค ๊ถค์ ์ด ์‹ค์ œ ์†์‹ค ๊ถค์ ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํฌ๋กค๋ง๋œ ์›น์‚ฌ์ดํŠธ ํ…์ŠคํŠธ๋กœ ๊ตฌ์„ฑ๋œ C4์™€ ์›นํ…์ŠคํŠธ์˜ ์œ ์‚ฌํ•œ ๋ถ„ํฌ๋กœ ์ธํ•ด ์ถ”์ •๋œ ์ƒ์ˆ˜ ํ•ญ์ด [10]์—์„œ ์ถ”์ •๋œ ๊ฒƒ๊ณผ ํฌ๊ฒŒ ๋‹ค๋ฅด์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ํ•ญ์ด ์ฃผ๋กœ ๋ฐ์ดํ„ฐ ๋งค๋‹ˆํด๋“œ์— ์˜์กดํ•œ๋‹ค๋Š” [11]์˜ ์ฃผ์žฅ์„ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค.
**Deduplication** After implementing a filtration process, we've developed a comprehensive deduplication pipeline. This pipeline includes document-level exact deduplication, document-level Minhash deduplication, and intra-document-level similar line deduplication, effectively identifying and removing duplicate content within documents. For exact deduplication, to reduce memory pressure we utilize a Bloom filter to approximate with a false positive rate set at 0.001. In the case of Minhash LSH, the signature is constructed from 128 hash functions and organized into 9 bands and 13 rows for LSH, achieving a Jaccard similarity of 0.8. The intra-document-level similar line deduplication targets removing repetitive lines within a single document. This approach was motivated by our observation that a significant portion of web-crawled data contained repetitions of 2 to 3 times within the same page, and due to the process of extracting text from HTML, some words might be lost, leading to slight variations in duplicates. For this deduplication, we employ edit distance to determine line similarity. The specific criterion is that two lines are considered similar if their edit distance is less than one-tenth of the length of the shorter line.
**Deduplication** ์—ฌ๊ณผ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ตฌํ˜„ํ•œ ํ›„ ์ข…ํ•ฉ์ ์ธ deduplication ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์—๋Š” ๋ฌธ์„œ ์ˆ˜์ค€์˜ ์ •ํ™•ํ•œ deduplication, ๋ฌธ์„œ ์ˆ˜์ค€์˜ Minhash deduplication, ๋ฌธ์„œ ๋‚ด ์ˆ˜์ค€์˜ ์œ ์‚ฌํ•œ ์ค„ deduplication์ด ํฌํ•จ๋˜์–ด ๋ฌธ์„œ ๋‚ด ์ค‘๋ณต ์ฝ˜ํ…์ธ ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ deduplication์˜ ๊ฒฝ์šฐ ๋ฉ”๋ชจ๋ฆฌ ์••๋ ฅ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด Bloom filter๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์˜คํƒ์ง€์œจ์„ 0.001๋กœ ์„ค์ •ํ•˜์—ฌ ๊ทผ์‚ฌ์น˜๋ฅผ ๊ตฌํ•ฉ๋‹ˆ๋‹ค. Minhash LSH์˜ ๊ฒฝ์šฐ, 128๊ฐœ์˜ ํ•ด์‹œ ํ•จ์ˆ˜๋กœ ์„œ๋ช…์„ ๊ตฌ์„ฑํ•˜๊ณ  LSH๋ฅผ ์œ„ํ•ด 9๊ฐœ์˜ ๋ฐด๋“œ์™€ 13๊ฐœ์˜ ํ–‰์œผ๋กœ ๊ตฌ์„ฑํ•˜์—ฌ Jaccard ์œ ์‚ฌ๋„ 0.8์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ ๋‚ด ์œ ์‚ฌ์„ฑ ์ค„ deduplication์€ ๋‹จ์ผ ๋ฌธ์„œ ๋‚ด์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ค„์„ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์›น ํฌ๋กค๋ง๋œ ๋ฐ์ดํ„ฐ์˜ ์ƒ๋‹น ๋ถ€๋ถ„์ด ๋™์ผํ•œ ํŽ˜์ด์ง€ ๋‚ด์—์„œ 2~3ํšŒ ๋ฐ˜๋ณต๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜๊ณ , HTML์—์„œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ณผ์ •์—์„œ ์ผ๋ถ€ ๋‹จ์–ด๊ฐ€ ์†์‹ค๋˜์–ด ์ค‘๋ณต์ด ์•ฝ๊ฐ„ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ด€์ฐฐ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด deduplication์„ ์œ„ํ•ด edit distance๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ค„ ์œ ์‚ฌ์„ฑ์„ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ๊ธฐ์ค€์€ edit distance๊ฐ€ ์งง์€ ์ค„์˜ ๊ธธ์ด์˜ 10๋ถ„์˜ 1 ๋ฏธ๋งŒ์ธ ๊ฒฝ์šฐ ๋‘ ์ค„์ด ์œ ์‚ฌํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
**์ค‘๋ณต ์ œ๊ฑฐ** ์—ฌ๊ณผ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ตฌํ˜„ํ•œ ํ›„ ์ข…ํ•ฉ์ ์ธ ์ค‘๋ณต ์ œ๊ฑฐ ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์—๋Š” ๋ฌธ์„œ ์ˆ˜์ค€์˜ ์ •ํ™•ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ, ๋ฌธ์„œ ์ˆ˜์ค€์˜ Minhash ์ค‘๋ณต ์ œ๊ฑฐ, ๋ฌธ์„œ ๋‚ด ์ˆ˜์ค€์˜ ์œ ์‚ฌํ•œ ์ค„ ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ํฌํ•จ๋˜์–ด ๋ฌธ์„œ ๋‚ด ์ค‘๋ณต ์ฝ˜ํ…์ธ ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ์˜ ๊ฒฝ์šฐ ๋ฉ”๋ชจ๋ฆฌ ์••๋ ฅ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋ธ”๋ฃธ ํ•„ํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์˜คํƒ์ง€์œจ์„ 0.001๋กœ ์„ค์ •ํ•˜์—ฌ ๊ทผ์‚ฌ์น˜๋ฅผ ๊ตฌํ•ฉ๋‹ˆ๋‹ค. Minhash LSH์˜ ๊ฒฝ์šฐ, 128๊ฐœ์˜ ํ•ด์‹œ ํ•จ์ˆ˜๋กœ ์„œ๋ช…์„ ๊ตฌ์„ฑํ•˜๊ณ  LSH๋ฅผ ์œ„ํ•ด 9๊ฐœ์˜ ๋ฐด๋“œ์™€ 13๊ฐœ์˜ ํ–‰์œผ๋กœ ๊ตฌ์„ฑํ•˜์—ฌ Jaccard ์œ ์‚ฌ๋„ 0.8์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ ๋‚ด ์œ ์‚ฌ์„ฑ ์ค„ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋‹จ์ผ ๋ฌธ์„œ ๋‚ด์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ค„์„ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์›น ํฌ๋กค๋ง๋œ ๋ฐ์ดํ„ฐ์˜ ์ƒ๋‹น ๋ถ€๋ถ„์ด ๋™์ผํ•œ ํŽ˜์ด์ง€ ๋‚ด์—์„œ 2~3ํšŒ ๋ฐ˜๋ณต๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜๊ณ , HTML์—์„œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ณผ์ •์—์„œ ์ผ๋ถ€ ๋‹จ์–ด๊ฐ€ ์†์‹ค๋˜์–ด ์ค‘๋ณต์ด ์•ฝ๊ฐ„ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ด€์ฐฐ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์œ„ํ•ด ํŽธ์ง‘ ๊ฑฐ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ค„ ์œ ์‚ฌ์„ฑ์„ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ๊ธฐ์ค€์€ ํŽธ์ง‘ ๊ฑฐ๋ฆฌ๊ฐ€ ์งง์€ ์ค„์˜ ๊ธธ์ด์˜ 10๋ถ„์˜ 1 ๋ฏธ๋งŒ์ธ ๊ฒฝ์šฐ ๋‘ ์ค„์ด ์œ ์‚ฌํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Retrieval-Augmented Generation (RAG) merges retrieval methods with deep learning advancements to address the static limitations of large language models (LLMs) by enabling the dynamic integration of up-to-date external information. This methodology, focusing primarily on the text domain, provides a cost-effective solution to the generation of plausible but incorrect responses by LLMs, thereby enhancing the accuracy and reliability of their outputs through the use of real-world data. As RAG grows in complexity and incorporates multiple concepts that can influence its performance, this paper organizes the RAG paradigm into four categories: pre-retrieval, retrieval, post-retrieval, and generation, offering a detailed perspective from the retrieval viewpoint. It outlines RAG's evolution and discusses the field's progression through the analysis of significant studies. Additionally, the paper introduces evaluation methods for RAG, addressing the challenges faced and proposing future research directions. By offering an organized framework and categorization, the study aims to consolidate existing research on RAG, clarify its technological underpinnings, and highlight its potential to broaden the adaptability and applications of LLMs.
๋ฆฌํŠธ๋ฆฌ๋ฒ„-์ฆ๊ฐ• ์ƒ์„ฑ(Retrieval-Augmented Generation, RAG)์€ retrieval ๋ฐฉ๋ฒ•๊ณผ deep learning ๋ฐœ์ „์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์ตœ์‹  ์™ธ๋ถ€ ์ •๋ณด๋ฅผ ๋™์ ์œผ๋กœ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ Large Language Models (LLMs)์˜ ์ •์  ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ฃผ๋กœ ํ…์ŠคํŠธ ์˜์—ญ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ด ๋ฐฉ๋ฒ•๋ก ์€ LLM์ด ๊ทธ๋Ÿด๋“ฏํ•˜์ง€๋งŒ ์ž˜๋ชป๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋น„์šฉ ํšจ์œจ์ ์ธ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜์—ฌ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์˜ ์ •ํ™•์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ณต์žก์„ฑ์ด ์ปค์ง€๊ณ  ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ๋Š” ์—ฌ๋Ÿฌ ๊ฐœ๋…์ด ํ†ตํ•ฉ๋จ์— ๋”ฐ๋ผ ์ด ๋ฐฑ์„œ์—์„œ๋Š” retrieval ๊ด€์ ์—์„œ ์ž์„ธํ•œ ๊ด€์ ์„ ์ œ๊ณตํ•˜๋Š” pre-retrieval, retrieval, post-retrieval, generation์˜ ๋„ค ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ RAG ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ๋ฐฑ์„œ๋Š” RAG์˜ ์ง„ํ™”๋ฅผ ๊ฐœ๊ด„ํ•˜๊ณ  ์ฃผ์š” ์—ฐ๊ตฌ ๋ถ„์„์„ ํ†ตํ•ด ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ๋ฐฑ์„œ๋Š” RAG์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๊ณ  ์ง๋ฉดํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ฒด๊ณ„์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ์™€ ๋ถ„๋ฅ˜๋ฅผ ์ œ๊ณตํ•จ์œผ๋กœ์จ RAG์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜๊ณ , ๊ธฐ์ˆ ์  ํ† ๋Œ€๋ฅผ ๋ช…ํ™•ํžˆ ํ•˜๋ฉฐ, LLM์˜ ์ ์‘์„ฑ๊ณผ ์‘์šฉ ๋ถ„์•ผ๋ฅผ ๋„“ํž ์ˆ˜ ์žˆ๋Š” RAG์˜ ์ž ์žฌ๋ ฅ์„ ๊ฐ•์กฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
๋ฆฌํŠธ๋ฆฌ๋ฒ„-์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์€ ๋ฆฌํŠธ๋ฆฌ๋ฒ„ ๋ฐฉ๋ฒ•๊ณผ ๋”ฅ๋Ÿฌ๋‹ ๋ฐœ์ „์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์ตœ์‹  ์™ธ๋ถ€ ์ •๋ณด๋ฅผ ๋™์ ์œผ๋กœ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ์ •์  ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ฃผ๋กœ ํ…์ŠคํŠธ ์˜์—ญ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ด ๋ฐฉ๋ฒ•๋ก ์€ LLM์ด ๊ทธ๋Ÿด๋“ฏํ•˜์ง€๋งŒ ์ž˜๋ชป๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋น„์šฉ ํšจ์œจ์ ์ธ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜์—ฌ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์˜ ์ •ํ™•์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ณต์žก์„ฑ์ด ์ปค์ง€๊ณ  ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ๋Š” ์—ฌ๋Ÿฌ ๊ฐœ๋…์ด ํ†ตํ•ฉ๋จ์— ๋”ฐ๋ผ ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๊ฒ€์ƒ‰ ๊ด€์ ์—์„œ ์ž์„ธํ•œ ๊ด€์ ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒ€์ƒ‰ ์ „, ๊ฒ€์ƒ‰, ๊ฒ€์ƒ‰ ํ›„, ์ƒ์„ฑ์˜ ๋„ค ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ RAG ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ๋ฐฑ์„œ๋Š” RAG์˜ ์ง„ํ™”๋ฅผ ๊ฐœ๊ด„ํ•˜๊ณ  ์ฃผ์š” ์—ฐ๊ตฌ ๋ถ„์„์„ ํ†ตํ•ด ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ๋ฐฑ์„œ๋Š” RAG์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๊ณ  ์ง๋ฉดํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ฒด๊ณ„์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ์™€ ๋ถ„๋ฅ˜๋ฅผ ์ œ๊ณตํ•จ์œผ๋กœ์จ RAG์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜๊ณ , ๊ธฐ์ˆ ์  ํ† ๋Œ€๋ฅผ ๋ช…ํ™•ํžˆ ํ•˜๋ฉฐ, LLM์˜ ์ ์‘์„ฑ๊ณผ ์‘์šฉ ๋ถ„์•ผ๋ฅผ ๋„“ํž ์ˆ˜ ์žˆ๋Š” RAG์˜ ์ž ์žฌ๋ ฅ์„ ๊ฐ•์กฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
**Small Language Models.** "Small Language Models" (SLMs) is an evolving concept that has undergone significant transformations over time. Presently, SLMs are generally construed as models that are smaller in scale compared to the well-known LLMs, typically not exceeding 7 billion parameters. These models are distinguished by their capacity for deployment on end-user devices, such as personal computers and smartphones, even in the absence of a GPU. Notable examples within the current landscape of SLMs include the Phi series (Gunasekar et al., 2023; Li et al., 2023; Javaheripi and Bubeck, 2023), TinyLlama (Zhang et al., 2024), MobileLLM (Liu et al., 2024), and Gemma (Banks and Warkentin, 2024), etc. A variety of methodologies have been explored to augment the efficacy of SLMs. These include the incorporation of high quality data (Gunasekar et al., 2023; Li et al., 2023; Javaheripi and Bubeck, 2023), the application of structure pruning techniques (Xia et al., 2023), and the reconfiguration of model architectures (Liu et al., 2024), among others. MiniCPMenhances the capabilities of SLMs through a meticulous amalgamation of hyper-parameter optimization, strategic training methodologies, architectural design, and high quality data.
**์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ**. "์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ"(SLM)์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ƒ๋‹นํ•œ ๋ณ€ํ™”๋ฅผ ๊ฒช์–ด์˜จ ์ง„ํ™”ํ•˜๋Š” ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ SLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ž˜ ์•Œ๋ ค์ง„ Large Language Models (LLMs)์— ๋น„ํ•ด ๊ทœ๋ชจ๊ฐ€ ์ž‘์€ ๋ชจ๋ธ๋กœ, ์ผ๋ฐ˜์ ์œผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 70์–ต ๊ฐœ๋ฅผ ๋„˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ GPU๊ฐ€ ์—†์–ด๋„ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ๋‚˜ ์Šค๋งˆํŠธํฐ๊ณผ ๊ฐ™์€ ์ตœ์ข… ์‚ฌ์šฉ์ž ์žฅ์น˜์— ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด ํŠน์ง•์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ SLM์˜ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” Phi ์‹œ๋ฆฌ์ฆˆ(Gunasekar et al., 2023; Li et al., 2023; Javaheripi and Bubeck, 2023), TinyLlama(Zhang et al., 2024), MobileLLM(Liu et al., 2024), Gemma(Banks and Warkentin, 2024) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. SLM์˜ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋ก ์ด ํƒ์ƒ‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ํ†ตํ•ฉ(Gunasekar et al., 2023; Li et al., 2023; Javaheripi and Bubeck, 2023), ๊ตฌ์กฐ ๊ฐ€์ง€์น˜ ๊ธฐ๋ฒ• ์ ์šฉ(Xia et al., 2023), ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ์žฌ๊ตฌ์„ฑ(Liu et al., 2024) ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. MiniCPM์€ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ตœ์ ํ™”, ์ „๋žต์  ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก , ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„, ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์˜ ์„ธ์‹ฌํ•œ ์กฐํ•ฉ์„ ํ†ตํ•ด SLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
**์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ**. "์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ"(SLM)์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ƒ๋‹นํ•œ ๋ณ€ํ™”๋ฅผ ๊ฒช์–ด์˜จ ์ง„ํ™”ํ•˜๋Š” ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ SLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ž˜ ์•Œ๋ ค์ง„ LLM์— ๋น„ํ•ด ๊ทœ๋ชจ๊ฐ€ ์ž‘์€ ๋ชจ๋ธ๋กœ, ์ผ๋ฐ˜์ ์œผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 70์–ต ๊ฐœ๋ฅผ ๋„˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ GPU๊ฐ€ ์—†์–ด๋„ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ๋‚˜ ์Šค๋งˆํŠธํฐ๊ณผ ๊ฐ™์€ ์ตœ์ข… ์‚ฌ์šฉ์ž ์žฅ์น˜์— ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด ํŠน์ง•์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ SLM์˜ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” Phi ์‹œ๋ฆฌ์ฆˆ(Gunasekar et al., 2023; Li et al., 2023; Javaheripi and Bubeck, 2023), TinyLlama(Zhang et al., 2024), MobileLLM(Liu et al., 2024), Gemma(Banks and Warkentin, 2024) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. SLM์˜ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋ก ์ด ํƒ์ƒ‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ํ†ตํ•ฉ(Gunasekar et al., 2023; Li et al., 2023; Javaheripi and Bubeck, 2023), ๊ตฌ์กฐ ๊ฐ€์ง€์น˜ ๊ธฐ๋ฒ• ์ ์šฉ(Xia et al., 2023), ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ์žฌ๊ตฌ์„ฑ(Liu et al., 2024) ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. MiniCPM์€ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ตœ์ ํ™”, ์ „๋žต์  ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก , ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„, ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์˜ ์„ธ์‹ฌํ•œ ์กฐํ•ฉ์„ ํ†ตํ•ด SLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
\(\bullet\)_Multi-agent based applications._ Different from single-agent systems where agents work independently, multi-agent systems work in collaboration to unleash collective intelligence. Typically, multiple agents can be instantiated from the same or different LLMs, each with their respective roles and functions. According to the coordinating strategies among these agents, multi-agent systems can be divided into two categories: cooperation-based and competition-based. In the cooperation-based mode, to share information and seek collaborative actions among agents, various communication protocols have been proposed, including free-form dialogue [883], structured document [884], and data embedding [885]. Based on the communication protocol, agents can be effectively organized for downstream applications, such as software engineering [884], user behavior analysis [821, 819], and society simulation [533]. In the competition-based mode, debate serves as one of the popular communication protocols to foster divergent thinking and elicit valuable external feedback among agents. Such a way is beneficial for domains that demand precise decision-making and accurate responses, such as mathematical reasoning [886] and evaluation [732].
\(\bullet\)_Multi-agent ๊ธฐ๋ฐ˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜._ ์—์ด์ „ํŠธ๊ฐ€ ๋…๋ฆฝ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” ๋‹จ์ผ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ๊ณผ๋Š” ๋‹ฌ๋ฆฌ Multi-agent ์‹œ์Šคํ…œ์€ ํ˜‘๋ ฅํ•˜์—ฌ ์ง‘๋‹จ ์ง€๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋™์ผํ•˜๊ฑฐ๋‚˜ ๋‹ค๋ฅธ LLM์—์„œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ ์—์ด์ „ํŠธ๋Š” ๊ฐ์ž์˜ ์—ญํ• ๊ณผ ๊ธฐ๋Šฅ์„ ๊ฐ–์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์กฐ์ • ์ „๋žต์— ๋”ฐ๋ผ Multi-agent ์‹œ์Šคํ…œ์€ ํ˜‘๋ ฅ ๊ธฐ๋ฐ˜๊ณผ ๊ฒฝ์Ÿ ๊ธฐ๋ฐ˜์˜ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜‘๋ ฅ ๊ธฐ๋ฐ˜ ๋ชจ๋“œ์—์„œ๋Š” ์—์ด์ „ํŠธ ๊ฐ„์— ์ •๋ณด๋ฅผ ๊ณต์œ ํ•˜๊ณ  ํ˜‘๋ ฅ์ ์ธ ํ–‰๋™์„ ๋ชจ์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด ์ž์œ  ํ˜•์‹์˜ ๋Œ€ํ™”[883], ๊ตฌ์กฐํ™”๋œ ๋ฌธ์„œ[884], ๋ฐ์ดํ„ฐ embedding[885] ๋“ฑ ๋‹ค์–‘ํ•œ ํ†ต์‹  ํ”„๋กœํ† ์ฝœ์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ†ต์‹  ํ”„๋กœํ† ์ฝœ์— ๋”ฐ๋ผ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง[884], ์‚ฌ์šฉ์ž ํ–‰๋™ ๋ถ„์„[821, 819], ์‚ฌํšŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜[533] ๋“ฑ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์œ„ํ•ด ์—์ด์ „ํŠธ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์Ÿ ๊ธฐ๋ฐ˜ ๋ชจ๋“œ์—์„œ๋Š” ํ† ๋ก ์ด ์—์ด์ „ํŠธ ๊ฐ„์— ๋‹ค์–‘ํ•œ ์‚ฌ๊ณ ๋ฅผ ์ด‰์ง„ํ•˜๊ณ  ๊ฐ€์น˜ ์žˆ๋Š” ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ํ”„๋กœํ† ์ฝœ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์  ์ถ”๋ก [886] ๋ฐ ํ‰๊ฐ€[732]์™€ ๊ฐ™์ด ์ •ํ™•ํ•œ ์˜์‚ฌ ๊ฒฐ์ •๊ณผ ์ •ํ™•ํ•œ ์‘๋‹ต์ด ํ•„์š”ํ•œ ์˜์—ญ์— ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜._ ์—์ด์ „ํŠธ๊ฐ€ ๋…๋ฆฝ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” ๋‹จ์ผ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ๊ณผ๋Š” ๋‹ฌ๋ฆฌ ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ํ˜‘๋ ฅํ•˜์—ฌ ์ง‘๋‹จ ์ง€๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋™์ผํ•˜๊ฑฐ๋‚˜ ๋‹ค๋ฅธ LLM์—์„œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ ์—์ด์ „ํŠธ๋Š” ๊ฐ์ž์˜ ์—ญํ• ๊ณผ ๊ธฐ๋Šฅ์„ ๊ฐ–์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์กฐ์ • ์ „๋žต์— ๋”ฐ๋ผ ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ํ˜‘๋ ฅ ๊ธฐ๋ฐ˜๊ณผ ๊ฒฝ์Ÿ ๊ธฐ๋ฐ˜์˜ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜‘๋ ฅ ๊ธฐ๋ฐ˜ ๋ชจ๋“œ์—์„œ๋Š” ์—์ด์ „ํŠธ ๊ฐ„์— ์ •๋ณด๋ฅผ ๊ณต์œ ํ•˜๊ณ  ํ˜‘๋ ฅ์ ์ธ ํ–‰๋™์„ ๋ชจ์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด ์ž์œ  ํ˜•์‹์˜ ๋Œ€ํ™”[883], ๊ตฌ์กฐํ™”๋œ ๋ฌธ์„œ[884], ๋ฐ์ดํ„ฐ ์ž„๋ฒ ๋”ฉ[885] ๋“ฑ ๋‹ค์–‘ํ•œ ํ†ต์‹  ํ”„๋กœํ† ์ฝœ์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ†ต์‹  ํ”„๋กœํ† ์ฝœ์— ๋”ฐ๋ผ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง[884], ์‚ฌ์šฉ์ž ํ–‰๋™ ๋ถ„์„[821, 819], ์‚ฌํšŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜[533] ๋“ฑ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์œ„ํ•ด ์—์ด์ „ํŠธ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์Ÿ ๊ธฐ๋ฐ˜ ๋ชจ๋“œ์—์„œ๋Š” ํ† ๋ก ์ด ์—์ด์ „ํŠธ ๊ฐ„์— ๋‹ค์–‘ํ•œ ์‚ฌ๊ณ ๋ฅผ ์ด‰์ง„ํ•˜๊ณ  ๊ฐ€์น˜ ์žˆ๋Š” ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ํ”„๋กœํ† ์ฝœ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์  ์ถ”๋ก [886] ๋ฐ ํ‰๊ฐ€[732]์™€ ๊ฐ™์ด ์ •ํ™•ํ•œ ์˜์‚ฌ ๊ฒฐ์ •๊ณผ ์ •ํ™•ํ•œ ์‘๋‹ต์ด ํ•„์š”ํ•œ ์˜์—ญ์— ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
FLARE (Jiang et al., 2023) uses StrategyQA, which contains a significant number of yes/no questions from a diverse range of sources. In addition, the authors request that the model provide the exact reasoning process and the final answer that determines the yes/no answer, ensuring that the answer matches the gold answer exactly. The incorporation of in-context samples into the retrieved content, along with training using data from the COPA, HellaSwag, and PIQA datasets, has resulted in the development of LLM-R (Wang et al., 2023) model that exhibits excellent performance. The fundamental concept of the ITER-RETGEN (Gao et al., 2022) model employs an iterative methodology to integrate retrievers and language models. The model was trained for the Commonsense Reasoning task using the StrategyQA dataset and achieved its optimal performance at seven iterations. In contrast, KG-BART (Shao et al., 2023) is designed to prioritize the Commonsense Reasoning task and employs knowledge graphs to enhance its performance in this area. This approach has proven effective in significantly improving the model's ability to complete the Commonsense Reasoning task, with performance approaching that of human beings under certain evaluation metrics.
FLARE(Jiang et al., 2023)๋Š” ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์ƒ๋‹น์ˆ˜์˜ ์˜ˆ/์•„๋‹ˆ์š” ์งˆ๋ฌธ์ด ํฌํ•จ๋œ StrategyQA๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ €์ž๋Š” ๋ชจ๋ธ์ด ์˜ˆ/์•„๋‹ˆ์š” ๋‹ต๋ณ€์„ ๊ฒฐ์ •ํ•˜๋Š” ์ •ํ™•ํ•œ ์ถ”๋ก  ๊ณผ์ •๊ณผ ์ตœ์ข… ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์š”์ฒญํ•˜์—ฌ ๋‹ต๋ณ€์ด ๊ณจ๋“œ ๋‹ต๋ณ€๊ณผ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ In-Context ์ƒ˜ํ”Œ์„ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์— ํ†ตํ•ฉํ•˜๊ณ  COPA, HellaSwag ๋ฐ PIQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” LLM-R(Wang et al., 2023) ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ITER-RETGEN(Gao et al., 2022) ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ๊ฐœ๋…์€ ๋ฐ˜๋ณต์ ์ธ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์—ฌ Retrieval๊ณผ Language Models๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ StrategyQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Commonsense Reasoning ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ 7๋ฒˆ์˜ ๋ฐ˜๋ณต์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ KG-BART(Shao et al., 2023)๋Š” Commonsense Reasoning ์ž‘์—…์˜ ์šฐ์„  ์ˆœ์œ„๋ฅผ ์ •ํ•˜๊ณ  Knowledge Graphs๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด ๋ถ„์•ผ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํŠน์ • ํ‰๊ฐ€ ์ง€ํ‘œ์—์„œ ์ธ๊ฐ„ ์ˆ˜์ค€์— ๊ทผ์ ‘ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ชจ๋ธ์˜ Commonsense Reasoning ์ž‘์—… ์™„๋ฃŒ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
FLARE(Jiang et al., 2023)๋Š” ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์ƒ๋‹น์ˆ˜์˜ ์˜ˆ/์•„๋‹ˆ์š” ์งˆ๋ฌธ์ด ํฌํ•จ๋œ StrategyQA๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ €์ž๋Š” ๋ชจ๋ธ์ด ์˜ˆ/์•„๋‹ˆ์š” ๋‹ต๋ณ€์„ ๊ฒฐ์ •ํ•˜๋Š” ์ •ํ™•ํ•œ ์ถ”๋ก  ๊ณผ์ •๊ณผ ์ตœ์ข… ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์š”์ฒญํ•˜์—ฌ ๋‹ต๋ณ€์ด ๊ณจ๋“œ ๋‹ต๋ณ€๊ณผ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒํ™ฉ ์ƒ˜ํ”Œ์„ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์— ํ†ตํ•ฉํ•˜๊ณ  COPA, HellaSwag ๋ฐ PIQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” LLM-R(Wang et al., 2023) ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ITER-RETGEN(Gao et al., 2022) ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ๊ฐœ๋…์€ ๋ฐ˜๋ณต์ ์ธ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๊ธฐ์™€ ์–ธ์–ด ๋ชจ๋ธ์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ StrategyQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์‹ ์ถ”๋ก  ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ 7๋ฒˆ์˜ ๋ฐ˜๋ณต์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ KG-BART(Shao et al., 2023)๋Š” ์ƒ์‹ ์ถ”๋ก  ์ž‘์—…์˜ ์šฐ์„  ์ˆœ์œ„๋ฅผ ์ •ํ•˜๊ณ  ์ง€์‹ ๊ทธ๋ž˜ํ”„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด ๋ถ„์•ผ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํŠน์ • ํ‰๊ฐ€ ์ง€ํ‘œ์—์„œ ์ธ๊ฐ„ ์ˆ˜์ค€์— ๊ทผ์ ‘ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ชจ๋ธ์˜ ์ƒ์‹ ์ถ”๋ก  ์ž‘์—… ์™„๋ฃŒ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
The advent of ChatGPT has significantly impacted both academia and industry due to its interactive capabilities and widespread application, establishing itself as a leading artificial intelligence tool (Laskar et al., 2023; Jahan et al., 2023; Huang and Huang, 2024). At the core of ChatGPT is the large language model (LLM) GPT-4, as detailed by (OpenAI et al., 2023), which has seen numerous enhancements to its predecessors, showcasing exceptional abilities in a variety of Natural Language Processing (NLP) tasks (Laskar et al., 2020). Despite these advancements, the adoption of LLMs has highlighted several critical issues primarily due to their reliance on extensive datasets. This reliance restricts their ability to incorporate new information post-training, leading to three primary challenges. First, the focus on broad and general data to maximize accessibility and applicability results in subpar performance in specialized areas. Second, the rapid creation of online data, combined with the significant resources required for data annotation and model training, hinders LLMs' ability to stay updated. Third, LLMs are susceptible to generating convincing yet inaccurate responses, known as "hallucinations", which can mislead users.
ChatGPT์˜ ๋“ฑ์žฅ์€ ๋Œ€ํ™”ํ˜• ๊ธฐ๋Šฅ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ์œผ๋กœ ํ•™๊ณ„์™€ ์‚ฐ์—…๊ณ„์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์ณค์œผ๋ฉฐ, ์ธ๊ณต์ง€๋Šฅ ๋„๊ตฌ๋กœ์„œ ์„ ๋„์ ์ธ ์œ„์น˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Laskar et al., 2023; Jahan et al., 2023; Huang๊ณผ Huang, 2024). ChatGPT์˜ ํ•ต์‹ฌ์€ Large Language Model (LLM)์ธ GPT-4์ด๋ฉฐ, (OpenAI et al., 2023)์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์ด์ „ ๋ชจ๋ธ์— ๋น„ํ•ด ๋‹ค์–‘ํ•œ Natural Language Processing (NLP) ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ฐœ์„  ์‚ฌํ•ญ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค(Laskar et al., 2020). ์ด๋Ÿฌํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์˜ ์ฑ„ํƒ์€ ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์˜์กด์œผ๋กœ ์ธํ•ด ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ฌธ์ œ๋ฅผ ๋ถ€๊ฐ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜์กด์„ฑ์€ ํ•™์Šต ํ›„ ์ƒˆ๋กœ์šด ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ œํ•œํ•˜์—ฌ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ณผ์ œ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ ‘๊ทผ์„ฑ๊ณผ ์ ์šฉ์„ฑ์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ณ  ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ์— ์ดˆ์ ์„ ๋งž์ถ”๋‹ค ๋ณด๋‹ˆ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ์˜ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์˜จ๋ผ์ธ ๋ฐ์ดํ„ฐ์˜ ๊ธ‰์†ํ•œ ์ƒ์„ฑ๊ณผ ๋ฐ์ดํ„ฐ ์ฃผ์„๊ณผ ๋ชจ๋ธ ํ•™์Šต์— ํ•„์š”ํ•œ ์ƒ๋‹นํ•œ ๋ฆฌ์†Œ์Šค๋กœ ์ธํ•ด LLM์˜ ์—…๋ฐ์ดํŠธ ๋Šฅ๋ ฅ์ด ์ €ํ•ด๋ฉ๋‹ˆ๋‹ค. ์…‹์งธ, LLM์€ ์‚ฌ์šฉ์ž๋ฅผ ์˜ค๋„ํ•  ์ˆ˜ ์žˆ๋Š” 'Hallucination'์œผ๋กœ ์•Œ๋ ค์ง„ ์„ค๋“๋ ฅ ์žˆ์ง€๋งŒ ๋ถ€์ •ํ™•ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค.
ChatGPT์˜ ๋“ฑ์žฅ์€ ๋Œ€ํ™”ํ˜• ๊ธฐ๋Šฅ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ์œผ๋กœ ํ•™๊ณ„์™€ ์‚ฐ์—…๊ณ„์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์ณค์œผ๋ฉฐ, ์ธ๊ณต์ง€๋Šฅ ๋„๊ตฌ๋กœ์„œ ์„ ๋„์ ์ธ ์œ„์น˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Laskar et al., 2023; Jahan et al., 2023; Huang๊ณผ Huang, 2024). ChatGPT์˜ ํ•ต์‹ฌ์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ธ GPT-4์ด๋ฉฐ, (OpenAI et al., 2023)์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์ด์ „ ๋ชจ๋ธ์— ๋น„ํ•ด ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ฐœ์„  ์‚ฌํ•ญ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค(Laskar et al., 2020). ์ด๋Ÿฌํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์˜ ์ฑ„ํƒ์€ ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์˜์กด์œผ๋กœ ์ธํ•ด ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ฌธ์ œ๋ฅผ ๋ถ€๊ฐ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜์กด์„ฑ์€ ํ•™์Šต ํ›„ ์ƒˆ๋กœ์šด ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ œํ•œํ•˜์—ฌ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ณผ์ œ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ ‘๊ทผ์„ฑ๊ณผ ์ ์šฉ์„ฑ์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ณ  ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ์— ์ดˆ์ ์„ ๋งž์ถ”๋‹ค ๋ณด๋‹ˆ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ์˜ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์˜จ๋ผ์ธ ๋ฐ์ดํ„ฐ์˜ ๊ธ‰์†ํ•œ ์ƒ์„ฑ๊ณผ ๋ฐ์ดํ„ฐ ์ฃผ์„๊ณผ ๋ชจ๋ธ ํ•™์Šต์— ํ•„์š”ํ•œ ์ƒ๋‹นํ•œ ๋ฆฌ์†Œ์Šค๋กœ ์ธํ•ด LLM์˜ ์—…๋ฐ์ดํŠธ ๋Šฅ๋ ฅ์ด ์ €ํ•ด๋ฉ๋‹ˆ๋‹ค. ์…‹์งธ, LLM์€ ์‚ฌ์šฉ์ž๋ฅผ ์˜ค๋„ํ•  ์ˆ˜ ์žˆ๋Š” 'ํ™˜๊ฐ'์œผ๋กœ ์•Œ๋ ค์ง„ ์„ค๋“๋ ฅ ์žˆ์ง€๋งŒ ๋ถ€์ •ํ™•ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค.
Large "instruction-tuned" language models (i.e., finetuned to respond to instructions) have demonstrated a remarkable ability to generalize zero-shot to new tasks. Nevertheless, they depend heavily on human-written instruction data that is often limited in quantity, diversity, and creativity, therefore hindering the generality of the tuned model. We introduce Self-Instruct, a framework for improving the instruction-following capabilities of pretrained language models by bootstrapping off their own generations. Our pipeline generates instructions, input, and output samples from a language model, then filters invalid or similar ones before using them to finetune the original model. Applying our method to the vanilla GPT3, we demonstrate a 33% absolute improvement over the original model on Super-NaturalInstructions, on par with the performance of InstructGPT\({}_{001}\),1 which was trained with private user data and human annotations. For further evaluation, we curate a set of expert-written instructions for novel tasks, and show through human evaluation that tuning GPT3 with Self-Instruct outperforms using existing public instruction datasets by a large margin, leaving only a 5% absolute gap behind InstructGPT\({}_{001}\).
๋Œ€๊ทœ๋ชจ "Instruction-tuned" ์–ธ์–ด ๋ชจ๋ธ(์ฆ‰, Instruction์— ์‘๋‹ตํ•˜๋„๋ก Fine-tuning๋œ ๋ชจ๋ธ)์€ ์ƒˆ๋กœ์šด ์ž‘์—…์— zero-shot์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋†€๋ผ์šด ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ˆ˜๋Ÿ‰, ๋‹ค์–‘์„ฑ, ์ฐฝ์˜์„ฑ์ด ์ œํ•œ์ ์ธ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ Instruction ๋ฐ์ดํ„ฐ์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํŠœ๋‹๋œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์„ฑ์„ ์ €ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์ž์ฒด ์ƒ์„ฑ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ Instruction ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ ์–ธ์–ด ๋ชจ๋ธ์—์„œ Instruction, ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์œ ํšจํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ์œ ์‚ฌํ•œ ์ƒ˜ํ”Œ์„ ํ•„ํ„ฐ๋งํ•œ ํ›„ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ๋ฐ”๋‹๋ผ GPT3์— ์ ์šฉํ•˜์—ฌ Super-NaturalInstructions์—์„œ ์›๋ž˜ ๋ชจ๋ธ๋ณด๋‹ค 33% ์ ˆ๋Œ€์ ์œผ๋กœ ๊ฐœ์„ ๋˜์—ˆ์œผ๋ฉฐ, ๋น„๊ณต๊ฐœ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€ ์ธ๊ฐ„ ์ฃผ์„์œผ๋กœ ํ•™์Šต๋œ InstructGPT\({}_{001}\)์˜ ์„ฑ๋Šฅ๊ณผ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•œ Instruction ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•˜๊ณ , ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด Self-Instruct๋กœ GPT3๋ฅผ ํŠœ๋‹ํ•˜๋Š” ๊ฒƒ์ด ๊ธฐ์กด ๊ณต๊ฐœ Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, InstructGPT\({}_{001}\) ๋’ค์— 5%์˜ ์ ˆ๋Œ€ ๊ฒฉ์ฐจ๋งŒ ๋‚จ์•˜์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ "๋ช…๋ น์–ด ํŠœ๋‹" ์–ธ์–ด ๋ชจ๋ธ(์ฆ‰, ๋ช…๋ น์–ด์— ์‘๋‹ตํ•˜๋„๋ก ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ)์€ ์ƒˆ๋กœ์šด ์ž‘์—…์— 0์ƒท์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋†€๋ผ์šด ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ˆ˜๋Ÿ‰, ๋‹ค์–‘์„ฑ, ์ฐฝ์˜์„ฑ์ด ์ œํ•œ์ ์ธ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ๋ช…๋ น ๋ฐ์ดํ„ฐ์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํŠœ๋‹๋œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์„ฑ์„ ์ €ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์ž์ฒด ์ƒ์„ฑ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋ช…๋ น์–ด, ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์œ ํšจํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ์œ ์‚ฌํ•œ ์ƒ˜ํ”Œ์„ ํ•„ํ„ฐ๋งํ•œ ํ›„ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ๋ฐ”๋‹๋ผ GPT3์— ์ ์šฉํ•˜์—ฌ Super-NaturalInstructions์—์„œ ์›๋ž˜ ๋ชจ๋ธ๋ณด๋‹ค 33% ์ ˆ๋Œ€์ ์œผ๋กœ ๊ฐœ์„ ๋˜์—ˆ์œผ๋ฉฐ, ๋น„๊ณต๊ฐœ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€ ์ธ๊ฐ„ ์ฃผ์„์œผ๋กœ ํ•™์Šต๋œ InstructGPT\({}_{001}\)1์˜ ์„ฑ๋Šฅ๊ณผ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•œ ์ง€์นจ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•˜๊ณ , ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด Self-Instruct๋กœ GPT3๋ฅผ ํŠœ๋‹ํ•˜๋Š” ๊ฒƒ์ด ๊ธฐ์กด ๊ณต๊ฐœ ์ง€์นจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, InstructGPT\({}_{001}\) ๋’ค์— 5%์˜ ์ ˆ๋Œ€ ๊ฒฉ์ฐจ๋งŒ ๋‚จ์•˜์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
Despite the progress and impact, the underlying principles of LLMs are still not well explored. Firstly, it is mysterious why emergent abilities occur in LLMs, instead of smaller PLMs. As a more general issue, there lacks a deep, detailed investigation of the key factors that contribute to the superior abilities of LLMs. It is important to study when and how LLMs obtain such abilities [47]. Although there are some meaningful discussions about this problem [47, 31], more principled investigations are needed to uncover the _"secrets"_ of LLMs. Secondly, it is difficult for the research community to train capable LLMs. Due to the huge demand of computation resources, it is very costly to carry out repetitive, ablating studies for investigating the effect of various strategies for training LLMs. Indeed, LLMs are mainly trained by industry, where many important training details (_e.g.,_ data collection and cleaning) are not revealed to the public. Thirdly, it is challenging to align LLMs with human values or preferences. Despite the capacities, LLMs are also likely to produce toxic, fictitious, or harmful contents. It requires effective and efficient control approaches to eliminating the potential risk of the use of LLMs [46].
์ง„์ „๊ณผ ์˜ํ–ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์˜ ๊ธฐ๋ณธ ์›๋ฆฌ๋Š” ์•„์ง ์ž˜ ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋” ์ž‘์€ PLM์ด ์•„๋‹Œ LLM์—์„œ emergent abilities๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ์ด์œ ๊ฐ€ ์‹ ๋น„ํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ์ผ๋ฐ˜์ ์ธ ๋ฌธ์ œ๋กœ์„œ, LLM์˜ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ์— ๊ธฐ์—ฌํ•˜๋Š” ํ•ต์‹ฌ ์š”์ธ์— ๋Œ€ํ•œ ์‹ฌ๋„ ์žˆ๊ณ  ์ƒ์„ธํ•œ ์กฐ์‚ฌ๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. LLM์ด ์–ธ์ œ ์–ด๋–ป๊ฒŒ ๊ทธ๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋Š”์ง€ ์—ฐ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[47]. ์ด ๋ฌธ์ œ์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ๋…ผ์˜๊ฐ€ ์žˆ์ง€๋งŒ[47, 31], LLM์˜ '๋น„๋ฐ€'์„ ๋ฐํžˆ๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋ณด๋‹ค ์›์น™์ ์ธ ์กฐ์‚ฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ์œ ๋Šฅํ•œ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋ง‰๋Œ€ํ•œ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM ํ›ˆ๋ จ์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ „๋žต์˜ ํšจ๊ณผ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•œ ๋ฐ˜๋ณต์ ์ด๊ณ  ์ ˆ์ œ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ญ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ LLM์€ ์ฃผ๋กœ ์—…๊ณ„์—์„œ ํ›ˆ๋ จํ•˜๋ฉฐ, ๋งŽ์€ ์ค‘์š”ํ•œ ํ›ˆ๋ จ ์„ธ๋ถ€ ์‚ฌํ•ญ(_์˜ˆ:_ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ์ •๋ฆฌ)์ด ๋Œ€์ค‘์—๊ฒŒ ๊ณต๊ฐœ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์…‹์งธ, LLM์„ ์ธ๊ฐ„์˜ ๊ฐ€์น˜ ๋˜๋Š” ์„ ํ˜ธ๋„์— ๋งž์ถ”๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋Šฅ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์€ ์œ ๋…ํ•˜๊ฑฐ๋‚˜ ํ—ˆ๊ตฌ์ ์ด๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. LLM ์‚ฌ์šฉ์˜ ์ž ์žฌ์  ์œ„ํ—˜์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํšจ๊ณผ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ์ œ์–ด ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค[46].
์ง„์ „๊ณผ ์˜ํ–ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์˜ ๊ธฐ๋ณธ ์›๋ฆฌ๋Š” ์•„์ง ์ž˜ ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋” ์ž‘์€ PLM์ด ์•„๋‹Œ LLM์—์„œ ์—๋จธ์ „ํŠธ ๋Šฅ๋ ฅ์ด ๋ฐœ์ƒํ•˜๋Š” ์ด์œ ๊ฐ€ ์‹ ๋น„ํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ์ผ๋ฐ˜์ ์ธ ๋ฌธ์ œ๋กœ์„œ, LLM์˜ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ์— ๊ธฐ์—ฌํ•˜๋Š” ํ•ต์‹ฌ ์š”์ธ์— ๋Œ€ํ•œ ์‹ฌ๋„ ์žˆ๊ณ  ์ƒ์„ธํ•œ ์กฐ์‚ฌ๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. LLM์ด ์–ธ์ œ ์–ด๋–ป๊ฒŒ ๊ทธ๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋Š”์ง€ ์—ฐ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[47]. ์ด ๋ฌธ์ œ์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ๋…ผ์˜๊ฐ€ ์žˆ์ง€๋งŒ[47, 31], LLM์˜ '๋น„๋ฐ€'์„ ๋ฐํžˆ๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋ณด๋‹ค ์›์น™์ ์ธ ์กฐ์‚ฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ์œ ๋Šฅํ•œ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋ง‰๋Œ€ํ•œ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM ํ›ˆ๋ จ์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ „๋žต์˜ ํšจ๊ณผ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•œ ๋ฐ˜๋ณต์ ์ด๊ณ  ์ ˆ์ œ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ญ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ LLM์€ ์ฃผ๋กœ ์—…๊ณ„์—์„œ ํ›ˆ๋ จํ•˜๋ฉฐ, ๋งŽ์€ ์ค‘์š”ํ•œ ํ›ˆ๋ จ ์„ธ๋ถ€ ์‚ฌํ•ญ(_์˜ˆ:_ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ์ •๋ฆฌ)์ด ๋Œ€์ค‘์—๊ฒŒ ๊ณต๊ฐœ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์…‹์งธ, LLM์„ ์ธ๊ฐ„์˜ ๊ฐ€์น˜ ๋˜๋Š” ์„ ํ˜ธ๋„์— ๋งž์ถ”๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋Šฅ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์€ ์œ ๋…ํ•˜๊ฑฐ๋‚˜ ํ—ˆ๊ตฌ์ ์ด๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. LLM ์‚ฌ์šฉ์˜ ์ž ์žฌ์  ์œ„ํ—˜์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํšจ๊ณผ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ์ œ์–ด ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค[46].
**Prompt Tuning**. Different from prefix tuning, prompt tuning [397, 402] mainly focuses on incorporating trainable prompt vectors at the input layer32. Based on the discrete prompting methods [404, 405], it augments the input text by including a group of soft prompt tokens (either in a free form [402] or a prefix form [397]), and then takes the prompt-augmented input to solve specific downstream tasks. In implementation, task-specific prompt embeddings are combined with the input text embeddings, which are subsequently fed into language models. P-tuning [402] has proposed a free form to combine the context, prompt and target tokens, which can be applied to the architectures for both natural language understanding and generation. They further learn the representations of soft prompt tokens by a bidirectional LSTM. Another representative approach [397] named _prompt tuning_ directly prepends prefix prompts to the input. During training, only the prompt embeddings would be learned according to task-specific supervisions. Since this method only includes a small number of trainable parameters at the input layer, it has been found that the performance highly relies on the model capacity of the underlying language models [397].
**Prompt Tuning**. Prompt Tuning[397, 402]์€ Prefix Tuning๊ณผ๋Š” ๋‹ฌ๋ฆฌ ์ฃผ๋กœ ์ž…๋ ฅ ๋ ˆ์ด์–ด32์— ํ•™์Šต ๊ฐ€๋Šฅํ•œ prompt ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. Discrete Prompting Methods[404, 405]์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ soft prompt ํ† ํฐ ๊ทธ๋ฃน์„ ํฌํ•จํ•˜์—ฌ ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ฆ๊ฐ•ํ•œ ๋‹ค์Œ, prompt-augmented ์ž…๋ ฅ์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ๊ตฌํ˜„ ์‹œ, ์ž‘์—…๋ณ„ prompt embeddings์€ ์ž…๋ ฅ ํ…์ŠคํŠธ embeddings๊ณผ ๊ฒฐํ•ฉ๋˜์–ด ์–ธ์–ด ๋ชจ๋ธ์— ๊ณต๊ธ‰๋ฉ๋‹ˆ๋‹ค. P-tuning[402]์€ ์ž์—ฐ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ ์•„ํ‚คํ…์ฒ˜์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ปจํ…์ŠคํŠธ, prompt ๋ฐ ๋Œ€์ƒ ํ† ํฐ์„ ๊ฒฐํ•ฉํ•˜๋Š” ์ž์œ  ํ˜•์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์–‘๋ฐฉํ–ฅ LSTM์„ ํ†ตํ•ด soft prompt ํ† ํฐ์˜ ํ‘œํ˜„์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๋Œ€ํ‘œ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹[397]์ธ 'Prompt Tuning'์€ ์ž…๋ ฅ์— prefix prompts๋ฅผ ์ง์ ‘ ์•ž์— ๋ถ™์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์— prompt embeddings๋งŒ ์ž‘์—…๋ณ„ ๊ฐ๋…์— ๋”ฐ๋ผ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์†Œ์ˆ˜๋งŒ ํฌํ•จ๋˜๋ฏ€๋กœ ์„ฑ๋Šฅ์ด ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ํฌ๊ฒŒ ์˜์กดํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[397].
ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹**. ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹[397, 402]์€ ์ ‘๋‘์‚ฌ ํŠœ๋‹๊ณผ๋Š” ๋‹ฌ๋ฆฌ ์ฃผ๋กœ ์ž…๋ ฅ ๋ ˆ์ด์–ด32์— ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํ”„๋กฌํ”„ํŠธ ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด์‚ฐ ํ”„๋กฌํ”„ํŠธ ๋ฐฉ๋ฒ•[404, 405]์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ ์†Œํ”„ํŠธ ํ”„๋กฌํ”„ํŠธ ํ† ํฐ ๊ทธ๋ฃน์„ ํฌํ•จํ•˜์—ฌ ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ฆ๊ฐ•ํ•œ ๋‹ค์Œ, ํ”„๋กฌํ”„ํŠธ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ๊ตฌํ˜„ ์‹œ, ์ž‘์—…๋ณ„ ํ”„๋กฌํ”„ํŠธ ์ž„๋ฒ ๋”ฉ์€ ์ž…๋ ฅ ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ๊ณผ ๊ฒฐํ•ฉ๋˜์–ด ์–ธ์–ด ๋ชจ๋ธ์— ๊ณต๊ธ‰๋ฉ๋‹ˆ๋‹ค. P-ํŠœ๋‹[402]์€ ์ž์—ฐ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ ์•„ํ‚คํ…์ฒ˜์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ปจํ…์ŠคํŠธ, ํ”„๋กฌํ”„ํŠธ ๋ฐ ๋Œ€์ƒ ํ† ํฐ์„ ๊ฒฐํ•ฉํ•˜๋Š” ์ž์œ  ํ˜•์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์–‘๋ฐฉํ–ฅ LSTM์„ ํ†ตํ•ด ์†Œํ”„ํŠธ ํ”„๋กฌํ”„ํŠธ ํ† ํฐ์˜ ํ‘œํ˜„์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๋Œ€ํ‘œ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹[397]์ธ 'ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹'์€ ์ž…๋ ฅ์— ์ ‘๋‘์‚ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ง์ ‘ ์•ž์— ๋ถ™์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์— ํ”„๋กฌํ”„ํŠธ ์ž„๋ฒ ๋”ฉ๋งŒ ์ž‘์—…๋ณ„ ๊ฐ๋…์— ๋”ฐ๋ผ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์†Œ์ˆ˜๋งŒ ํฌํ•จ๋˜๋ฏ€๋กœ ์„ฑ๋Šฅ์ด ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ํฌ๊ฒŒ ์˜์กดํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[397].
In our DFS merging experiments, \(M=64,r=3\), and consequently, \(T=M\times r=192\). We kept the last 200 examples in the training data as our validation set and optimize on the rest of the data with a batch size of 200. We report the performance of the snapshot that achieved the highest accuracy in the validation set, and the test set is strictly isolated from the optimization process. We adopted CMA-ES in EvoJAX [46], it optimized \(\mathcal{I}\) and \(W\) for a total of 100 generations with a popluation size of 128, and we used the default hyper-parameters. We limited our DFS merging to two models \(A\) and \(B\) to ensure that the final model remains modest in size and can be run on a single GPU, but in principle, the methodology can scale to merging multiple models. During the merging, model \(A\)'stokenizer, and input/output embeddings are utilized. Furthermore, to maintain compatibility with the embedding layers, we mandate that the initial and final transformer layers of model \(A\) define the start and the end of the inference path. We initialized the indicator array \(\mathcal{I}\) so that all layers in model \(A\) are more likely to be included as initial hops in the inference path to shorten the search time.
DFS ๋ณ‘ํ•ฉ ์‹คํ—˜์—์„œ๋Š” \(M=64,r=3\)์ด๋ฉฐ, ๊ฒฐ๊ณผ์ ์œผ๋กœ \(T=M\times r=192\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ๋งˆ์ง€๋ง‰ 200๊ฐœ ์˜ˆ์ œ๋ฅผ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋กœ ์œ ์ง€ํ•˜๊ณ  ๋‚˜๋จธ์ง€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด Batch Size 200์œผ๋กœ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์—์„œ ๊ฐ€์žฅ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•œ ์Šค๋ƒ…์ƒท์˜ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋ฉฐ, ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” ์—„๊ฒฉํ•˜๊ฒŒ ์ตœ์ ํ™” ํ”„๋กœ์„ธ์Šค์—์„œ ๊ฒฉ๋ฆฌ๋ฉ๋‹ˆ๋‹ค. EvoJAX [46]์—์„œ CMA-ES๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ์ด 100์„ธ๋Œ€์— ๊ฑธ์ณ \(\mathcal{I}\)์™€ \(W\)๋ฅผ ์ตœ์ ํ™”ํ–ˆ์œผ๋ฉฐ, ๋ชจ์ง‘๋‹จ ํฌ๊ธฐ๋Š” 128๋กœ ํ•˜์—ฌ ๊ธฐ๋ณธ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์ข… ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ์ ๋‹นํ•˜๊ฒŒ ์œ ์ง€ํ•˜๊ณ  ๋‹จ์ผ GPU์—์„œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก DFS ๋ณ‘ํ•ฉ์„ ๋‘ ๊ฐœ์˜ ๋ชจ๋ธ \(A\)์™€ \(B\)๋กœ ์ œํ•œํ–ˆ์ง€๋งŒ, ์›์น™์ ์œผ๋กœ ์ด ๋ฐฉ๋ฒ•๋ก ์€ ์—ฌ๋Ÿฌ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์œผ๋กœ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉํ•˜๋Š” ๋™์•ˆ ๋ชจ๋ธ \(A\)์˜ ํ† ํฌ๋‚˜์ด์ € ๋ฐ ์ž…๋ ฅ/์ถœ๋ ฅ ์ž„๋ฒ ๋”ฉ์ด ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž„๋ฒ ๋”ฉ ๋ ˆ์ด์–ด์™€์˜ ํ˜ธํ™˜์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ \(A\)์˜ ์ดˆ๊ธฐ ๋ฐ ์ตœ์ข… Transformer ๋ ˆ์ด์–ด๊ฐ€ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์‹œ์ž‘๊ณผ ๋์„ ์ •์˜ํ•˜๋„๋ก ์˜๋ฌดํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๋ก  ์‹œ๊ฐ„์„ ๋‹จ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ \(A\)์˜ ๋ชจ๋“  ๋ ˆ์ด์–ด๊ฐ€ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ดˆ๊ธฐ ํ™‰์œผ๋กœ ํฌํ•จ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋„๋ก ์ธ๋””์ผ€์ดํ„ฐ ๋ฐฐ์—ด \(\mathcal{I}\)๋ฅผ ์ดˆ๊ธฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.
DFS ๋ณ‘ํ•ฉ ์‹คํ—˜์—์„œ๋Š” \(M=64,r=3\)์ด๋ฉฐ, ๊ฒฐ๊ณผ์ ์œผ๋กœ \(T=M\๋ฐฐ r=192\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ๋งˆ์ง€๋ง‰ 200๊ฐœ ์˜ˆ์ œ๋ฅผ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋กœ ์œ ์ง€ํ•˜๊ณ  ๋‚˜๋จธ์ง€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฐฐ์น˜ ํฌ๊ธฐ 200์œผ๋กœ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์—์„œ ๊ฐ€์žฅ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•œ ์Šค๋ƒ…์ƒท์˜ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋ฉฐ, ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” ์—„๊ฒฉํ•˜๊ฒŒ ์ตœ์ ํ™” ํ”„๋กœ์„ธ์Šค์—์„œ ๊ฒฉ๋ฆฌ๋ฉ๋‹ˆ๋‹ค. EvoJAX [46]์—์„œ CMA-ES๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ์ด 100์„ธ๋Œ€์— ๊ฑธ์ณ \(\mathcal{I}\)์™€ \(W\)๋ฅผ ์ตœ์ ํ™”ํ–ˆ์œผ๋ฉฐ, ๋ชจ์ง‘๋‹จ ํฌ๊ธฐ๋Š” 128๋กœ ํ•˜์—ฌ ๊ธฐ๋ณธ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์ข… ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ์ ๋‹นํ•˜๊ฒŒ ์œ ์ง€ํ•˜๊ณ  ๋‹จ์ผ GPU์—์„œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก DFS ๋ณ‘ํ•ฉ์„ ๋‘ ๊ฐœ์˜ ๋ชจ๋ธ \(A\)์™€ \(B\)๋กœ ์ œํ•œํ–ˆ์ง€๋งŒ, ์›์น™์ ์œผ๋กœ ์ด ๋ฐฉ๋ฒ•๋ก ์€ ์—ฌ๋Ÿฌ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์œผ๋กœ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉํ•˜๋Š” ๋™์•ˆ ๋ชจ๋ธ \(A\)์˜ ํ† ํฐํ™” ๋ฐ ์ž…๋ ฅ/์ถœ๋ ฅ ์ž„๋ฒ ๋”ฉ์ด ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž„๋ฒ ๋”ฉ ๋ ˆ์ด์–ด์™€์˜ ํ˜ธํ™˜์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ \(A\)์˜ ์ดˆ๊ธฐ ๋ฐ ์ตœ์ข… ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด๊ฐ€ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์‹œ์ž‘๊ณผ ๋์„ ์ •์˜ํ•˜๋„๋ก ์˜๋ฌดํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๋ก  ์‹œ๊ฐ„์„ ๋‹จ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ \(A\)์˜ ๋ชจ๋“  ๋ ˆ์ด์–ด๊ฐ€ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ดˆ๊ธฐ ํ™‰์œผ๋กœ ํฌํ•จ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋„๋ก ์ธ๋””์ผ€์ดํ„ฐ ๋ฐฐ์—ด \(\mathcal{I}\)๋ฅผ ์ดˆ๊ธฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.
**Overview.** The basic idea of non-RL alignment approaches is to directly fine-tune LLMs with _supervised learning_ on high-quality _alignment dataset_. It basically assumes that response feedback or golden rules to avert unsafe behaviors have been injected or included in the specially curated alignment dataset, so that LLMs can directly learn aligned behaviors from these demonstration data via suitable fine-tuning strategies. Thus, to implement this approach, two key issues are the construction of alignment dataset and the design of fine-tuning loss. For the first issue, the alignment dataset can be automatically constructed by an aligned LLMs according to human-written safety principles [347] or refining existing examples using edits operations [383]. In addition, we can also reuse existing reward models to select high-rated responses from existing human feedback data [373]. For the second issue, non-RL alignment approaches mainly fine-tune LLMs in a supervised learning way (the same as the original instruction tuning loss) on a high-quality alignment dataset, meanwhile auxiliary learning objectives can be used to enhance the alignment performance, _e.g._, ranking responses or contrasting instruction-response pairs.
**๊ฐœ์š”**. ๋น„RL ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๊ธฐ๋ณธ ์•„์ด๋””์–ด๋Š” ๊ณ ํ’ˆ์งˆ _alignment dataset_์— ๋Œ€ํ•œ _supervised learning_์„ ํ†ตํ•ด LLM์„ ์ง์ ‘ Fine-tuningํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ธฐ๋ณธ์ ์œผ๋กœ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ํ–‰๋™์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•œ ์‘๋‹ต ํ”ผ๋“œ๋ฐฑ ๋˜๋Š” ํ™ฉ๊ธˆ๋ฅ ์ด ํŠน๋ณ„ํžˆ ํ๋ ˆ์ด์…˜๋œ alignment dataset์— ์ฃผ์ž…๋˜๊ฑฐ๋‚˜ ํฌํ•จ๋˜์–ด ์žˆ์–ด, LLM์ด ์ ์ ˆํ•œ Fine-tuning ์ „๋žต์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ์—์„œ ์ •๋ ฌ๋œ ํ–‰๋™์„ ์ง์ ‘ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” alignment dataset์˜ ๊ตฌ์„ฑ๊ณผ Fine-tuning loss์˜ ์„ค๊ณ„๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฌธ์ œ์— ๋Œ€ํ•ด alignment dataset์€ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ์•ˆ์ „ ์›์น™์— ๋”ฐ๋ผ ์ •๋ ฌ๋œ LLM์— ์˜ํ•ด ์ž๋™์œผ๋กœ ๊ตฌ์„ฑ๋˜๊ฑฐ๋‚˜ [347] ํŽธ์ง‘ ์ž‘์—…์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ์กด ์˜ˆ์ œ๋ฅผ ๊ฐœ์„ ํ•˜๊ฑฐ๋‚˜ [383] ๊ธฐ์กด ๋ณด์ƒ ๋ชจ๋ธ์„ ์žฌ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ์กด ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ์—์„œ ๋†’์€ ๋“ฑ๊ธ‰์˜ ์‘๋‹ต์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋น„RL ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ฃผ๋กœ ๊ณ ํ’ˆ์งˆ alignment dataset์—์„œ supervised learning ๋ฐฉ์‹์œผ๋กœ LLM์„ Fine-tuningํ•˜๋Š” ๋™์‹œ์— ๋ณด์กฐ ํ•™์Šต ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •๋ ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ์‘๋‹ต ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๋˜๋Š” ๋ช…๋ น์–ด-์‘๋‹ต ์Œ ๋Œ€์กฐ).
**๊ฐœ์š”**. ๋น„RL ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๊ธฐ๋ณธ ์•„์ด๋””์–ด๋Š” ๊ณ ํ’ˆ์งˆ _์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ_์— ๋Œ€ํ•œ _์ง€๋„ ํ•™์Šต_์„ ํ†ตํ•ด LLM์„ ์ง์ ‘ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ธฐ๋ณธ์ ์œผ๋กœ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ํ–‰๋™์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•œ ์‘๋‹ต ํ”ผ๋“œ๋ฐฑ ๋˜๋Š” ํ™ฉ๊ธˆ๋ฅ ์ด ํŠน๋ณ„ํžˆ ํ๋ ˆ์ด์…˜๋œ ์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ฃผ์ž…๋˜๊ฑฐ๋‚˜ ํฌํ•จ๋˜์–ด ์žˆ์–ด, LLM์ด ์ ์ ˆํ•œ ๋ฏธ์„ธ ์กฐ์ • ์ „๋žต์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ์—์„œ ์ •๋ ฌ๋œ ํ–‰๋™์„ ์ง์ ‘ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ตฌ์„ฑ๊ณผ ๋ฏธ์„ธ ์กฐ์ • ์†์‹ค์˜ ์„ค๊ณ„๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ์•ˆ์ „ ์›์น™์— ๋”ฐ๋ผ ์ •๋ ฌ๋œ LLM์— ์˜ํ•ด ์ž๋™์œผ๋กœ ๊ตฌ์„ฑ๋˜๊ฑฐ๋‚˜ [347] ํŽธ์ง‘ ์ž‘์—…์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ์กด ์˜ˆ์ œ๋ฅผ ๊ฐœ์„ ํ•˜๊ฑฐ๋‚˜ [383] ๊ธฐ์กด ๋ณด์ƒ ๋ชจ๋ธ์„ ์žฌ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ์กด ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ์—์„œ ๋†’์€ ๋“ฑ๊ธ‰์˜ ์‘๋‹ต์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋น„RL ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ฃผ๋กœ ๊ณ ํ’ˆ์งˆ ์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ง€๋„ ํ•™์Šต ๋ฐฉ์‹์œผ๋กœ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋™์‹œ์— ๋ณด์กฐ ํ•™์Šต ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •๋ ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ์‘๋‹ต ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๋˜๋Š” ๋ช…๋ น์–ด-์‘๋‹ต ์Œ ๋Œ€์กฐ).
To ensure a fair comparison between Adapt-LLM and the popularity-based method, we replicated their approach using the Llama-2 TB model to determine the best popularity score threshold (found to be 707,000) using the same PopQA development set. This allowed us to obtain results consistent with their methodology while utilizing our base LLM. Similar to the original results in Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22] when using smaller models, the popularity score threshold is almost equivalent to always retrieving contextual information for Llama-2 7B. The IR usage is of 99.86% as presented in Table 5. This clearly shows how the popularity score method struggles with smaller size models, being GPT-3 davinci-003 the only model to get a IR usage below 80% in the original paper when using adaptive retrieval with the Contriever. Subsequently, we evaluated our Adapt-LLM configuration on the same 25% test set split and compared the outcomes with those obtained using the method described by Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22] with the Llama-2 LLM as the base model.
์–ด๋Œ‘ํŠธ-LLM๊ณผ ์ธ๊ธฐ๋„ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ• ๊ฐ„์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด Llama-2 TB ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋™์ผํ•œ PopQA ๊ฐœ๋ฐœ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ์˜ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ์ž„๊ณ„๊ฐ’(707,000์œผ๋กœ ํ™•์ธ๋จ)์„ ์ฐพ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณต์ œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๊ธฐ๋ณธ LLM์„ ํ™œ์šฉํ•˜๋ฉด์„œ๋„ ๊ทธ๋“ค์˜ ๋ฐฉ๋ฒ•๋ก ๊ณผ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์˜ ์›๋ž˜ ๊ฒฐ๊ณผ์™€ ์œ ์‚ฌํ•˜๊ฒŒ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ๋•Œ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ์ž„๊ณ„๊ฐ’์€ Llama-2 7B์— ๋Œ€ํ•œ ๋ฌธ๋งฅ ์ •๋ณด๋ฅผ ํ•ญ์ƒ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฑฐ์˜ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 5์— ์ œ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด IR ์‚ฌ์šฉ๋ฅ ์€ 99.86%์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ธ๊ธฐ๋„ ์ ์ˆ˜ ๋ฐฉ์‹์ด ์ž‘์€ ํฌ๊ธฐ์˜ ๋ชจ๋ธ, ํŠนํžˆ Contriever๋กœ adaptive retrieval์„ ์‚ฌ์šฉํ•  ๋•Œ ์›๋ž˜ ๋…ผ๋ฌธ์—์„œ GPT-3 davinci-003์ด ์œ ์ผํ•˜๊ฒŒ IR ์‚ฌ์šฉ๋ฅ ์ด 80% ๋ฏธ๋งŒ์ธ ๋ชจ๋ธ์ด๋ผ๋Š” ์ ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ดํ›„ ๋™์ผํ•œ 25% ํ…Œ์ŠคํŠธ ์„ธํŠธ ๋ถ„ํ• ์— ๋Œ€ํ•ด Adapt-LLM ๊ตฌ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ทธ ๊ฒฐ๊ณผ๋ฅผ Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์—์„œ ์„ค๋ช…ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ป์€ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. base ๋ชจ๋ธ๋กœ Llama-2 LLM์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
์–ด๋Œ‘ํŠธ-LLM๊ณผ ์ธ๊ธฐ๋„ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ• ๊ฐ„์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด Llama-2 TB ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋™์ผํ•œ PopQA ๊ฐœ๋ฐœ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ์˜ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ์ž„๊ณ„๊ฐ’(707,000์œผ๋กœ ํ™•์ธ๋จ)์„ ์ฐพ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณต์ œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๊ธฐ๋ณธ LLM์„ ํ™œ์šฉํ•˜๋ฉด์„œ๋„ ๊ทธ๋“ค์˜ ๋ฐฉ๋ฒ•๋ก ๊ณผ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์˜ ์›๋ž˜ ๊ฒฐ๊ณผ์™€ ์œ ์‚ฌํ•˜๊ฒŒ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ๋•Œ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ์ž„๊ณ„๊ฐ’์€ Llama-2 7B์— ๋Œ€ํ•œ ๋ฌธ๋งฅ ์ •๋ณด๋ฅผ ํ•ญ์ƒ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฑฐ์˜ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 5์— ์ œ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด IR ์‚ฌ์šฉ๋ฅ ์€ 99.86%์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ธ๊ธฐ๋„ ์ ์ˆ˜ ๋ฐฉ์‹์ด ์ž‘์€ ํฌ๊ธฐ์˜ ๋ชจ๋ธ, ํŠนํžˆ Contriever๋กœ ์ ์‘ํ˜• ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•  ๋•Œ ์›๋ž˜ ๋…ผ๋ฌธ์—์„œ GPT-3 davinci-003์ด ์œ ์ผํ•˜๊ฒŒ IR ์‚ฌ์šฉ๋ฅ ์ด 80% ๋ฏธ๋งŒ์ธ ๋ชจ๋ธ์ด๋ผ๋Š” ์ ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ดํ›„ ๋™์ผํ•œ 25% ํ…Œ์ŠคํŠธ ์„ธํŠธ ๋ถ„ํ• ์— ๋Œ€ํ•ด Adapt-LLM ๊ตฌ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ทธ ๊ฒฐ๊ณผ๋ฅผ Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์—์„œ ์„ค๋ช…ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ป์€ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. base ๋ชจ๋ธ๋กœ Llama-2 LLM์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
The OLMo framework encompasses the tools and resources required for building and researching language models. For training and modeling, it includes full model weights, training code, training logs, ablations, training metrics in the form of Weights & Biases logs, and inference code. This first release includes four variants of our language model at the 7B scale corresponding to different architectures, optimizers, and training hardware, and one model at the 1B scale, all trained on at least 2T tokens. We are also releasing hundreds of intermediate checkpoints available as revisions on HuggingFace. For dataset building and analysis, it includes the full training data used for these models, including code that produces the training data, from AI2's Dolma (Soldaini et al., 2024), and WIMBD (Elazar et al., 2023) for analyzing pretraining data. For evaluation, it includes AI2's Catwalk (Groeneveld et al., 2023) for downstream evaluation and Paloma (Magnusson et al., 2023) for perplexity-based evaluation. For instruction-tuning, we released Open Instruct (Ivison et al., 2023; Wang et al., 2023), and we are currently using it to produce an adapted (instruction-tuned and RLHFed) version of OLMo, which we will release soon.
OLMo ํ”„๋ ˆ์ž„์›Œํฌ์—๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์—ฐ๊ตฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋„๊ตฌ์™€ ๋ฆฌ์†Œ์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ ๋ชจ๋ธ๋ง์—๋Š” ์ „์ฒด ๋ชจ๋ธ ๊ฐ€์ค‘์น˜, ํ›ˆ๋ จ ์ฝ”๋“œ, ํ›ˆ๋ จ ๋กœ๊ทธ, ์ ˆ์ œ, Weights & Biases ๋กœ๊ทธ ํ˜•์‹์˜ ํ›ˆ๋ จ ๋ฉ”ํŠธ๋ฆญ, ์ถ”๋ก  ์ฝ”๋“œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ์ฒซ ๋ฒˆ์งธ ๋ฆด๋ฆฌ์Šค์—๋Š” ๋‹ค์–‘ํ•œ ์•„ํ‚คํ…์ฒ˜, optimizers ๋ฐ ํ›ˆ๋ จ ํ•˜๋“œ์›จ์–ด์— ํ•ด๋‹นํ•˜๋Š” 7B ๊ทœ๋ชจ์˜ ์–ธ์–ด ๋ชจ๋ธ 4๊ฐ€์ง€ ๋ณ€ํ˜•๊ณผ 1B ๊ทœ๋ชจ์˜ ๋ชจ๋ธ 1๊ฐ€์ง€๊ฐ€ ํฌํ•จ๋˜๋ฉฐ, ๋ชจ๋‘ ์ตœ์†Œ 2T ํ† ํฐ ์ด์ƒ์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ HuggingFace์—์„œ ์ˆ˜์ •๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜๋ฐฑ ๊ฐœ์˜ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ๋„ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ตฌ์ถ• ๋ฐ ๋ถ„์„์—๋Š” AI2์˜ Dolma(Soldaini et al., 2024)์—์„œ ์ƒ์„ฑ๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ์‚ฌ์šฉ๋œ ์ „์ฒด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ฝ”๋“œ, ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๋ถ„์„์„ ์œ„ํ•œ WIMBD(Elazar et al., 2023)๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€์—๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ AI2์˜ Catwalk(Groeneveld et al., 2023)๊ณผ Perplexity ๊ธฐ๋ฐ˜ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ Paloma(Magnusson et al., 2023)๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Instruction Tuning์„ ์œ„ํ•ด Open Instruct(Ivison et al., 2023; Wang et al., 2023)์„ ๊ณต๊ฐœํ–ˆ์œผ๋ฉฐ, ํ˜„์žฌ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ OLMo์˜ ์กฐ์ •๋œ(Instruction Tuning ๋ฐ RLHF) ๋ฒ„์ „์„ ์ œ์ž‘ ์ค‘์ด๋ฉฐ, ๊ณง ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
OLMo ํ”„๋ ˆ์ž„์›Œํฌ์—๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์—ฐ๊ตฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋„๊ตฌ์™€ ๋ฆฌ์†Œ์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ ๋ชจ๋ธ๋ง์—๋Š” ์ „์ฒด ๋ชจ๋ธ ๊ฐ€์ค‘์น˜, ํ›ˆ๋ จ ์ฝ”๋“œ, ํ›ˆ๋ จ ๋กœ๊ทธ, ์ ˆ์ œ, ์›จ์ดํŠธ & ๋ฐ”์ด์ฆˆ ๋กœ๊ทธ ํ˜•์‹์˜ ํ›ˆ๋ จ ๋ฉ”ํŠธ๋ฆญ, ์ถ”๋ก  ์ฝ”๋“œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ์ฒซ ๋ฒˆ์งธ ๋ฆด๋ฆฌ์Šค์—๋Š” ๋‹ค์–‘ํ•œ ์•„ํ‚คํ…์ฒ˜, ์ตœ์ ํ™” ๋„๊ตฌ ๋ฐ ํ›ˆ๋ จ ํ•˜๋“œ์›จ์–ด์— ํ•ด๋‹นํ•˜๋Š” 7B ๊ทœ๋ชจ์˜ ์–ธ์–ด ๋ชจ๋ธ 4๊ฐ€์ง€ ๋ณ€ํ˜•๊ณผ 1B ๊ทœ๋ชจ์˜ ๋ชจ๋ธ 1๊ฐ€์ง€๊ฐ€ ํฌํ•จ๋˜๋ฉฐ, ๋ชจ๋‘ ์ตœ์†Œ 2T ํ† ํฐ ์ด์ƒ์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ HuggingFace์—์„œ ์ˆ˜์ •๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜๋ฐฑ ๊ฐœ์˜ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ๋„ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ตฌ์ถ• ๋ฐ ๋ถ„์„์—๋Š” AI2์˜ Dolma(Soldaini et al., 2024)์—์„œ ์ƒ์„ฑ๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ์‚ฌ์šฉ๋œ ์ „์ฒด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ฝ”๋“œ, ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๋ถ„์„์„ ์œ„ํ•œ WIMBD(Elazar et al., 2023)๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€์—๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ AI2์˜ Catwalk(Groeneveld et al., 2023)๊ณผ ๋‚œํ•ด๋„ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ Paloma(Magnusson et al., 2023)๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ช…๋ น์–ด ํŠœ๋‹์„ ์œ„ํ•ด Open Instruct(Ivison et al., 2023; Wang et al., 2023)์„ ๊ณต๊ฐœํ–ˆ์œผ๋ฉฐ, ํ˜„์žฌ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ OLMo์˜ ์กฐ์ •๋œ(๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ RLHF) ๋ฒ„์ „์„ ์ œ์ž‘ ์ค‘์ด๋ฉฐ, ๊ณง ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
In this work, we limit ourselves to prompts that express a composition of two facts such as _"The mother of the singer of 'Superstition' is"_ that humans can complete with two hops by (1) inferring a _bridge entity_ (e.g., Stevie Wonder) and (2) inferring an attribute of that entity (e.g., who his mother is). Then, we study how often LLMs process the prompt using a similar latent two-hop reasoning pathway, although this pathway may not be the most salient pathway that largely determines the predicted output. To this end, we first study these hops individually, as shown in Figure 1. To study the first hop, we propose _entity recall score_ to approximate LLM's internal recall of the bridge entity by projecting specific hidden representations to vocabulary space. We test how changes to the input prompt affect this score. To study the second hop, we measure _consistency score_ between the distribution for completions of the two-hop prompt and an equivalent recall-based one-hop prompt (e.g., _"The mother of Stevie Wonder is"_). We check how often an intervention to increase the entity recall score increases consistency as an indication of second-hop utilization. Finally, we investigate how frequently both steps coincide.
์ด ์ž‘์—…์—์„œ๋Š” ์ธ๊ฐ„์ด (1) _bridge entity_ (์˜ˆ: 'Superstition'์˜ ๊ฐ€์ˆ˜) ๋ฅผ ์ถ”๋ก ํ•˜๊ณ  (2) ํ•ด๋‹น ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ (์˜ˆ: ๊ทธ์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ˆ„๊ตฌ์ธ์ง€)์„ ์ถ”๋ก ํ•˜์—ฌ ๋‘ ๋ฒˆ์˜ ํ™‰์œผ๋กœ ์™„์„ฑํ•  ์ˆ˜ ์žˆ๋Š” _๋‘ ๊ฐ€์ง€ ์‚ฌ์‹ค์˜ ๊ตฌ์„ฑ์„ ํ‘œํ˜„ํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์— ๊ตญํ•œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, LLM์ด ์˜ˆ์ธก๋œ ์ถœ๋ ฅ์„ ํฌ๊ฒŒ ๊ฒฐ์ •ํ•˜๋Š” ๊ฐ€์žฅ ๋ˆˆ์— ๋„๋Š” ๊ฒฝ๋กœ๋Š” ์•„๋‹ˆ๋”๋ผ๋„ ์œ ์‚ฌํ•œ ์ž ์žฌ์  2-hop ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋นˆ๋„๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋จผ์ € ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ด๋Ÿฌํ•œ ํ™‰์„ ๊ฐœ๋ณ„์ ์œผ๋กœ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ํ™‰์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ์ˆจ๊ฒจ์ง„ ํ‘œํ˜„์„ ์–ดํœ˜ ๊ณต๊ฐ„์— ํˆฌ์˜ํ•˜์—ฌ LLM์˜ ๋‚ด๋ถ€์—์„œ bridge entity๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ์„ ๊ทผ์‚ฌํ™”ํ•˜๋Š” _entity recall score_๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๋ณ€๊ฒฝ์ด ์ด ์ ์ˆ˜์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ™‰์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด 2-hop ํ”„๋กฌํ”„ํŠธ์˜ ์™„์„ฑ์„ ์œ„ํ•œ ์™„์„ฑ๊ณผ ๋™๋“ฑํ•œ ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ 1-hop ํ”„๋กฌํ”„ํŠธ(_"Stevie Wonder์˜ ์–ด๋จธ๋‹ˆ๋Š”"_) ๊ฐ„์˜ _consistency score_๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. entity recall score๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•œ ๊ฐœ์ž…์ด ๋‘ ๋ฒˆ์งธ ํ™‰ ํ™œ์šฉ์˜ ์ง€ํ‘œ๋กœ์„œ ์ผ๊ด€์„ฑ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ฆ๊ฐ€์‹œํ‚ค๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๋‘ ๋‹จ๊ณ„๊ฐ€ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ผ์น˜ํ•˜๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ด ์ž‘์—…์—์„œ๋Š” ์ธ๊ฐ„์ด (1) _๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ_ (์˜ˆ: '๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜)๋ฅผ ์ถ”๋ก ํ•˜๊ณ  (2) ํ•ด๋‹น ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ (์˜ˆ: ๊ทธ์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ˆ„๊ตฌ์ธ์ง€)์„ ์ถ”๋ก ํ•˜์—ฌ ๋‘ ๋ฒˆ์˜ ํ™‰์œผ๋กœ ์™„์„ฑํ•  ์ˆ˜ ์žˆ๋Š” _๋‘ ๊ฐ€์ง€ ์‚ฌ์‹ค์˜ ๊ตฌ์„ฑ์„ ํ‘œํ˜„ํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์— ๊ตญํ•œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, LLM์ด ์˜ˆ์ธก๋œ ์ถœ๋ ฅ์„ ํฌ๊ฒŒ ๊ฒฐ์ •ํ•˜๋Š” ๊ฐ€์žฅ ๋ˆˆ์— ๋„๋Š” ๊ฒฝ๋กœ๋Š” ์•„๋‹ˆ๋”๋ผ๋„ ์œ ์‚ฌํ•œ ์ž ์žฌ์  2ํ™‰ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋นˆ๋„๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋จผ์ € ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ด๋Ÿฌํ•œ ํ™‰์„ ๊ฐœ๋ณ„์ ์œผ๋กœ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ํ™‰์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ์ˆจ๊ฒจ์ง„ ํ‘œํ˜„์„ ์–ดํœ˜ ๊ณต๊ฐ„์— ํˆฌ์˜ํ•˜์—ฌ LLM์˜ ๋‚ด๋ถ€์—์„œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ์„ ๊ทผ์‚ฌํ™”ํ•˜๋Š” _์—”ํ‹ฐํ‹ฐ ๊ฒ€์ƒ‰ ์ ์ˆ˜_๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๋ณ€๊ฒฝ์ด ์ด ์ ์ˆ˜์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ™‰์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด 2ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ์™„์„ฑ์„ ์œ„ํ•œ ์™„์„ฑ๊ณผ ๋™๋“ฑํ•œ ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ 1ํ™‰ ํ”„๋กฌํ”„ํŠธ(_"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š”"_) ๊ฐ„์˜ _์ผ๊ด€์„ฑ ์ ์ˆ˜_๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์—”ํ‹ฐํ‹ฐ ๊ฒ€์ƒ‰ ์ ์ˆ˜๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•œ ๊ฐœ์ž…์ด ๋‘ ๋ฒˆ์งธ ํ™‰ ํ™œ์šฉ์˜ ์ง€ํ‘œ๋กœ์„œ ์ผ๊ด€์„ฑ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ฆ๊ฐ€์‹œํ‚ค๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๋‘ ๋‹จ๊ณ„๊ฐ€ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ผ์น˜ํ•˜๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
In this section, we answer RQ2 of _how often an LLM performs the second-hop reasoning while processing two-hop prompts_. We view the second hop of reasoning as the LLM's utilization of what it knows about the bridge entity's attribute (Stevie Wonder's mother) to answer the two-hop prompt about the same attribute of the entity referred to by the descriptive mention (the singer of 'Superstition"s mother). Therefore, when an LLM performs the second hop, we expect to see a connection between its recall of the bridge entity (i.e. resolving the first hop) and its similarity in responding to a two-hop prompt and a corresponding one-hop prompt about the bridge entity's attribute, e.g., the two-hop prompt _"The mother of the singer of 'Superstition's"_ and the one-hop prompt _"The mother of Stevie Wonder is"_. Namely, the more strongly the model recalls the bridge entity (e.g., Stevie Wonder) while processing the two-hop prompt, the more similar the completion of this prompt should be to the completion of the one-hop prompt. In the following, we describe our approach for testing how often such a causal connection exists between entity recall and the _similarity_ in the prompt completions, which we refer to as _consistency_.
์ด ์„น์…˜์—์„œ๋Š” _๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ LLM์ด ๋‘ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€_์— ๋Œ€ํ•œ RQ2์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ถ”๋ก ์˜ ๋‘ ๋ฒˆ์งธ ํ™‰์„ LLM์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ(Stevie Wonder์˜ ์–ด๋จธ๋‹ˆ)์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ํ™œ์šฉํ•˜์—ฌ ์„ค๋ช…์  ์–ธ๊ธ‰์ด ๊ฐ€๋ฆฌํ‚ค๋Š” ์—”ํ‹ฐํ‹ฐ์˜ ๋™์ผํ•œ ์†์„ฑ์— ๋Œ€ํ•œ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ(์˜ˆ: 'Superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ)์— ๋‹ตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ๋‘ ๋ฒˆ์งธ ํ™‰์„ ์ˆ˜ํ–‰ํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ธฐ์–ตํ•˜๋Š” ๊ฒƒ(์ฆ‰, ์ฒซ ๋ฒˆ์งธ ํ™‰์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ)๊ณผ ๋‘ ๋ฒˆ์งธ ํ™‰ ํ”„๋กฌํ”„ํŠธ์— ์‘๋‹ตํ•˜๋Š” ๊ฒƒ๊ณผ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ์— ๋Œ€ํ•œ ํ•ด๋‹น ํ•œ ํ™‰ ํ”„๋กฌํ”„ํŠธ(์˜ˆ: ๋‘ ๋ฒˆ์งธ ํ™‰ ํ”„๋กฌํ”„ํŠธ _"Superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ"์™€ ํ•œ ํ™‰ ํ”„๋กฌํ”„ํŠธ _"Stevie Wonder์˜ ์–ด๋จธ๋‹ˆ๋Š”"_. ์ฆ‰, ๋ชจ๋ธ์ด ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ(์˜ˆ: Stevie Wonder)๋ฅผ ๋” ๊ฐ•ํ•˜๊ฒŒ ๊ธฐ์–ตํ• ์ˆ˜๋ก ์ด ํ”„๋กฌํ”„ํŠธ์˜ ์™„์„ฑ์€ ํ•œ ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ์™„์„ฑ๊ณผ ๋” ์œ ์‚ฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์—”ํ‹ฐํ‹ฐ ๋ฆฌ์ฝœ๊ณผ ํ”„๋กฌํ”„ํŠธ ์™„๋ฃŒ์˜ _similarity_ ์‚ฌ์ด์— ์ด๋Ÿฌํ•œ ์ธ๊ณผ ๊ด€๊ณ„๊ฐ€ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์กด์žฌํ•˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค(์—ฌ๊ธฐ์„œ _consistency_๋ผ๊ณ  ํ•จ).
์ด ์„น์…˜์—์„œ๋Š” _๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ LLM์ด ๋‘ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€_์— ๋Œ€ํ•œ RQ2์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ถ”๋ก ์˜ ๋‘ ๋ฒˆ์งธ ํ™‰์„ LLM์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ(์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ)์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ํ™œ์šฉํ•˜์—ฌ ์„ค๋ช…์  ์–ธ๊ธ‰์ด ๊ฐ€๋ฆฌํ‚ค๋Š” ์—”ํ‹ฐํ‹ฐ์˜ ๋™์ผํ•œ ์†์„ฑ์— ๋Œ€ํ•œ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ(์˜ˆ: 'Superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ)์— ๋‹ตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ๋‘ ๋ฒˆ์งธ ํ™‰์„ ์ˆ˜ํ–‰ํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ธฐ์–ตํ•˜๋Š” ๊ฒƒ(์ฆ‰, ์ฒซ ๋ฒˆ์งธ ํ™‰์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ)๊ณผ ๋‘ ๋ฒˆ์งธ ํ™‰ ํ”„๋กฌํ”„ํŠธ์— ์‘๋‹ตํ•˜๋Š” ๊ฒƒ๊ณผ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ์— ๋Œ€ํ•œ ํ•ด๋‹น ํ•œ ํ™‰ ํ”„๋กฌํ”„ํŠธ(์˜ˆ: ๋‘ ๋ฒˆ์งธ ํ™‰ ํ”„๋กฌํ”„ํŠธ _"Superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ"์™€ ํ•œ ํ™‰ ํ”„๋กฌํ”„ํŠธ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š”"_. ์ฆ‰, ๋ชจ๋ธ์ด ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ(์˜ˆ: ์Šคํ‹ฐ๋น„ ์›๋”)๋ฅผ ๋” ๊ฐ•ํ•˜๊ฒŒ ๊ธฐ์–ตํ• ์ˆ˜๋ก ์ด ํ”„๋กฌํ”„ํŠธ์˜ ์™„์„ฑ์€ ํ•œ ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ์™„์„ฑ๊ณผ ๋” ์œ ์‚ฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์—”ํ‹ฐํ‹ฐ ๋ฆฌ์ฝœ๊ณผ ํ”„๋กฌํ”„ํŠธ ์™„๋ฃŒ์˜ _์œ ์‚ฌ์„ฑ_ ์‚ฌ์ด์— ์ด๋Ÿฌํ•œ ์ธ๊ณผ ๊ด€๊ณ„๊ฐ€ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์กด์žฌํ•˜๋Š”์ง€ ํ…Œ์ŠคํŠธํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค(์—ฌ๊ธฐ์„œ _์ผ๊ด€์„ฑ_์ด๋ผ๊ณ  ํ•จ).
**Healthcare** is a vital application field closely related to human life. Ever since the advent of ChatGPT, a number of studies have applied ChatGPT or other LLMs to the medical domain. It has been shown that LLMs are capable of handling a variety of healthcare tasks, _e.g.,_ biology information extraction [763], medical advice consultation [902], mental health analysis [903], and report simplification [904]. As the major technical approach, researchers typically design specific prompts or instructions to guide LLMs to perform a wide range of medical tasks. To further harness the power of LLMs in the healthcare domain, researchers propose to develop healthcare-related LLMs [905, 356, 906]. Specifically, the Med-PaLM models [356, 905] achieves expert-level performance on the United States Medical Licensing Examination (USMLE), and earns greater approval from physicians in answering consumer's medical questions. However, LLMs may fabricate medical misinformation [904, 907], _e.g.,_ misinterpreting medical terms and suggesting advice inconsistent with medical guidelines. In addition, it would also raise privacy concerns to upload the health information of patients [763] into a commercial server that support the LLM.
**ํ—ฌ์Šค์ผ€์–ด**๋Š” ์ธ๊ฐ„์˜ ์ƒ๋ช…๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ๋Š” ์ค‘์š”ํ•œ ์‘์šฉ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ChatGPT๊ฐ€ ๋“ฑ์žฅํ•œ ์ด๋ž˜๋กœ ๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ChatGPT ๋˜๋Š” ๋‹ค๋ฅธ LLMs์„ ์˜๋ฃŒ ์˜์—ญ์— ์ ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLMs์€ ์ƒ๋ฌผํ•™ ์ •๋ณด ์ถ”์ถœ[763], ์˜๋ฃŒ ์ž๋ฌธ ์ƒ๋‹ด[902], ์ •์‹  ๊ฑด๊ฐ• ๋ถ„์„[903], ๋ณด๊ณ ์„œ ๊ฐ„์†Œํ™”[904] ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๊ธฐ์ˆ  ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์—ฐ๊ตฌ์ž๋“ค์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • prompts ๋˜๋Š” instructions์„ ์„ค๊ณ„ํ•˜์—ฌ LLMs์ด ๊ด‘๋ฒ”์œ„ํ•œ ์˜๋ฃŒ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ์˜๋ฃŒ ์˜์—ญ์—์„œ LLMs์˜ ํž˜์„ ๋”์šฑ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์˜๋ฃŒ ๊ด€๋ จ LLMs [905, 356, 906]์„ ๊ฐœ๋ฐœํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, Med-PaLM ๋ชจ๋ธ [356, 905]์€ ๋ฏธ๊ตญ ์˜์‚ฌ ๋ฉดํ—ˆ ์‹œํ—˜(USMLE)์—์„œ ์ „๋ฌธ๊ฐ€ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ์†Œ๋น„์ž์˜ ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์˜์‚ฌ์˜ ์Šน์ธ์„ ๋” ๋งŽ์ด ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLMs์€ ์˜๋ฃŒ ์ •๋ณด๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•˜๊ฑฐ๋‚˜ ์˜๋ฃŒ ์ง€์นจ๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ์กฐ์–ธ์„ ์ œ์•ˆํ•˜๋Š” ๋“ฑ ์˜๋ฃŒ ์˜ค๋ณด๋ฅผ ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[904, 907]. ๋˜ํ•œ ํ™˜์ž์˜ ๊ฑด๊ฐ• ์ •๋ณด [763]๋ฅผ LLMs์„ ์ง€์›ํ•˜๋Š” ์ƒ์šฉ ์„œ๋ฒ„์— ์—…๋กœ๋“œํ•˜๋Š” ๊ฒƒ๋„ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**ํ—ฌ์Šค์ผ€์–ด**๋Š” ์ธ๊ฐ„์˜ ์ƒ๋ช…๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ๋Š” ์ค‘์š”ํ•œ ์‘์šฉ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ChatGPT๊ฐ€ ๋“ฑ์žฅํ•œ ์ด๋ž˜๋กœ ๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ChatGPT ๋˜๋Š” ๋‹ค๋ฅธ LLM์„ ์˜๋ฃŒ ์˜์—ญ์— ์ ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์ƒ๋ฌผํ•™ ์ •๋ณด ์ถ”์ถœ[763], ์˜๋ฃŒ ์ž๋ฌธ ์ƒ๋‹ด[902], ์ •์‹  ๊ฑด๊ฐ• ๋ถ„์„[903], ๋ณด๊ณ ์„œ ๊ฐ„์†Œํ™”[904] ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๊ธฐ์ˆ  ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์—ฐ๊ตฌ์ž๋“ค์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ํ”„๋กฌํ”„ํŠธ ๋˜๋Š” ์ง€์นจ์„ ์„ค๊ณ„ํ•˜์—ฌ LLM์ด ๊ด‘๋ฒ”์œ„ํ•œ ์˜๋ฃŒ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ์˜๋ฃŒ ์˜์—ญ์—์„œ LLM์˜ ํž˜์„ ๋”์šฑ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์˜๋ฃŒ ๊ด€๋ จ LLM [905, 356, 906]์„ ๊ฐœ๋ฐœํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, Med-PaLM ๋ชจ๋ธ [356, 905]์€ ๋ฏธ๊ตญ ์˜์‚ฌ ๋ฉดํ—ˆ ์‹œํ—˜(USMLE)์—์„œ ์ „๋ฌธ๊ฐ€ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ์†Œ๋น„์ž์˜ ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์˜์‚ฌ์˜ ์Šน์ธ์„ ๋” ๋งŽ์ด ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLM์€ ์˜๋ฃŒ ์ •๋ณด๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•˜๊ฑฐ๋‚˜ ์˜๋ฃŒ ์ง€์นจ๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ์กฐ์–ธ์„ ์ œ์•ˆํ•˜๋Š” ๋“ฑ ์˜๋ฃŒ ์˜ค๋ณด๋ฅผ ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[904, 907]. ๋˜ํ•œ ํ™˜์ž์˜ ๊ฑด๊ฐ• ์ •๋ณด [763]๋ฅผ LLM์„ ์ง€์›ํ•˜๋Š” ์ƒ์šฉ ์„œ๋ฒ„์— ์—…๋กœ๋“œํ•˜๋Š” ๊ฒƒ๋„ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Reasoning without Observation (ReWOO) [175] aims to decouple reasoning from direct observations. ReWOO operates by enabling LLMs to formulate comprehensive reasoning plans or meta-plans without immediate reliance on external data or tools. This approach allows the agent to create a structured framework for reasoning that can be executed once the necessary data or observations are available. In ReWOO, the LLM initially develops a plan (a series of steps) that outlines how to approach and solve a given problem. This meta-planning phase is crucial as it sets the stage for the agent to process information once it becomes available. The execution phase then involves integrating actual data or observations into the pre-specified plan, leading to coherent and contextually relevant responses. ReWOO offers significant advantages in terms of token efficiency and robustness to tool failure. It enables LLMs to handle tasks where immediate access to external data is not available, relying instead on a well-structured reasoning framework. This method is particularly advantageous in scenarios where data retrieval is costly, slow, or uncertain, allowing the LLM-based agent to maintain a high level of performance and reliability.
๊ด€์ฐฐ ์—†์ด ์ถ”๋ก ํ•˜๊ธฐ(ReWOO) [175]๋Š” ์ถ”๋ก ์„ ์ง์ ‘ ๊ด€์ฐฐ๊ณผ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ReWOO๋Š” LLM์ด ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ๋‚˜ ๋„๊ตฌ์— ์ฆ‰๊ฐ์ ์œผ๋กœ ์˜์กดํ•˜์ง€ ์•Š๊ณ  ํฌ๊ด„์ ์ธ ์ถ”๋ก  ๊ณ„ํš ๋˜๋Š” ๋ฉ”ํƒ€ ๊ณ„ํš์„ ์ˆ˜๋ฆฝํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ๋‚˜ ๊ด€์ฐฐ์ด ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐํ™”๋œ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ReWOO์—์„œ LLM์€ ๋จผ์ € ์ฃผ์–ด์ง„ ๋ฌธ์ œ์— ์ ‘๊ทผํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋Š” ๊ณ„ํš(์ผ๋ จ์˜ ๋‹จ๊ณ„)์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด meta-planning ๋‹จ๊ณ„๋Š” ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐœํŒ์„ ๋งˆ๋ จํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ execution ๋‹จ๊ณ„์—์„œ๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋˜๋Š” ๊ด€์ฐฐ์„ ์‚ฌ์ „ ์ง€์ •๋œ ๊ณ„ํš์— ํ†ตํ•ฉํ•˜์—ฌ ์ผ๊ด€์„ฑ ์žˆ๊ณ  ์ƒํ™ฉ์— ๋งž๋Š” ์‘๋‹ต์„ ์ด๋Œ์–ด๋ƒ…๋‹ˆ๋‹ค. ReWOO๋Š” ํ† ํฐ ํšจ์œจ์„ฑ๊ณผ ๋„๊ตฌ ์‹คํŒจ์— ๋Œ€ํ•œ robustness ์ธก๋ฉด์—์„œ ์ƒ๋‹นํ•œ ์ด์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ์— ์ฆ‰์‹œ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†๋Š” ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋Œ€์‹  ์ž˜ ๊ตฌ์กฐํ™”๋œ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋ฐ์ดํ„ฐ retrieval ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฑฐ๋‚˜ ๋Š๋ฆฌ๊ฑฐ๋‚˜ ๋ถˆํ™•์‹คํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ํŠนํžˆ ์œ ๋ฆฌํ•˜๋ฉฐ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๊ฐ€ ๋†’์€ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ๊ณผ ์•ˆ์ •์„ฑ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
๊ด€์ฐฐ ์—†์ด ์ถ”๋ก ํ•˜๊ธฐ(ReWOO) [175]๋Š” ์ถ”๋ก ์„ ์ง์ ‘ ๊ด€์ฐฐ๊ณผ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ReWOO๋Š” LLM์ด ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ๋‚˜ ๋„๊ตฌ์— ์ฆ‰๊ฐ์ ์œผ๋กœ ์˜์กดํ•˜์ง€ ์•Š๊ณ  ํฌ๊ด„์ ์ธ ์ถ”๋ก  ๊ณ„ํš ๋˜๋Š” ๋ฉ”ํƒ€ ๊ณ„ํš์„ ์ˆ˜๋ฆฝํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ๋‚˜ ๊ด€์ฐฐ์ด ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐํ™”๋œ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ReWOO์—์„œ LLM์€ ๋จผ์ € ์ฃผ์–ด์ง„ ๋ฌธ์ œ์— ์ ‘๊ทผํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋Š” ๊ณ„ํš(์ผ๋ จ์˜ ๋‹จ๊ณ„)์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”ํƒ€ ๊ณ„ํš ๋‹จ๊ณ„๋Š” ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐœํŒ์„ ๋งˆ๋ จํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์‹คํ–‰ ๋‹จ๊ณ„์—์„œ๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋˜๋Š” ๊ด€์ฐฐ์„ ์‚ฌ์ „ ์ง€์ •๋œ ๊ณ„ํš์— ํ†ตํ•ฉํ•˜์—ฌ ์ผ๊ด€์„ฑ ์žˆ๊ณ  ์ƒํ™ฉ์— ๋งž๋Š” ์‘๋‹ต์„ ์ด๋Œ์–ด๋ƒ…๋‹ˆ๋‹ค. ReWOO๋Š” ํ† ํฐ ํšจ์œจ์„ฑ๊ณผ ๋„๊ตฌ ์‹คํŒจ์— ๋Œ€ํ•œ ๊ฒฌ๊ณ ์„ฑ ์ธก๋ฉด์—์„œ ์ƒ๋‹นํ•œ ์ด์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ์— ์ฆ‰์‹œ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†๋Š” ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋Œ€์‹  ์ž˜ ๊ตฌ์กฐํ™”๋œ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋ฐ์ดํ„ฐ ๊ฒ€์ƒ‰ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฑฐ๋‚˜ ๋Š๋ฆฌ๊ฑฐ๋‚˜ ๋ถˆํ™•์‹คํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ํŠนํžˆ ์œ ๋ฆฌํ•˜๋ฉฐ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๊ฐ€ ๋†’์€ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ๊ณผ ์•ˆ์ •์„ฑ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
Pretraining Large Language Models (LLMs) on large corpora of textual data is now a standard paradigm. When using these LLMs for many downstream applications, it is common to additionally bake in new knowledge (e.g., time-critical news, or private domain knowledge) into the pretrained model either through RAG-based-prompting, or finetuning. However, the optimal methodology for the model to gain such new knowledge remains an open question. In this paper, we present Retrieval Augmented Fine Tuning (RAFT), a training recipe that improves the model's ability to answer questions in an "open-book" in-domain setting. In RAFT, given a question, and a set of retrieved documents, we train the model to ignore those documents that don't help in answering the question, which we call, distractor documents. RAFT accomplishes this by citing verbatim the right sequence from the relevant document that would help answer the question. This coupled with RAFT's chain-of-thought-style response helps improve the model's ability to reason. In domain specific RAG, RAFT consistently improves the model's performance across PubMed, HotpotQA, and Gorilla datasets, presenting a post-training recipe to improve pre-trained LLMs to in-domain RAG.
๋Œ€๊ทœ๋ชจ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ ์ฝ”ํผ๋ผ์— ๋Œ€ํ•ด Large Language Models (LLMs)์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ํ˜„์žฌ ํ‘œ์ค€ ํŒจ๋Ÿฌ๋‹ค์ž„์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ LLM์„ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์‚ฌ์šฉํ•  ๋•Œ, Retrieval-augmented generation (RAG) ๊ธฐ๋ฐ˜ ํ”„๋กฌํ”„ํŠธ ๋˜๋Š” Fine-tuning์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์— ์ƒˆ๋กœ์šด ์ง€์‹(์˜ˆ: ์‹œ๊ฐ„์— ๋ฏผ๊ฐํ•œ ๋‰ด์Šค ๋˜๋Š” ๋น„๊ณต๊ฐœ ๋„๋ฉ”์ธ ์ง€์‹)์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ์ตœ์ ์˜ ๋ฐฉ๋ฒ•๋ก ์€ ์—ฌ์ „ํžˆ ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” '์˜คํ”ˆ ๋ถ' ๋„๋ฉ”์ธ ์„ค์ •์—์„œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ํ›ˆ๋ จ ๋ ˆ์‹œํ”ผ์ธ Retrieval Augmented Fine Tuning (RAFT)์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ์งˆ๋ฌธ๊ณผ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ ์ง‘ํ•ฉ์ด ์ฃผ์–ด์ง€๋ฉด ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€ ์•Š๋Š” ๋ฌธ์„œ, ์ฆ‰ ์‚ฐ๋งŒ ๋ฌธ์„œ๋ฅผ ๋ฌด์‹œํ•˜๋„๋ก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๊ด€๋ จ ๋ฌธ์„œ์—์„œ ์˜ฌ๋ฐ”๋ฅธ ์ˆœ์„œ๋ฅผ ์ •ํ™•ํžˆ ์ธ์šฉํ•˜์—ฌ ์ด๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” RAFT์˜ ์—ฐ์‡„์  ์‚ฌ๊ณ  ์Šคํƒ€์ผ์˜ ์‘๋‹ต๊ณผ ๊ฒฐํ•ฉ๋˜์–ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ RAG์—์„œ RAFT๋Š” PubMed, HotpotQA, Gorilla ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ผ๊ด€๋˜๊ฒŒ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๋„๋ฉ”์ธ๋ณ„ RAG๋กœ ๊ฐœ์„ ํ•˜๋Š” ์‚ฌํ›„ ํ•™์Šต ๋ ˆ์‹œํ”ผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ ์ฝ”ํผ๋ผ์— ๋Œ€ํ•ด ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ํ˜„์žฌ ํ‘œ์ค€ ํŒจ๋Ÿฌ๋‹ค์ž„์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ LLM์„ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์‚ฌ์šฉํ•  ๋•Œ, RAG ๊ธฐ๋ฐ˜ ํ”„๋กฌํ”„ํŠธ ๋˜๋Š” ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์— ์ƒˆ๋กœ์šด ์ง€์‹(์˜ˆ: ์‹œ๊ฐ„์— ๋ฏผ๊ฐํ•œ ๋‰ด์Šค ๋˜๋Š” ๋น„๊ณต๊ฐœ ๋„๋ฉ”์ธ ์ง€์‹)์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ์ตœ์ ์˜ ๋ฐฉ๋ฒ•๋ก ์€ ์—ฌ์ „ํžˆ ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” '์˜คํ”ˆ ๋ถ' ๋„๋ฉ”์ธ ์„ค์ •์—์„œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ํ›ˆ๋ จ ๋ ˆ์‹œํ”ผ์ธ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ๋ฏธ์„ธ ์กฐ์ •(RAFT)์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ์งˆ๋ฌธ๊ณผ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ ์ง‘ํ•ฉ์ด ์ฃผ์–ด์ง€๋ฉด ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€ ์•Š๋Š” ๋ฌธ์„œ, ์ฆ‰ ์‚ฐ๋งŒ ๋ฌธ์„œ๋ฅผ ๋ฌด์‹œํ•˜๋„๋ก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๊ด€๋ จ ๋ฌธ์„œ์—์„œ ์˜ฌ๋ฐ”๋ฅธ ์ˆœ์„œ๋ฅผ ์ •ํ™•ํžˆ ์ธ์šฉํ•˜์—ฌ ์ด๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” RAFT์˜ ์—ฐ์‡„์  ์‚ฌ๊ณ  ์Šคํƒ€์ผ์˜ ์‘๋‹ต๊ณผ ๊ฒฐํ•ฉ๋˜์–ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ RAG์—์„œ RAFT๋Š” PubMed, HotpotQA, Gorilla ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ผ๊ด€๋˜๊ฒŒ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๋„๋ฉ”์ธ๋ณ„ RAG๋กœ ๊ฐœ์„ ํ•˜๋Š” ์‚ฌํ›„ ํ•™์Šต ๋ ˆ์‹œํ”ผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
**Knowledge Completion.** In knowledge completion tasks, LLMs might be (to some extent) considered as a knowledge base [576], which can be leveraged to complete or predict the missing parts of knowledge units (_e.g._, knowledge triples). Such tasks can probe and evaluate _how much_ and _what kind of_ knowledge LLMs have learned from the pre-training data. Existing knowledge completion tasks can be roughly divided into knowledge graph completion tasks (_e.g._, FB15k-237 [572] and WN18RR [574]) and fact completion tasks (_e.g._, WikiFact [571]), which aim to complete the triples from a knowledge graph and incomplete sentences about specific facts, respectively. Empirical studies have revealed that it is difficult for existing LLMs to accomplish knowledge completion tasks related to specific relation types [520]. As shown in the evaluation results on WikiFact, LLMs perform well on several frequent relations that occur in the pre-training data (_e.g._, currency and author), while not well on rare ones (_e.g._, discoverer_or_inventor and place_of_birth). Interestingly, under the same evaluation settings (_e.g._, in-context learning), InstructGPT (_i.e._, text-davinci-002) outperforms GPT-3 in all subsets of WikiFact.
**Knowledge Completion**. ์ง€์‹ ์™„์„ฑ ์ž‘์—…์—์„œ LLM์€ (์–ด๋Š ์ •๋„) ์ง€์‹ ๋ฒ ์ด์Šค [576]๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ง€์‹ ๋‹จ์œ„์˜ ๋ˆ„๋ฝ๋œ ๋ถ€๋ถ„(_์˜ˆ: ์ง€์‹ ํŠธ๋ฆฌํ”Œ)์„ ์™„์„ฑํ•˜๊ฑฐ๋‚˜ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์€ LLM์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์–‘์˜ ์ง€์‹์„ ํ•™์Šตํ–ˆ๋Š”์ง€, ์–ด๋–ค ์ข…๋ฅ˜์˜ ์ง€์‹์„ ํ•™์Šตํ–ˆ๋Š”์ง€๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ ์ง€์‹ ์™„์„ฑ ์ž‘์—…์€ ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ ํŠธ๋ฆฌํ”Œ์„ ์™„์„ฑํ•˜๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์™„์„ฑ ์ž‘์—…(_์˜ˆ: FB15k-237 [572] ๋ฐ WN18RR [574])๊ณผ ํŠน์ • ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๋ถˆ์™„์ „ํ•œ ๋ฌธ์žฅ์„ ์™„์„ฑํ•˜๋Š” ์‚ฌ์‹ค ์™„์„ฑ ์ž‘์—…(_์˜ˆ: WikiFact [571])์œผ๋กœ ๋Œ€๋ณ„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ๊ฐ ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ ํŠธ๋ฆฌํ”Œ์„ ์™„์„ฑํ•˜๊ณ  ํŠน์ • ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๋ถˆ์™„์ „ํ•œ ๋ฌธ์žฅ์„ ์™„์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ฆ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๊ธฐ์กด LLM์€ ํŠน์ • ๊ด€๊ณ„ ์œ ํ˜•๊ณผ ๊ด€๋ จ๋œ ์ง€์‹ ์™„์„ฑ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค[520]. WikiFact์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ฒฐ๊ณผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, LLM์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋ฐœ์ƒํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋นˆ๋ฒˆํ•œ ๊ด€๊ณ„(์˜ˆ: ํ†ตํ™” ๋ฐ ์ž‘์„ฑ์ž)์— ๋Œ€ํ•ด์„œ๋Š” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ํฌ๊ท€ํ•œ ๊ด€๊ณ„(์˜ˆ: ๋ฐœ๊ฒฌ์ž_๋˜๋Š”_๋ฐœ๋ช…์ž ๋ฐ ์ถœ์ƒ์ง€)์— ๋Œ€ํ•ด์„œ๋Š” ๊ทธ๋‹ค์ง€ ์šฐ์ˆ˜ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋™์ผํ•œ ํ‰๊ฐ€ ์„ค์ •(์˜ˆ: In-Context Learning)์—์„œ InstructGPT(_์ฆ‰, text-davinci-002)๋Š” WikiFact์˜ ๋ชจ๋“  ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ GPT-3๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
**์ง€์‹ ์™„์„ฑ**. ์ง€์‹ ์™„์„ฑ ์ž‘์—…์—์„œ LLM์€ (์–ด๋Š ์ •๋„) ์ง€์‹ ๋ฒ ์ด์Šค [576]๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ง€์‹ ๋‹จ์œ„์˜ ๋ˆ„๋ฝ๋œ ๋ถ€๋ถ„(_์˜ˆ: ์ง€์‹ ํŠธ๋ฆฌํ”Œ)์„ ์™„์„ฑํ•˜๊ฑฐ๋‚˜ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์€ LLM์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์–‘์˜ ์ง€์‹์„ ํ•™์Šตํ–ˆ๋Š”์ง€, ์–ด๋–ค ์ข…๋ฅ˜์˜ ์ง€์‹์„ ํ•™์Šตํ–ˆ๋Š”์ง€๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ ์ง€์‹ ์™„์„ฑ ์ž‘์—…์€ ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ ํŠธ๋ฆฌํ”Œ์„ ์™„์„ฑํ•˜๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์™„์„ฑ ์ž‘์—…(_์˜ˆ: FB15k-237 [572] ๋ฐ WN18RR [574])๊ณผ ํŠน์ • ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๋ถˆ์™„์ „ํ•œ ๋ฌธ์žฅ์„ ์™„์„ฑํ•˜๋Š” ์‚ฌ์‹ค ์™„์„ฑ ์ž‘์—…(_์˜ˆ: ์œ„ํ‚คํŒฉํŠธ [571])์œผ๋กœ ๋Œ€๋ณ„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ๊ฐ ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ ํŠธ๋ฆฌํ”Œ์„ ์™„์„ฑํ•˜๊ณ  ํŠน์ • ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๋ถˆ์™„์ „ํ•œ ๋ฌธ์žฅ์„ ์™„์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ฆ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๊ธฐ์กด LLM์€ ํŠน์ • ๊ด€๊ณ„ ์œ ํ˜•๊ณผ ๊ด€๋ จ๋œ ์ง€์‹ ์™„์„ฑ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค[520]. ์œ„ํ‚คํŒฉํŠธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ฒฐ๊ณผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, LLM์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋ฐœ์ƒํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋นˆ๋ฒˆํ•œ ๊ด€๊ณ„(์˜ˆ: ํ†ตํ™” ๋ฐ ์ž‘์„ฑ์ž)์— ๋Œ€ํ•ด์„œ๋Š” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ํฌ๊ท€ํ•œ ๊ด€๊ณ„(์˜ˆ: ๋ฐœ๊ฒฌ์ž_๋˜๋Š”_๋ฐœ๋ช…์ž ๋ฐ ์ถœ์ƒ์ง€)์— ๋Œ€ํ•ด์„œ๋Š” ๊ทธ๋‹ค์ง€ ์šฐ์ˆ˜ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋™์ผํ•œ ํ‰๊ฐ€ ์„ค์ •(์˜ˆ: ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต)์—์„œ InstructGPT(_์ฆ‰, ํ…์ŠคํŠธ-๋‹ค๋นˆ์น˜-002)๋Š” ์œ„ํ‚คํŒฉํŠธ์˜ ๋ชจ๋“  ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ GPT-3๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
* **Three datasets, no shift** : We consider an \(N=3\) setting, where \(\mathcal{D}_{0},\mathcal{D}_{1},\mathcal{D}_{2}\) are each district 100B token splits of SlimPajama. This setting is primarily used to evaluate the ability of our techniques to scale to many future updates and to assess the performance of our proposed infinite learning rate schedules. * **Domain incremental continual pre-training**: This setting considers consuming the tokens of SlimPajama sequentially ordered by domain. That is, we train on a sequence of \(N\) future datasets \(\{\mathcal{D}_{0},\mathcal{D}_{1},\ldots,\mathcal{D}_{N-1}\}\) each of is a distinct domain of SlimPajama 300B. We note that this is similar to DACPT (Ke et al., 2022), however, we consider much larger datasets for each domain. This setting is particularly challenging due to the distribution shift experience at the transition between each domain. While it is certainly interesting, we believe it is unnecessarily difficult compare to mixing the SlimPajama data before training on it. The poor results in this setting (Sec. A.1 of the appendix) suggest that general-purpose LLMs should be continually pre-trained on a mixture of domains if possible, not updated per domain.
* **์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์ „ํ™˜ ์—†์Œ**: N=3 ์„ค์ •์„ ๊ณ ๋ คํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\mathcal{D}_{0},\mathcal{D}_{1},\mathcal{D}_{2}\)๋Š” ๊ฐ๊ฐ SlimPajama์˜ 100B token ๋ถ„ํ• ์ž…๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ์ฃผ๋กœ ๋งŽ์€ ๋ฏธ๋ž˜ ์—…๋ฐ์ดํŠธ๋กœ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ˆ ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ œ์•ˆ๋œ ๋ฌดํ•œ Learning Rate ์ผ์ •์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. * **๋„๋ฉ”์ธ ์ฆ๋ถ„ ์—ฐ์† ์‚ฌ์ „ ํ•™์Šต**: ์ด ์„ค์ •์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„๋กœ ์ˆœ์ฐจ์ ์œผ๋กœ ์ •๋ ฌ๋œ SlimPajama์˜ token์„ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๊ฐ๊ฐ SlimPajama 300B์˜ ๊ณ ์œ ํ•œ ๋„๋ฉ”์ธ์ธ \(N\) ๋ฏธ๋ž˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\{\mathcal{D}_{0},\mathcal{D}_{1},\ldots,\mathcal{D}_{N-1}\}\)์˜ ์‹œํ€€์Šค์— ๋Œ€ํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” DACPT(Ke et al., 2022)์™€ ์œ ์‚ฌํ•˜์ง€๋งŒ, ๊ฐ ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด ํ›จ์”ฌ ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ณ ๋ คํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ๊ฐ ๋„๋ฉ”์ธ ๊ฐ„์˜ ์ „ํ™˜ ์‹œ ๋ถ„ํฌ ์ด๋™ ๊ฒฝํ—˜์œผ๋กœ ์ธํ•ด ํŠนํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ๋ถ„๋ช… ํฅ๋ฏธ๋กญ์ง€๋งŒ, SlimPajama ๋ฐ์ดํ„ฐ๋ฅผ ํ›ˆ๋ จํ•˜๊ธฐ ์ „์— ํ˜ผํ•ฉํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๋ถˆํ•„์š”ํ•˜๊ฒŒ ์–ด๋ ต๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์˜ ์—ด์•…ํ•œ ๊ฒฐ๊ณผ(๋ถ€๋ก์˜ A.1์ ˆ)๋Š” ๋ฒ”์šฉ LLM์ด ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ ๋„๋ฉ”์ธ๋ณ„๋กœ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์„ ํ˜ผํ•ฉํ•˜์—ฌ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
* **์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์ „ํ™˜ ์—†์Œ**: N=3$ ์„ค์ •์„ ๊ณ ๋ คํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\mathcal{D}_{0},\mathcal{D}_{1},\mathcal{D}_{2}\)๋Š” ๊ฐ๊ฐ SlimPajama์˜ 100B ํ† ํฐ ๋ถ„ํ• ์ž…๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ์ฃผ๋กœ ๋งŽ์€ ๋ฏธ๋ž˜ ์—…๋ฐ์ดํŠธ๋กœ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ˆ ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ œ์•ˆ๋œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์ผ์ •์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. * **๋„๋ฉ”์ธ ์ฆ๋ถ„ ์—ฐ์† ์‚ฌ์ „ ํ•™์Šต**: ์ด ์„ค์ •์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„๋กœ ์ˆœ์ฐจ์ ์œผ๋กœ ์ •๋ ฌ๋œ SlimPajama์˜ ํ† ํฐ์„ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๊ฐ๊ฐ SlimPajama 300B์˜ ๊ณ ์œ ํ•œ ๋„๋ฉ”์ธ์ธ \(N\) ๋ฏธ๋ž˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\{\mathcal{D}_{0},\mathcal{D}_{1},\ldots,\mathcal{D}_{N-1}\}\)์˜ ์‹œํ€€์Šค์— ๋Œ€ํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” DACPT(Ke et al., 2022)์™€ ์œ ์‚ฌํ•˜์ง€๋งŒ, ๊ฐ ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด ํ›จ์”ฌ ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ณ ๋ คํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ๊ฐ ๋„๋ฉ”์ธ ๊ฐ„์˜ ์ „ํ™˜ ์‹œ ๋ถ„ํฌ ์ด๋™ ๊ฒฝํ—˜์œผ๋กœ ์ธํ•ด ํŠนํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ๋ถ„๋ช… ํฅ๋ฏธ๋กญ์ง€๋งŒ, SlimPajama ๋ฐ์ดํ„ฐ๋ฅผ ํ›ˆ๋ จํ•˜๊ธฐ ์ „์— ํ˜ผํ•ฉํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๋ถˆํ•„์š”ํ•˜๊ฒŒ ์–ด๋ ต๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์˜ ์—ด์•…ํ•œ ๊ฒฐ๊ณผ(๋ถ€๋ก์˜ A.1์ ˆ)๋Š” ๋ฒ”์šฉ LLM์ด ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ ๋„๋ฉ”์ธ๋ณ„๋กœ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์„ ํ˜ผํ•ฉํ•˜์—ฌ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Determining \(B\)As long as all hyperparameters are well-tuned (especially the learning rate and regularization hyperparameters) and the number of training steps is sufficient, it is believed that the same final performance should be attainable using any batch size [19], so the batch size mainly influences the training speed of language models. Often, when training large language models, the ideal batch size is suggested to be set as the largest batch size supported by the available hardware [1], so as to maximize the training speed without considering the computational cost. In Eq 3.12, we show that the critical batch size with the optimal speed/computation trade-off can be analytically computed from the loss value. Under the guidance of this formula, we would be able to estimate the preferred batch size under any loss trajectory. Furthermore, this optimal batch size in Eq 3.12 is determined by equally minimizing the training time and required computation, as shown in Eq 3.9. In practice, if we would like to prioritize one over the other, we can follow the same process to derive the optimal batch size. By this means, we are able to obtain the optimal batch size based on our customized need in a systematic way.
๊ฒฐ์ • \(B\)๋ชจ๋“  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ(ํŠนํžˆ Learning Rate ๋ฐ ์ •๊ทœํ™” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ)๊ฐ€ ์ž˜ ์กฐ์ •๋˜๊ณ  ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ถฉ๋ถ„ํ•œ ํ•œ, ๋ชจ๋“  Batch Size๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ตœ์ข… ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์–ด์ง€๋ฏ€๋กœ Batch Size๋Š” ์ฃผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ์†๋„์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ๋•Œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ•˜๋“œ์›จ์–ด์—์„œ ์ง€์›ํ•˜๋Š” ์ตœ๋Œ€ Batch Size๋กœ ์ด์ƒ์ ์ธ Batch Size๋ฅผ ์„ค์ •ํ•˜์—ฌ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ํ›ˆ๋ จ ์†๋„๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[1]. 3.12 ๋ฐฉ์ •์‹์—์„œ๋Š” Loss ๊ฐ’์—์„œ ์ตœ์ ์˜ ์†๋„/๊ณ„์‚ฐ ๋น„์šฉ ์ ˆ๊ฐ์ด ๊ฐ€๋Šฅํ•œ ์ž„๊ณ„ Batch Size๋ฅผ ๋ถ„์„์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๊ณต์‹์„ ํ†ตํ•ด ๋ชจ๋“  Loss ๊ถค์ ์—์„œ ์„ ํ˜ธํ•˜๋Š” Batch Size๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 3.12 ๋ฐฉ์ •์‹์˜ ์ตœ์  Batch Size๋Š” 3.9 ๋ฐฉ์ •์‹๊ณผ ๊ฐ™์ด ํ›ˆ๋ จ ์‹œ๊ฐ„๊ณผ ํ•„์š”ํ•œ ๊ณ„์‚ฐ์„ ๋™๋“ฑํ•˜๊ฒŒ ์ตœ์†Œํ™”ํ•˜์—ฌ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ•œ์ชฝ์„ ๋‹ค๋ฅธ ์ชฝ๋ณด๋‹ค ์šฐ์„ ์ˆœ์œ„์— ๋”ฐ๋ผ ์ตœ์ ์˜ Batch Size๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋งž์ถคํ˜• ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ฒด๊ณ„์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ตœ์ ์˜ Batch Size๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฒฐ์ • \(B\)๋ชจ๋“  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ(ํŠนํžˆ ํ•™์Šต ์†๋„ ๋ฐ ์ •๊ทœํ™” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ)๊ฐ€ ์ž˜ ์กฐ์ •๋˜๊ณ  ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ถฉ๋ถ„ํ•œ ํ•œ, ๋ชจ๋“  ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ตœ์ข… ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์–ด์ง€๋ฏ€๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” ์ฃผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ์†๋„์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ๋•Œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ•˜๋“œ์›จ์–ด์—์„œ ์ง€์›ํ•˜๋Š” ์ตœ๋Œ€ ๋ฐฐ์น˜ ํฌ๊ธฐ๋กœ ์ด์ƒ์ ์ธ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์„ค์ •ํ•˜์—ฌ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ํ›ˆ๋ จ ์†๋„๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[1]. 3.12 ๋ฐฉ์ •์‹์—์„œ๋Š” ์†์‹ค ๊ฐ’์—์„œ ์ตœ์ ์˜ ์†๋„/๊ณ„์‚ฐ ๋น„์šฉ ์ ˆ๊ฐ์ด ๊ฐ€๋Šฅํ•œ ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋ถ„์„์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๊ณต์‹์„ ํ†ตํ•ด ๋ชจ๋“  ์†์‹ค ๊ถค์ ์—์„œ ์„ ํ˜ธํ•˜๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 3.12 ๋ฐฉ์ •์‹์˜ ์ตœ์  ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 3.9 ๋ฐฉ์ •์‹๊ณผ ๊ฐ™์ด ํ›ˆ๋ จ ์‹œ๊ฐ„๊ณผ ํ•„์š”ํ•œ ๊ณ„์‚ฐ์„ ๋™๋“ฑํ•˜๊ฒŒ ์ตœ์†Œํ™”ํ•˜์—ฌ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ•œ์ชฝ์„ ๋‹ค๋ฅธ ์ชฝ๋ณด๋‹ค ์šฐ์„ ์ˆœ์œ„์— ๋”ฐ๋ผ ์ตœ์ ์˜ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋งž์ถคํ˜• ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ฒด๊ณ„์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ตœ์ ์˜ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The results of this experiment are presented in Table 5. We observe comparable performance between the replicated approach of Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22] and Adapt-LLM when trained on NQ and SQuAD datasets and tested on the 25% subset of PopQA. It's worth mentioning that Adapt-LLM does not utilize any information from PopQA, unlike Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22], who directly use the popularity score and a 75% portion of PopQA dataset to find an optimal value for that popularity score. This methodology is not generalizable to other open-domain question answering tasks since the popularity score is a unique feature of PopQA. However, Adapt-LLM can be applied to any similar dataset. Given these characteristics, we believe that the results obtained by Adapt-LLM are even more significant, offering comparable performance to an approach that utilizes dataset-specific information. These findings substantiate the validity of our approach, demonstrating its effectiveness even when trained on datasets different from the one used for testing.
์ด ์‹คํ—˜์˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 5์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. NQ ๋ฐ SQuAD ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•™์Šตํ•˜๊ณ  PopQA์˜ 25% ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ํ…Œ์ŠคํŠธํ–ˆ์„ ๋•Œ ๋ณต์ œ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์ธ Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์™€ Adapt-LLM ๊ฐ„์˜ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•๋ก ์€ ์ธ๊ธฐ๋„ ์ ์ˆ˜๋ฅผ ์ง์ ‘ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ๋Œ€ํ•œ ์ตœ์ ์˜ ๊ฐ’์„ ์ฐพ๋Š” Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์™€ ๋‹ฌ๋ฆฌ PopQA์˜ ์–ด๋–ค ์ •๋ณด๋„ ํ™œ์šฉํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•๋ก ์€ ์ธ๊ธฐ๋„ ์ ์ˆ˜๊ฐ€ PopQA์˜ ๊ณ ์œ ํ•œ ํŠน์ง•์ด๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ์˜คํ”ˆ ๋„๋ฉ”์ธ question answering ์ž‘์—…์—๋Š” ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Adapt-LLM์€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŠน์„ฑ์„ ๊ณ ๋ คํ•  ๋•Œ, Adapt-LLM์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„ ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์•„ Adapt-LLM์ด ์–ป์€ ๊ฒฐ๊ณผ๋Š” ๋”์šฑ ์˜๋ฏธ๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ์„ ์ž…์ฆํ•˜๋ฉฐ, ํ…Œ์ŠคํŠธ์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•™์Šตํ•œ ๊ฒฝ์šฐ์—๋„ ๊ทธ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
์ด ์‹คํ—˜์˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 5์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. NQ ๋ฐ SQuAD ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•™์Šตํ•˜๊ณ  PopQA์˜ 25% ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ํ…Œ์ŠคํŠธํ–ˆ์„ ๋•Œ ๋ณต์ œ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์ธ Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์™€ Adapt-LLM ๊ฐ„์˜ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•๋ก ์€ ์ธ๊ธฐ๋„ ์ ์ˆ˜๋ฅผ ์ง์ ‘ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ๋Œ€ํ•œ ์ตœ์ ์˜ ๊ฐ’์„ ์ฐพ๋Š” Mallen, Alex Troy ๋ฐ Asai, Akari ๋ฐ Zhong, Victor ๋ฐ Das, Rajarshi ๋ฐ Khashabi, Daniel ๋ฐ Hajishirzi, Hannaneh [22]์™€ ๋‹ฌ๋ฆฌ PopQA์˜ ์–ด๋–ค ์ •๋ณด๋„ ํ™œ์šฉํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•๋ก ์€ ์ธ๊ธฐ๋„ ์ ์ˆ˜๊ฐ€ PopQA์˜ ๊ณ ์œ ํ•œ ํŠน์ง•์ด๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ์˜คํ”ˆ ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์—๋Š” ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Adapt-LLM์€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŠน์„ฑ์„ ๊ณ ๋ คํ•  ๋•Œ, Adapt-LLM์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„ ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์•„ Adapt-LLM์ด ์–ป์€ ๊ฒฐ๊ณผ๋Š” ๋”์šฑ ์˜๋ฏธ๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ์„ ์ž…์ฆํ•˜๋ฉฐ, ํ…Œ์ŠคํŠธ์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•™์Šตํ•œ ๊ฒฝ์šฐ์—๋„ ๊ทธ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
RAG systems have a unique appeal over traditional search engines in that they can incorporate prior knowledge to fill in the gaps and extrapolate the retrieved information. We find that this comes with trade-offs - namely, that such priors can override information provided in documents. When perturbing RAG documents over a wide interval of values, the point at which models revert to their prior responses, or "tipping points", are latent and heterogeneous across different models and domains. While strong priors are not inherently problematic (and can often serve to safeguard models), the lack of explicit expectations around how models will mix reference documents with their priors can lead to downstream issues. For example, if RAG systems are used to extract nested financial data to be used in an algorithm, what will happen if there is a typo in the financial documents? Will the model notice the error and if so, what data will it provide in its place? Given that LLMs are soon to be widely deployed in many domains including medicine and law, users and developers alike should be cognizant of their unintended effects, especially if users have preconceptions that RAG-enabled systems are, by nature, always truthful.
RAG ์‹œ์Šคํ…œ์€ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ถ€์กฑํ•œ ๋ถ€๋ถ„์„ ์ฑ„์šฐ๊ณ  ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๊ธฐ์กด ๊ฒ€์ƒ‰ ์—”์ง„์— ๋น„ํ•ด ๋…ํŠนํ•œ ๋งค๋ ฅ์„ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ด์ ์—๋Š” trade-offs๊ฐ€ ์žˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ priors๊ฐ€ ๋ฌธ์„œ์— ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฎ์–ด์“ธ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๊ฐ’ ๊ฐ„๊ฒฉ์— ๊ฑธ์ณ RAG ๋ฌธ์„œ๋ฅผ ๊ต๋ž€ํ•  ๋•Œ, ๋ชจ๋ธ์ด ์ด์ „ ์‘๋‹ต์œผ๋กœ ๋˜๋Œ์•„๊ฐ€๋Š” ์‹œ์ , ์ฆ‰ 'tipping points'๋Š” ์ž ์žฌ์ ์ด๋ฉฐ ๋ชจ๋ธ๊ณผ ๋„๋ฉ”์ธ์— ๋”ฐ๋ผ ์ด์งˆ์ ์ž…๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ priors๊ฐ€ ๋ณธ์งˆ์ ์œผ๋กœ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ(์ข…์ข… ๋ชจ๋ธ์„ ๋ณดํ˜ธํ•˜๋Š” ์—ญํ• ์„ ํ•  ์ˆ˜๋„ ์žˆ์Œ), ๋ชจ๋ธ์ด ์ฐธ์กฐ ๋ฌธ์„œ์™€ priors๋ฅผ ํ˜ผํ•ฉํ•˜๋Š” ๋ฐฉ์‹์— ๋Œ€ํ•œ ๋ช…์‹œ์ ์ธ ๊ธฐ๋Œ€์น˜๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด downstream์—์„œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์‚ฌ์šฉํ•  ์ค‘์ฒฉ๋œ ์žฌ๋ฌด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๋ฐ RAG ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ ์žฌ๋ฌด ๋ฌธ์„œ์— ์˜คํƒ€๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๋ชจ๋ธ์ด ์˜ค๋ฅ˜๋ฅผ ๊ฐ์ง€ํ•˜๊ณ  ๊ฐ์ง€ํ•œ๋‹ค๋ฉด ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์‹  ์ œ๊ณตํ• ๊นŒ์š”? LLM์ด ์˜๋ฃŒ ๋ฐ ๋ฒ•๋ฅ ์„ ํฌํ•จํ•œ ๋งŽ์€ ์˜์—ญ์— ๊ณง ๋„๋ฆฌ ๋ฐฐํฌ๋  ์˜ˆ์ •์ด๋ฏ€๋กœ, ์‚ฌ์šฉ์ž์™€ ๊ฐœ๋ฐœ์ž ๋ชจ๋‘ ๋ชจ๋ธ์˜ ์˜๋„ํ•˜์ง€ ์•Š์€ ์˜ํ–ฅ, ํŠนํžˆ RAG ์ง€์› ์‹œ์Šคํ…œ์ด ๋ณธ์งˆ์ ์œผ๋กœ ํ•ญ์ƒ ์ง„์‹คํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ์‚ฌ์šฉ์ž์˜ ์˜คํ•ด์— ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
RAG ์‹œ์Šคํ…œ์€ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ถ€์กฑํ•œ ๋ถ€๋ถ„์„ ์ฑ„์šฐ๊ณ  ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๊ธฐ์กด ๊ฒ€์ƒ‰ ์—”์ง„์— ๋น„ํ•ด ๋…ํŠนํ•œ ๋งค๋ ฅ์„ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ด์ ์—๋Š” ์žฅ๋‹จ์ ์ด ์žˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ์„ ํ—˜์ด ๋ฌธ์„œ์— ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฎ์–ด์“ธ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๊ฐ’ ๊ฐ„๊ฒฉ์— ๊ฑธ์ณ RAG ๋ฌธ์„œ๋ฅผ ๊ต๋ž€ํ•  ๋•Œ, ๋ชจ๋ธ์ด ์ด์ „ ์‘๋‹ต์œผ๋กœ ๋˜๋Œ์•„๊ฐ€๋Š” ์‹œ์ , ์ฆ‰ '์„ ํ—˜์  ์ง€์ '์€ ์ž ์žฌ์ ์ด๋ฉฐ ๋ชจ๋ธ๊ณผ ๋„๋ฉ”์ธ์— ๋”ฐ๋ผ ์ด์งˆ์ ์ž…๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์„ ํ—˜์ด ๋ณธ์งˆ์ ์œผ๋กœ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ(์ข…์ข… ๋ชจ๋ธ์„ ๋ณดํ˜ธํ•˜๋Š” ์—ญํ• ์„ ํ•  ์ˆ˜๋„ ์žˆ์Œ), ๋ชจ๋ธ์ด ์ฐธ์กฐ ๋ฌธ์„œ์™€ ์„ ํ—˜์„ ํ˜ผํ•ฉํ•˜๋Š” ๋ฐฉ์‹์— ๋Œ€ํ•œ ๋ช…์‹œ์ ์ธ ๊ธฐ๋Œ€์น˜๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด ๋‹ค์šด์ŠคํŠธ๋ฆผ์—์„œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์‚ฌ์šฉํ•  ์ค‘์ฒฉ๋œ ์žฌ๋ฌด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๋ฐ RAG ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ ์žฌ๋ฌด ๋ฌธ์„œ์— ์˜คํƒ€๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๋ชจ๋ธ์ด ์˜ค๋ฅ˜๋ฅผ ๊ฐ์ง€ํ•˜๊ณ  ๊ฐ์ง€ํ•œ๋‹ค๋ฉด ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์‹  ์ œ๊ณตํ• ๊นŒ์š”? LLM์ด ์˜๋ฃŒ ๋ฐ ๋ฒ•๋ฅ ์„ ํฌํ•จํ•œ ๋งŽ์€ ์˜์—ญ์— ๊ณง ๋„๋ฆฌ ๋ฐฐํฌ๋  ์˜ˆ์ •์ด๋ฏ€๋กœ, ์‚ฌ์šฉ์ž์™€ ๊ฐœ๋ฐœ์ž ๋ชจ๋‘ ๋ชจ๋ธ์˜ ์˜๋„ํ•˜์ง€ ์•Š์€ ์˜ํ–ฅ, ํŠนํžˆ RAG ์ง€์› ์‹œ์Šคํ…œ์ด ๋ณธ์งˆ์ ์œผ๋กœ ํ•ญ์ƒ ์ง„์‹คํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ์‚ฌ์šฉ์ž์˜ ์˜คํ•ด์— ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Training on code data._ A major limitation of the original GPT-3 model (pre-trained on plain text) lies in the lack of the reasoning ability on complex tasks, _e.g._, completing the code and solving math problems. To enhance this ability, Codex [105] was introduced by OpenAI in July 2021, which was a GPT model fine-tuned on a large corpus of GitHub code. It demonstrated that Codex can solve very difficult programming problems, and also lead to a significant performance improvement in solving math problems [126]. Further, a contrastive approach [127] to training text and code embedding was reported in January 2022, which was shown to improve a series of related tasks (_i.e._, linear-probe classification, text search and code search). Actually, the GPT-3.5 models are developed based on a code-based GPT model (_i.e._, code-davinci-002), which indicates that training on code data is a very useful practice to improve the model capacity of GPT models, especially the reasoning ability. Furthermore, there is also a speculation that training on code data can greatly increase the chain-of-thought prompting abilities of LLMs [47], while it is still worth further investigation with more thorough verification.
\(\bullet\)_์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ._ ์›๋ž˜ GPT-3 ๋ชจ๋ธ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋กœ ์‚ฌ์ „ ํ•™์Šต)์˜ ์ฃผ์š” ํ•œ๊ณ„๋Š” ๋ณต์žกํ•œ ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด ์ฝ”๋“œ ์™„์„ฑ ๋ฐ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด 2021๋…„ 7์›”์— OpenAI์—์„œ GitHub ์ฝ”๋“œ์˜ ๋Œ€๊ทœ๋ชจ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด Fine-tuningํ•œ GPT ๋ชจ๋ธ์ธ Codex [105]๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. Codex๊ฐ€ ๋งค์šฐ ์–ด๋ ค์šด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์—์„œ๋„ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ด๋Œ์–ด๋ƒˆ์Šต๋‹ˆ๋‹ค[126]. ๋˜ํ•œ 2022๋…„ 1์›”์—๋Š” ํ…์ŠคํŠธ์™€ ์ฝ”๋“œ ์ž„๋ฒ ๋”ฉ์„ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์กฐ์  ์ ‘๊ทผ ๋ฐฉ์‹[127]์ด ๋ณด๊ณ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ผ๋ จ์˜ ๊ด€๋ จ ์ž‘์—…(์ฆ‰, ์„ ํ˜• ํ”„๋กœ๋ธŒ ๋ถ„๋ฅ˜, ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ๋ฐ ์ฝ”๋“œ ๊ฒ€์ƒ‰)์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ GPT-3.5 ๋ชจ๋ธ์€ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ GPT ๋ชจ๋ธ(_์ฆ‰, code-davinci-002)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์ด ํŠนํžˆ ์ถ”๋ก  ๋Šฅ๋ ฅ๊ณผ ๊ฐ™์€ GPT ๋ชจ๋ธ์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋งค์šฐ ์œ ์šฉํ•œ ์‹ค์Šต์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ LLM์˜ chain-of-thought prompting ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ถ”์ธก๋„ ์žˆ์ง€๋งŒ, ์ด๋Š” ์•„์ง ๋” ์ฒ ์ €ํ•œ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ._ ์›๋ž˜ GPT-3 ๋ชจ๋ธ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋กœ ์‚ฌ์ „ ํ•™์Šต)์˜ ์ฃผ์š” ํ•œ๊ณ„๋Š” ๋ณต์žกํ•œ ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด ์ฝ”๋“œ ์™„์„ฑ ๋ฐ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด 2021๋…„ 7์›”์— OpenAI์—์„œ GitHub ์ฝ”๋“œ์˜ ๋Œ€๊ทœ๋ชจ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•œ GPT ๋ชจ๋ธ์ธ Codex [105]๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. Codex๊ฐ€ ๋งค์šฐ ์–ด๋ ค์šด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์—์„œ๋„ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ด๋Œ์–ด๋ƒˆ์Šต๋‹ˆ๋‹ค[126]. ๋˜ํ•œ 2022๋…„ 1์›”์—๋Š” ํ…์ŠคํŠธ์™€ ์ฝ”๋“œ ์ž„๋ฒ ๋”ฉ์„ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์กฐ์  ์ ‘๊ทผ ๋ฐฉ์‹[127]์ด ๋ณด๊ณ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ผ๋ จ์˜ ๊ด€๋ จ ์ž‘์—…(์ฆ‰, ์„ ํ˜• ํ”„๋กœ๋ธŒ ๋ถ„๋ฅ˜, ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ๋ฐ ์ฝ”๋“œ ๊ฒ€์ƒ‰)์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ GPT-3.5 ๋ชจ๋ธ์€ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ GPT ๋ชจ๋ธ(_์ฆ‰, code-davinci-002)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์ด ํŠนํžˆ ์ถ”๋ก  ๋Šฅ๋ ฅ๊ณผ ๊ฐ™์€ GPT ๋ชจ๋ธ์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋งค์šฐ ์œ ์šฉํ•œ ์‹ค์Šต์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ LLM์˜ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ํ”„๋กฌํ”„ํŠธ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ถ”์ธก๋„ ์žˆ์ง€๋งŒ, ์ด๋Š” ์•„์ง ๋” ์ฒ ์ €ํ•œ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
Table 3 shows that by integrating our method with other methods, the performance can be further improved across different LLMs and tasks, despite these methods have different implementations. To be specific, in arithmetic reasoning tasks, our method enhances these methods to further improvement, yielding increases between \(10\%\) and \(21\%\) on the GSM8K dataset, and between \(1\%\) and \(15\%\) on the MATH dataset. In general reasoning tasks, integration with other methods generally achieves performance gains ranging from \(1\%\) to \(13\%\) in the Chess task and from \(1\%\) to \(11\%\) in the MMLU task. In code generation task, when combined with other methods, gains range from \(2\%\) to \(7\%\). However, two notable exceptions are observed when integrated with the debate method with the Llama2-13B and Llama2-70B models, which result in failed cases. This failure in performance is attributed primarily to the noise generated by referencing the answers of other agents during the debate process. The synthesized responses, which incorporate input from multiple agents, disrupt the coherence of the code logic, leading to the observed performance degradation. All accuracy curves are provided in the Appendix B.
ํ‘œ 3์€ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ฉํ•˜๋ฉด ๊ตฌํ˜„ ๋ฐฉ์‹์ด ๋‹ค๋ฅด์ง€๋งŒ ์—ฌ๋Ÿฌ LLM๊ณผ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ arithmetic reasoning ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์€ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ ํ–ฅ์ƒ์‹œ์ผœ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 10%์—์„œ 21%, MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 1%์—์„œ 15% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜ reasoning ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ์˜ ํ†ตํ•ฉ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ฒด์Šค ์ž‘์—…์—์„œ 1%์—์„œ 13%, MMLU ์ž‘์—…์—์„œ 1%์—์„œ 11% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. code generation ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ•˜๋ฉด 2%์—์„œ 7%๊นŒ์ง€ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๊ฐ€์ง€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์˜ˆ์™ธ๋Š” debate ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ–ˆ์„ ๋•Œ Llama2-13B ๋ฐ Llama2-70B ๋ชจ๋ธ์—์„œ ์‹คํŒจํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ ์ €ํ•˜๋Š” ์ฃผ๋กœ ํ† ๋ก  ๊ณผ์ •์—์„œ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ๋…ธ์ด์ฆˆ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ์ž…๋ ฅ์„ ํ†ตํ•ฉํ•œ ํ•ฉ์„ฑ๋œ ์‘๋‹ต์€ ์ฝ”๋“œ ๋กœ์ง์˜ ์ผ๊ด€์„ฑ์„ ๋ฐฉํ•ดํ•˜์—ฌ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋ชจ๋“  ์ •ํ™•๋„ ๊ณก์„ ์€ ๋ถ€๋ก B์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
ํ‘œ 3์€ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ฉํ•˜๋ฉด ๊ตฌํ˜„ ๋ฐฉ์‹์ด ๋‹ค๋ฅด์ง€๋งŒ ์—ฌ๋Ÿฌ LLM๊ณผ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ์‚ฐ์ˆ  ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์€ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ ํ–ฅ์ƒ์‹œ์ผœ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 10%์—์„œ 21%, MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 1%์—์„œ 15% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜ ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ์˜ ํ†ตํ•ฉ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ฒด์Šค ์ž‘์—…์—์„œ 1%์—์„œ 13%, MMLU ์ž‘์—…์—์„œ 1%์—์„œ 11% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ•˜๋ฉด 2%์—์„œ 7%๊นŒ์ง€ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๊ฐ€์ง€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์˜ˆ์™ธ๋Š” ํ† ๋ก  ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ–ˆ์„ ๋•Œ Llama2-13B ๋ฐ Llama2-70B ๋ชจ๋ธ์—์„œ ์‹คํŒจํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ ์ €ํ•˜๋Š” ์ฃผ๋กœ ํ† ๋ก  ๊ณผ์ •์—์„œ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ๋…ธ์ด์ฆˆ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ์ž…๋ ฅ์„ ํ†ตํ•ฉํ•œ ํ•ฉ์„ฑ๋œ ์‘๋‹ต์€ ์ฝ”๋“œ ๋กœ์ง์˜ ์ผ๊ด€์„ฑ์„ ๋ฐฉํ•ดํ•˜์—ฌ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋ชจ๋“  ์ •ํ™•๋„ ๊ณก์„ ์€ ๋ถ€๋ก B์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
RQ5: How does the style of synthetic data impact performance on specialized domains?We compare the performance of various models trained on different styles of synthetic data. In particular, we generate four styles of synthetic data (easy, medium, hard, and Q/A) and evaluate the performance of training on combinations of each style across File subsets. The prompts to generate these synthetic data styles are outlined in Appendix G. Results corresponding to generations from a Vicuna-v1.3 model, and for a 128M model trained for 3B tokens are summarized in Figure 7. We see that training with combinations of real C4 and synthetic data matching the style of the domain at evaluation improves performance. However, we find that no single synthetic data style performs the best across all domains, resulting in similar performance across training with combinations of real C4 data and each synthetic style variant. While knowing the best synthetic style to pre-train an LLM is impractical, an oracle that selects the best synthetic style across all domains will improve perplexity by 16%--indicating the importance of training with diverse data styles for LLM generalization, even when the underlying knowledge stays the same.
RQ5: ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์Šคํƒ€์ผ์ด ํŠน์ˆ˜ ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋‚˜์š”? ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋„ค ๊ฐ€์ง€ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(easy, medium, hard, and Q/A)๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ ์Šคํƒ€์ผ์˜ ์กฐํ•ฉ์— ๋Œ€ํ•œ ํ•™์Šต ์„ฑ๋Šฅ์„ ํŒŒ์ผ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ prompts๋Š” ๋ถ€๋ก G์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Vicuna-v1.3 ๋ชจ๋ธ์— ํ•ด๋‹นํ•˜๋Š” ๊ฒฐ๊ณผ์™€ 3B tokens์— ๋Œ€ํ•ด 128M ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 7์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€ ์‹œ ๋„๋ฉ”์ธ ์Šคํƒ€์ผ์— ๋งž๋Š” ์‹ค์ œ C4์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๋‹จ์ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์‹ค์ œ C4 ๋ฐ์ดํ„ฐ์™€ ๊ฐ ํ•ฉ์„ฑ ์Šคํƒ€์ผ ๋ณ€ํ˜•์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์•„๋Š” ๊ฒƒ์€ ๋น„ํ˜„์‹ค์ ์ด์ง€๋งŒ, ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์„ ํƒํ•˜๋Š” ์˜ค๋ผํด์€ perplexity๋ฅผ 16% ๊ฐœ์„ ํ•˜์—ฌ ๊ธฐ๋ณธ ์ง€์‹์€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋”๋ผ๋„ LLM ์ผ๋ฐ˜ํ™”๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
RQ5: ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์Šคํƒ€์ผ์ด ํŠน์ˆ˜ ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋‚˜์š”?๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋„ค ๊ฐ€์ง€ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(์‰ฌ์›€, ์ค‘๊ฐ„, ์–ด๋ ค์›€, Q/A)๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ ์Šคํƒ€์ผ์˜ ์กฐํ•ฉ์— ๋Œ€ํ•œ ํ•™์Šต ์„ฑ๋Šฅ์„ ํŒŒ์ผ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ถ€๋ก G์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Vicuna-v1.3 ๋ชจ๋ธ์— ํ•ด๋‹นํ•˜๋Š” ๊ฒฐ๊ณผ์™€ 3B ํ† ํฐ์— ๋Œ€ํ•ด 128M ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 7์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€ ์‹œ ๋„๋ฉ”์ธ ์Šคํƒ€์ผ์— ๋งž๋Š” ์‹ค์ œ C4์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๋‹จ์ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์‹ค์ œ C4 ๋ฐ์ดํ„ฐ์™€ ๊ฐ ํ•ฉ์„ฑ ์Šคํƒ€์ผ ๋ณ€ํ˜•์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์•„๋Š” ๊ฒƒ์€ ๋น„ํ˜„์‹ค์ ์ด์ง€๋งŒ, ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์„ ํƒํ•˜๋Š” ์˜ค๋ผํด์€ ๋‚œํ•ด์„ฑ์„ 16% ๊ฐœ์„ ํ•˜์—ฌ ๊ธฐ๋ณธ ์ง€์‹์€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋”๋ผ๋„ LLM ์ผ๋ฐ˜ํ™”๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**Tokenization.** Tokenization is also a crucial step for data preprocessing. It aims to segment raw text into sequences of individual tokens, which are subsequently used as the inputs of LLMs. In traditional NLP research (_e.g.,_ sequence labeling with conditional random fields [220]), word-based tokenization is the predominant approach, which is more aligned with human's language cognition. However, word-based tokenization can yield different segmentation results for the same input in some languages (_e.g.,_ Chinese word segmentation), generate a huge word vocabulary containing many low-frequency words, and also suffer from the _"out-of-vocabulary"_ issue. Thus, several neural network models employ _character_ as the minimum unit to derive the word representation (_e.g.,_ a CNN word encoder in ELMo [21]). Recently, _subword tokenizers_ have been widely used in Transformer based language models, typically including Byte-Pair Encoding tokenization, WordPiece tokenization and Unigram tokenization. HuggingFace has maintained an excellent online NLP course on tokenizer22 with running examples, and we refer to the beginners to this course. Next, we briefly describe the three representative tokenization methods.
ํ† ํฐํ™”.** Tokenization์€ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. Tokenization์€ ์›์‹œ ํ…์ŠคํŠธ๋ฅผ ๊ฐœ๋ณ„ ํ† ํฐ์˜ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ํ•˜์—ฌ LLM์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด NLP ์—ฐ๊ตฌ(์˜ˆ: ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹œํ€€์Šค ๋ผ๋ฒจ๋ง [220])์—์„œ๋Š” ๋‹จ์–ด ๊ธฐ๋ฐ˜ Tokenization์ด ์ฃผ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ด๋Š” ์ธ๊ฐ„์˜ ์–ธ์–ด ์ธ์ง€์™€ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์–ด ๊ธฐ๋ฐ˜ Tokenization์€ ์ผ๋ถ€ ์–ธ์–ด(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ถ„ํ• )์—์„œ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํ•  ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฐฉ๋Œ€ํ•œ ๋‹จ์–ด ์–ดํœ˜๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ณ , '์–ดํœ˜ ์™ธ ๋‹จ์–ด' ๋ฌธ์ œ๋„ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช‡๋ช‡ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์—์„œ๋Š” _๋ฌธ์ž_๋ฅผ ์ตœ์†Œ ๋‹จ์œ„๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ: ELMo [21]์˜ CNN ๋‹จ์–ด Encoder). ์ตœ๊ทผ์—๋Š” Transformer ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ์—์„œ _์„œ๋ธŒ์›Œ๋“œ Tokenization_๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ Byte-Pair Encoding Tokenization, WordPiece Tokenization, Unigram Tokenization์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. HuggingFace๋Š” Tokenizer22์— ๋Œ€ํ•œ ๋›ฐ์–ด๋‚œ ์˜จ๋ผ์ธ NLP ๊ฐ•์ขŒ๋ฅผ ์šด์˜ ์ค‘์ด๋ฉฐ, ์ดˆ๋ณด์ž๋Š” ์ด ๊ฐ•์ขŒ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ Tokenization ๋ฐฉ๋ฒ• ์„ธ ๊ฐ€์ง€๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
ํ† ํฐํ™”.** ํ† ํฐํ™”๋Š” ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฐํ™”๋Š” ์›์‹œ ํ…์ŠคํŠธ๋ฅผ ๊ฐœ๋ณ„ ํ† ํฐ์˜ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ํ•˜์—ฌ LLM์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด NLP ์—ฐ๊ตฌ(์˜ˆ: ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹œํ€€์Šค ๋ผ๋ฒจ๋ง [220])์—์„œ๋Š” ๋‹จ์–ด ๊ธฐ๋ฐ˜ ํ† ํฐํ™”๊ฐ€ ์ฃผ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ด๋Š” ์ธ๊ฐ„์˜ ์–ธ์–ด ์ธ์ง€์™€ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์–ด ๊ธฐ๋ฐ˜ ํ† ํฐํ™”๋Š” ์ผ๋ถ€ ์–ธ์–ด(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ถ„ํ• )์—์„œ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํ•  ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฐฉ๋Œ€ํ•œ ๋‹จ์–ด ์–ดํœ˜๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ณ , '์–ดํœ˜ ์™ธ ๋‹จ์–ด' ๋ฌธ์ œ๋„ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช‡๋ช‡ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์—์„œ๋Š” _๋ฌธ์ž_๋ฅผ ์ตœ์†Œ ๋‹จ์œ„๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ: ELMo [21]์˜ CNN ๋‹จ์–ด ์ธ์ฝ”๋”). ์ตœ๊ทผ์—๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ์—์„œ _์„œ๋ธŒ์›Œ๋“œ ํ† ํฐํ™”_๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ”์ดํŠธ ์Œ ์ธ์ฝ”๋”ฉ ํ† ํฐํ™”, ์›Œ๋“œํ”ผ์Šค ํ† ํฐํ™”, ์œ ๋‹ˆ๊ทธ๋žจ ํ† ํฐํ™”๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. HuggingFace๋Š” ํ† ํฐ๋ผ์ด์ €22์— ๋Œ€ํ•œ ๋›ฐ์–ด๋‚œ ์˜จ๋ผ์ธ NLP ๊ฐ•์ขŒ๋ฅผ ์šด์˜ ์ค‘์ด๋ฉฐ, ์ดˆ๋ณด์ž๋Š” ์ด ๊ฐ•์ขŒ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ ํ† ํฐํ™” ๋ฐฉ๋ฒ• ์„ธ ๊ฐ€์ง€๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
However, to align the model with human preferences, RLHF (Ouyang et al., 2022) presents a more complex challenge. It involves the development of alignment criteria, the acquisition of human feedback, and final learning adjustments based on this feedback. LLaMA2 (Touvron et al., 2023) is currently one of the publicly available models utilizing RLHF, with other models such as WebGPT (Nakano et al., 2021), InstructGPT (Ouyang et al., 2022), and GPT-4 (OpenAI, 2023) being proprietary. Implementing RLHF is intricate, stemming not only from the need for human annotations but also due to technical challenges. These include overfitting in reward models and instabilities during the Reinforcement Learning training phase (Gao et al., 2022). Additionally, the tedious procedure of training multiple LMs including the model being aligned, the reward model, and the inference model at the same time substantially amplifies memory and computational demands, particularly for larger models. Rafailov et al. (2023) proposed direct preference optimization (DPO) to align models with human preferences instead of complex and unstable RLHF. Nevertheless, one still needs to collect human-labeled preference data in the target language.
๊ทธ๋Ÿฌ๋‚˜ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ๋” ๋ณต์žกํ•œ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ •๋ ฌ ๊ธฐ์ค€์˜ ๊ฐœ๋ฐœ, ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์˜ ํš๋“, ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ตœ์ข… ํ•™์Šต ์กฐ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ RLHF๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋Š” LLaMA2(Touvron et al., 2023)์ด๋ฉฐ, WebGPT(Nakano et al., 2021), InstructGPT(Ouyang et al., 2022), GPT-4(OpenAI, 2023)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋…์  ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. RLHF๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์ฃผ์„์ด ํ•„์š”ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์ ์ธ ๋ฌธ์ œ๋กœ ์ธํ•ด ๋ณต์žกํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ณผ์ ํ•ฉ๊ณผ Reinforcement Learning ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ๋ถˆ์•ˆ์ •์„ฑ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Gao et al., 2022). ๋˜ํ•œ ์ •๋ ฌ ๋Œ€์ƒ ๋ชจ๋ธ, ๋ณด์ƒ ๋ชจ๋ธ, ์ถ”๋ก  ๋ชจ๋ธ์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ LM์„ ๋™์‹œ์— ํ›ˆ๋ จํ•ด์•ผ ํ•˜๋Š” ๋ฒˆ๊ฑฐ๋กœ์šด ์ ˆ์ฐจ๋กœ ์ธํ•ด ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ์š”๊ตฌ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Rafailov et al. (2023)์€ ๋ณต์žกํ•˜๊ณ  ๋ถˆ์•ˆ์ •ํ•œ RLHF ๋Œ€์‹  ๋ชจ๋ธ์„ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด Direct Preference Optimization (DPO)๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋Œ€์ƒ ์–ธ์–ด์—์„œ ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ๋ถ™์ธ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ๋” ๋ณต์žกํ•œ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ •๋ ฌ ๊ธฐ์ค€์˜ ๊ฐœ๋ฐœ, ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์˜ ํš๋“, ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ตœ์ข… ํ•™์Šต ์กฐ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ RLHF๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋Š” LLaMA2(Touvron et al., 2023)์ด๋ฉฐ, WebGPT(Nakano et al., 2021), InstructGPT(Ouyang et al., 2022), GPT-4(OpenAI, 2023)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋…์  ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. RLHF๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์ฃผ์„์ด ํ•„์š”ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์ ์ธ ๋ฌธ์ œ๋กœ ์ธํ•ด ๋ณต์žกํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ณผ์ ํ•ฉ๊ณผ ๊ฐ•ํ™” ํ•™์Šต ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ๋ถˆ์•ˆ์ •์„ฑ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Gao et al., 2022). ๋˜ํ•œ ์ •๋ ฌ ๋Œ€์ƒ ๋ชจ๋ธ, ๋ณด์ƒ ๋ชจ๋ธ, ์ถ”๋ก  ๋ชจ๋ธ์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ LM์„ ๋™์‹œ์— ํ›ˆ๋ จํ•ด์•ผ ํ•˜๋Š” ๋ฒˆ๊ฑฐ๋กœ์šด ์ ˆ์ฐจ๋กœ ์ธํ•ด ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ์š”๊ตฌ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Rafailov et al. (2023)์€ ๋ณต์žกํ•˜๊ณ  ๋ถˆ์•ˆ์ •ํ•œ RLHF ๋Œ€์‹  ๋ชจ๋ธ์„ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ์ง์ ‘ ์„ ํ˜ธ๋„ ์ตœ์ ํ™”(DPO)๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋Œ€์ƒ ์–ธ์–ด์—์„œ ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ๋ถ™์ธ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Results.Figure 6 shows the performance of GPT3 model and its instruction-tuned counterparts on this newly written instruction set (w. inter-rater agreement \(\kappa=0.57\) on the 4-class categorical scale, see Appendix B for details). As anticipated, the vanilla GPT3 LM is largely unable to respond to instructions, and all instruction-tuned models demonstrate comparatively higher performance. Nonetheless, GPT3SELF-INST (i.e., GPT3 model finetuned with Self-INstruct) outperforms those counterparts trained on T0 or SuperNI data by a large margin, demonstrating the value of the generated data despite the noise. Compared with InstructGPT001, GPT3SELF-INST is quite close in performance--if we count acceptable response with minor imperfections (Rating-B) as valid, GPT3SELF-INST is only 5% behind InstructGPT001. Lastly, our evaluation confirms the impressive instruction-following ability of InstructGPT002 and InstructGPT003. Although there are many factors behind this success, we conjecture that future work can largely benefit from improving the quality of our generated data by using human annotators or training a reward model to select better generations, similar to the algorithm used by Ouyang et al. (2022).
๊ฒฐ๊ณผ. ๊ทธ๋ฆผ 6์€ ์ƒˆ๋กœ ์ž‘์„ฑ๋œ instruction ์ง‘ํ•ฉ์— ๋Œ€ํ•ด GPT3 ๋ชจ๋ธ๊ณผ ๊ทธ instruction-tuned ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(4๋“ฑ๊ธ‰ ๋ฒ”์ฃผํ˜• ์ฒ™๋„์—์„œ rater ๊ฐ„ ์ผ์น˜๋„ \(\kappa=0.57\), ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B ์ฐธ์กฐ). ์˜ˆ์ƒ๋Œ€๋กœ, ๋ฐ”๋‹๋ผ GPT3 LM์€ instruction์— ๋Œ€์‘ํ•˜์ง€ ๋ชปํ•˜๋ฉฐ, ๋ชจ๋“  instruction-tuned ๋ชจ๋ธ์€ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  GPT3SELF-INST(์ฆ‰, Self-INstruct๋กœ Fine-tuning๋œ GPT3 ๋ชจ๋ธ)๋Š” T0 ๋˜๋Š” SuperNI ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋Œ€์‘๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋…ธ์ด์ฆˆ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์น˜๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. InstructGPT001๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, GPT3SELF-INST๋Š” ์„ฑ๋Šฅ์ด ์ƒ๋‹นํžˆ ๊ทผ์ ‘ํ•ฉ๋‹ˆ๋‹ค. ์•ฝ๊ฐ„์˜ ๋ถˆ์™„์ „์„ฑ(๋“ฑ๊ธ‰-B)์„ ๊ฐ–์ถ˜ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์‘๋‹ต์„ ์œ ํšจํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•œ๋‹ค๋ฉด, GPT3SELF-INST๋Š” InstructGPT001๋ณด๋‹ค 5% ๋’ค์ฒ˜์งˆ ๋ฟ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด๋ฒˆ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด InstructGPT002์™€ InstructGPT003์˜ ์ธ์ƒ์ ์ธ instruction ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๊ณต์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์ธ์ด ์žˆ์ง€๋งŒ, ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๋ฅผ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ Ouyang ๋“ฑ(2022)์ด ์‚ฌ์šฉํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ๋‚˜์€ ์ƒ์„ฑ์„ ์„ ํƒํ•จ์œผ๋กœ์จ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
๊ฒฐ๊ณผ. ๊ทธ๋ฆผ 6์€ ์ƒˆ๋กœ ์ž‘์„ฑ๋œ ๋ช…๋ น์–ด ์ง‘ํ•ฉ์— ๋Œ€ํ•ด GPT3 ๋ชจ๋ธ๊ณผ ๊ทธ ๋ช…๋ น์–ด์— ๋งž๊ฒŒ ์กฐ์ •๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(4๋“ฑ๊ธ‰ ๋ฒ”์ฃผํ˜• ์ฒ™๋„์—์„œ rater ๊ฐ„ ์ผ์น˜๋„ \(\kappa=0.57\), ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B ์ฐธ์กฐ). ์˜ˆ์ƒ๋Œ€๋กœ, ๋ฐ”๋‹๋ผ GPT3 LM์€ ๋ช…๋ น์— ๋Œ€์‘ํ•˜์ง€ ๋ชปํ•˜๋ฉฐ, ๋ชจ๋“  ๋ช…๋ น ์กฐ์ • ๋ชจ๋ธ์€ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  GPT3SELF-INST(์ฆ‰, Self-INstruct๋กœ ๋ฏธ์„ธ ์กฐ์ •๋œ GPT3 ๋ชจ๋ธ)๋Š” T0 ๋˜๋Š” SuperNI ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋Œ€์‘๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋…ธ์ด์ฆˆ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์น˜๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. InstructGPT001๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, GPT3SELF-INST๋Š” ์„ฑ๋Šฅ์ด ์ƒ๋‹นํžˆ ๊ทผ์ ‘ํ•ฉ๋‹ˆ๋‹ค. ์•ฝ๊ฐ„์˜ ๋ถˆ์™„์ „์„ฑ(๋“ฑ๊ธ‰-B)์„ ๊ฐ–์ถ˜ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์‘๋‹ต์„ ์œ ํšจํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•œ๋‹ค๋ฉด, GPT3SELF-INST๋Š” InstructGPT001๋ณด๋‹ค 5% ๋’ค์ฒ˜์งˆ ๋ฟ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด๋ฒˆ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด InstructGPT002์™€ InstructGPT003์˜ ์ธ์ƒ์ ์ธ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๊ณต์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์ธ์ด ์žˆ์ง€๋งŒ, ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๋ฅผ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ Ouyang ๋“ฑ(2022)์ด ์‚ฌ์šฉํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ๋‚˜์€ ์ƒ์„ฑ์„ ์„ ํƒํ•จ์œผ๋กœ์จ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
Given the poor performance of German models on all German evaluation tasks except HellaSwag (the same as English models on average), we further investigated their understanding of German by conducting a short qualitative study of model generations. In section A.4 of the appendix, we select five German prompts that contain various peculiarities of the German language (see Tab. 8 of the appendix). We then generate a fixed token-length response for each of the models trained German Common Crawl. As a baseline, we also evaluate the model trained only on the Pile. Despite the poor quality of generations at small model scale, we find that there is an observable improvement in the generative quality of German-language outputs from the models trained on German Common Crawl when compared to the Pile baseline, which tends to be systematically off-topic. This suggests that while our German-trained models have learned about the language, the evaluation tasks are too difficult to pick it up at the 405M parameter scale. Another reason is that the German dataset is smaller than the English datasets considered, and contains only web-scraped data, as opposed to the more sophisticated English datasets used in this work.
ํ—ฌ๋ผ์Šค์›ฉ(์˜์–ด ๋ชจ๋ธ๊ณผ ํ‰๊ท ์ ์œผ๋กœ ๋™์ผ)์„ ์ œ์™ธํ•œ ๋ชจ๋“  ๋…์ผ ํ‰๊ฐ€ ๊ณผ์ œ์—์„œ ๋…์ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ์ ์„ ๊ฐ์•ˆํ•˜์—ฌ ๋ชจ๋ธ ์ƒ์„ฑ์— ๋Œ€ํ•œ ์งง์€ ์ •์„ฑ์  ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋…์ผ์–ด์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก์˜ ์„น์…˜ A.4์—์„œ๋Š” ๋…์ผ์–ด์˜ ๋‹ค์–‘ํ•œ ํŠน์ด์ ์„ ํฌํ•จํ•˜๋Š” 5๊ฐœ์˜ ๋…์ผ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก์˜ ํ‘œ 8 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋…์ผ์–ด Common Crawl๋กœ ํ•™์Šต๋œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ณ ์ • ํ† ํฐ ๊ธธ์ด์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์„ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” the Pile๋งŒ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ฒด๊ณ„์ ์œผ๋กœ ์ฃผ์ œ์—์„œ ๋ฒ—์–ด๋‚˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” the Pile ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋…์ผ์–ด Common Crawl๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋…์ผ์–ด ์ถœ๋ ฅ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๊ฐœ์„ ์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋…์ผ์–ด๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์–ธ์–ด์— ๋Œ€ํ•ด ํ•™์Šตํ–ˆ์ง€๋งŒ ํ‰๊ฐ€ ๊ณผ์ œ๊ฐ€ ๋„ˆ๋ฌด ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ์—์„œ๋Š” ์ด๋ฅผ ํ”ฝ์—…ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ด์œ ๋Š” ๋…์ผ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณ ๋ ค๋œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์ž‘๊ณ , ์ด ์ž‘์—…์— ์‚ฌ์šฉ๋œ ๋” ์ •๊ตํ•œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ฌ๋ฆฌ ์›น ์Šคํฌ๋ž˜ํ•‘๋œ ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
ํ—ฌ๋ผ์Šค์›ฉ(์˜์–ด ๋ชจ๋ธ๊ณผ ํ‰๊ท ์ ์œผ๋กœ ๋™์ผ)์„ ์ œ์™ธํ•œ ๋ชจ๋“  ๋…์ผ ํ‰๊ฐ€ ๊ณผ์ œ์—์„œ ๋…์ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ์ ์„ ๊ฐ์•ˆํ•˜์—ฌ ๋ชจ๋ธ ์ƒ์„ฑ์— ๋Œ€ํ•œ ์งง์€ ์ •์„ฑ์  ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋…์ผ์–ด์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก์˜ ์„น์…˜ A.4์—์„œ๋Š” ๋…์ผ์–ด์˜ ๋‹ค์–‘ํ•œ ํŠน์ด์ ์„ ํฌํ•จํ•˜๋Š” 5๊ฐœ์˜ ๋…์ผ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก์˜ ํ‘œ 8 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง์œผ๋กœ ํ•™์Šต๋œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ณ ์ • ํ† ํฐ ๊ธธ์ด์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์„ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋”๋ฏธ๋งŒ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ฒด๊ณ„์ ์œผ๋กœ ์ฃผ์ œ์—์„œ ๋ฒ—์–ด๋‚˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๋”๋ฏธ ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋…์ผ์–ด ์ถœ๋ ฅ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๊ฐœ์„ ์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋…์ผ์–ด๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์–ธ์–ด์— ๋Œ€ํ•ด ํ•™์Šตํ–ˆ์ง€๋งŒ ํ‰๊ฐ€ ๊ณผ์ œ๊ฐ€ ๋„ˆ๋ฌด ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ์—์„œ๋Š” ์ด๋ฅผ ํ”ฝ์—…ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ด์œ ๋Š” ๋…์ผ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณ ๋ ค๋œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์ž‘๊ณ , ์ด ์ž‘์—…์— ์‚ฌ์šฉ๋œ ๋” ์ •๊ตํ•œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ฌ๋ฆฌ ์›น ์Šคํฌ๋ž˜ํ•‘๋œ ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
We validate our _Evol-Instruct_ by fine-tuning open-source LLaMA [4] with our evolved instructions and evaluating its performance similar to existing SOTA works (e.g., Alpaca [31] and Vicuna [22]) on instruction finetuning. The instruction datasets we compare with are the data used by Alpaca (generated using self-instruct [32]) and the 70k ShareGPT (shared by real users) used by Vicuna. To prove that the instruction dataset from our method is superior to human-created instruction datasets, we select Alpaca's training data (generated from only 175 human-created seed instructions) as the initial dataset. We execute four epochs of evolution using OpenAI ChatGPT API5 and finally obtain 250k instructions. To ensure a fair comparison with Vicuna's 70k real user data, we sampled an equal amount from the full 250k data and trained the LLaMA 7B model. We name our model _WizardLM_. Due to the low proportion of difficult instructions in the previous instruction-following test dataset, we manually created a new difficulty-balanced test dataset, named _Evol-Instruct_ testset. We hire annotators and leverage GPT-4 to evaluate Alpaca, Vicuna, ChatGPT, and _WizardLM_ on _Evol-Instruct_ testset and Vicuna's testset.
์šฐ๋ฆฌ๋Š” ์ง„ํ™”ํ•œ instructions๋กœ ์˜คํ”ˆ ์†Œ์Šค LLaMA [4]๋ฅผ Fine-tuningํ•˜๊ณ  instruction Fine-tuning์— ๋Œ€ํ•œ ๊ธฐ์กด SOTA ์ž‘์—…(์˜ˆ: Alpaca [31] ๋ฐ Vicuna [22])๊ณผ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์—ฌ _Evol-Instruct_์˜ ์œ ํšจ์„ฑ์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ๋น„๊ตํ•˜๋Š” instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Alpaca์—์„œ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ(self-instruct [32]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋จ)์™€ Vicuna์—์„œ ์‚ฌ์šฉํ•œ 70k ShareGPT(์‹ค์ œ ์‚ฌ์šฉ์ž๊ฐ€ ๊ณต์œ ํ•œ ๋ฐ์ดํ„ฐ)์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์‚ฌ๋žŒ์ด ๋งŒ๋“  instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์šฐ์›”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด Alpaca์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ(175๊ฐœ์˜ ์‚ฌ๋žŒ์ด ๋งŒ๋“  ์‹œ๋“œ instructions๋กœ๋งŒ ์ƒ์„ฑ๋จ)๋ฅผ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. OpenAI ChatGPT API5๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 4ํšŒ์˜ ์ง„ํ™”๋ฅผ ์‹คํ–‰ํ•˜๊ณ  ์ตœ์ข…์ ์œผ๋กœ 25๋งŒ ๊ฐœ์˜ instructions๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. Vicuna์˜ 70,000๊ฐœ์˜ ์‹ค์ œ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ „์ฒด 25๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์—์„œ ๋™์ผํ•œ ์–‘์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ LLaMA 7B ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์ด๋ฆ„์€ _WizardLM_์ž…๋‹ˆ๋‹ค. ์ด์ „ instruction ์ˆ˜ํ–‰ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‚œ์ด๋„๊ฐ€ ๋‚ฎ์•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‚œ์ด๋„ ๊ท ํ˜•์ด ์žกํžŒ ์ƒˆ๋กœ์šด ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ผ๊ณ  ๋ช…๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ฃผ์„๊ฐ€๋ฅผ ๊ณ ์šฉํ•˜๊ณ  GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ Alpaca, Vicuna, ChatGPT ๋ฐ _WizardLM_์„ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” ์ง„ํ™”ํ•œ ๋ช…๋ น์–ด๋กœ ์˜คํ”ˆ ์†Œ์Šค LLaMA [4]๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ณ  ๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ •์— ๋Œ€ํ•œ ๊ธฐ์กด SOTA ์ž‘์—…(์˜ˆ: ์•ŒํŒŒ์นด [31] ๋ฐ ๋น„์ฟ ๋‚˜ [22])๊ณผ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์—ฌ _Evol-Instruct_์˜ ์œ ํšจ์„ฑ์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ๋น„๊ตํ•˜๋Š” ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์•ŒํŒŒ์นด์—์„œ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ(self-instruct [32]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋จ)์™€ ๋น„์ฟ ๋‚˜์—์„œ ์‚ฌ์šฉํ•œ 70k ShareGPT(์‹ค์ œ ์‚ฌ์šฉ์ž๊ฐ€ ๊ณต์œ ํ•œ ๋ฐ์ดํ„ฐ)์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์‚ฌ๋žŒ์ด ๋งŒ๋“  ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์šฐ์›”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ์•ŒํŒŒ์นด์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ(175๊ฐœ์˜ ์‚ฌ๋žŒ์ด ๋งŒ๋“  ์‹œ๋“œ ๋ช…๋ น์–ด๋กœ๋งŒ ์ƒ์„ฑ๋จ)๋ฅผ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. OpenAI ChatGPT API5๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 4ํšŒ์˜ ์ง„ํ™”๋ฅผ ์‹คํ–‰ํ•˜๊ณ  ์ตœ์ข…์ ์œผ๋กœ 25๋งŒ ๊ฐœ์˜ ๋ช…๋ น์–ด๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ๋น„์ฟ ๋‚˜์˜ 70,000๊ฐœ์˜ ์‹ค์ œ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ „์ฒด 25๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์—์„œ ๋™์ผํ•œ ์–‘์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ LLaMA 7B ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์ด๋ฆ„์€ _WizardLM_์ž…๋‹ˆ๋‹ค. ์ด์ „ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‚œ์ด๋„๊ฐ€ ๋‚ฎ์•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‚œ์ด๋„ ๊ท ํ˜•์ด ์žกํžŒ ์ƒˆ๋กœ์šด ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ผ๊ณ  ๋ช…๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ฃผ์„๊ฐ€๋ฅผ ๊ณ ์šฉํ•˜๊ณ  GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์•ŒํŒŒ์นด, ๋น„์ฟ ๋‚˜, ChatGPT ๋ฐ _WizardLM_์„ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ๋น„์ฟ ๋‚˜์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
We therefore introduce Contrastive Filtering, a method to select the instructions that can effectively enhance the target LLM \(f_{t}\). For clarity, we define the space of all natural language sequences as \(\mathcal{N}\). We have the strong LLM \(f_{s}:\mathcal{N}\rightarrow\mathcal{N}\), the target LLM \(f_{t}:\mathcal{N}\rightarrow\mathcal{N}\), and a scoring function \(S:\mathcal{N}\rightarrow\mathbb{R}\) to evaluate response quality. In practice, \(S\) is obtained by reusing the strong LLM \(f_{s}\) with a prompt template (Figure 11, Appendix A.9) adapted from the Vicuna pairwise evaluation template (Taori et al., 2023; Chiang et al., 2023). To mitigate potential position bias, we average the scores obtained by exchanging the positions of two responses (Chiang et al., 2023). We observe using \(f_{s}\) for scoring works quite well in practice, so we prioritize this option for simplicity. Given an input instruction \(I\in\mathcal{N}\), we obtain responses from both LLMs as \(f_{s}(I)\) and \(f_{t}(I)\), respectively. We then define the _quality gap_\(G:\mathcal{N}\rightarrow\mathbb{R}\) between these responses to estimate the _effectiveness_ of the instruction: \(G(I)=S(f_{s}(I))-S(f_{t}(I))\).
๋”ฐ๋ผ์„œ ๋Œ€์ƒ LLM(f_{t})์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” instructions๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ Contrastive Filtering์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด ๋ชจ๋“  ์ž์—ฐ์–ด ์‹œํ€€์Šค์˜ ๊ณต๊ฐ„์„ \(\mathcal{N}\)์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ LLM \(f_{s}:\mathcal{N}\rightarrow\mathcal{N}\), ๋Œ€์ƒ LLM \(f_{t}:\mathcal{N}\rightarrow\mathcal{N}\), ์‘๋‹ต ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ scoring function \(S:\mathcal{N}\rightarrow\mathbb{R}\)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ \(S\)๋Š” Vicuna ์Œ๋ณ„ ํ‰๊ฐ€ ํ…œํ”Œ๋ฆฟ(Taori et al., 2023; Chiang et al., 2023)์—์„œ ์ˆ˜์ •ํ•œ prompt ํ…œํ”Œ๋ฆฟ(๊ทธ๋ฆผ 11, ๋ถ€๋ก A.9)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)๋ฅผ ์žฌ์‚ฌ์šฉํ•˜์—ฌ ์–ป์Šต๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ์œ„์น˜ ํŽธํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ์‘๋‹ต์˜ ์œ„์น˜๋ฅผ ๊ตํ™˜ํ•˜์—ฌ ์–ป์€ ์ ์ˆ˜๋ฅผ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค(Chiang et al., 2023). ์ ์ˆ˜ ๋ถ€์—ฌ์— \(f_{s}\)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์‹ค์ œ๋กœ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ์ˆœํ™”๋ฅผ ์œ„ํ•ด ์ด ์˜ต์…˜์„ ์šฐ์„ ์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ instruction \(I\in\mathcal{N}\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๋‘ LLM ๋ชจ๋‘์—์„œ ๊ฐ๊ฐ \(f_{s}(I)\)์™€ \(f_{t}(I)\)๋กœ ์‘๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์‘๋‹ต ๊ฐ„์˜ _quality gap_ \(G:\mathcal{N}\rightarrow\mathbb{R}\)์„ ์ •์˜ํ•˜์—ฌ instruction์˜ _effectiveness_๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค: \(G(I)=S(f_{s}(I))-S(f_{t}(I))\).
๋”ฐ๋ผ์„œ ๋Œ€์ƒ LLM(f_{t})์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ช…๋ น์„ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ ๋Œ€์กฐ ํ•„ํ„ฐ๋ง์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด ๋ชจ๋“  ์ž์—ฐ์–ด ์‹œํ€€์Šค์˜ ๊ณต๊ฐ„์„ \(\mathcal{N}\)์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ LLM \(f_{s}:\mathcal{N}\rightarrow\mathcal{N}\), ๋Œ€์ƒ LLM \(f_{t}:\mathcal{N}\rightarrow\mathcal{N}\), ์‘๋‹ต ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ ์ˆ˜ ํ•จ์ˆ˜ \(S:\mathcal{N}\rightarrow\mathbb{R}\)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ \(S\)๋Š” Vicuna ์Œ๋ณ„ ํ‰๊ฐ€ ํ…œํ”Œ๋ฆฟ(Taori et al., 2023; Chiang et al., 2023)์—์„œ ์ˆ˜์ •ํ•œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ(๊ทธ๋ฆผ 11, ๋ถ€๋ก A.9)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)๋ฅผ ์žฌ์‚ฌ์šฉํ•˜์—ฌ ์–ป์Šต๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ์œ„์น˜ ํŽธํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ์‘๋‹ต์˜ ์œ„์น˜๋ฅผ ๊ตํ™˜ํ•˜์—ฌ ์–ป์€ ์ ์ˆ˜๋ฅผ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค(Chiang et al., 2023). ์ ์ˆ˜ ๋ถ€์—ฌ์— \(f_{s}\)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์‹ค์ œ๋กœ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ์ˆœํ™”๋ฅผ ์œ„ํ•ด ์ด ์˜ต์…˜์„ ์šฐ์„ ์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ ๋ช…๋ น \(I\in\mathcal{N}\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๋‘ LLM ๋ชจ๋‘์—์„œ ๊ฐ๊ฐ \(f_{s}(I)\)์™€ \(f_{t}(I)\)๋กœ ์‘๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์‘๋‹ต ๊ฐ„์˜ _ํ’ˆ์งˆ ๊ฐญ_\(G:\mathcal{N}\rightarrow\mathbb{R}\)์„ ์ •์˜ํ•˜์—ฌ ๋ช…๋ น์˜ _ํšจ๊ณผ์„ฑ_์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค: \(G(I)=S(f_{s}(I))-S(f_{t}(I))\).
**LLM as Encoder with Instruction Metadata.** We begin by encoding the given seed instructions \(\mathcal{D}_{s}=\{I_{i}\}_{i=1}^{n}\) into instruction _metadata_\(\mathcal{M}\), _i.e._, keywords that capture the underlying target instruction distribution. Inspired by the task pool by Wang et al. (2022) and the post-hoc analysis on skill distribution by Xu et al. (2023), we define the metadata as encompassing two key aspects: _use case_ and _skills_. Use case describes the intended task (_e.g._, question answering or creative writing), while Skills are the knowledge the LLM required to have to successfully respond to the given instruction (_e.g._, algorithms or communication). Skills are often generalizable to different use cases. Therefore, each instruction has a single use case and may involve multiple skills. To extract this metadata, we leverage the strong LLM \(f_{s}\) following the prompt template in Figure 7, Appendix A.9. While richer definitions are possible based on finer-grained instruction-following metrics Zhou et al. (2023), we prioritize use case and skills for their broad applicability across diverse instruction distributions. Future work can explore extending this metadata further.
**LLM as Encoder with Instruction Metadata.** ์ฃผ์–ด์ง„ ์‹œ๋“œ ๋ช…๋ น์–ด \(\mathcal{D}_{s}=\{I_{i}\}_{i=1}^{n}\)์„ instruction _metadata_\(\mathcal{M}\), ์ฆ‰, ๊ธฐ๋ณธ ํƒ€๊ฒŸ ๋ช…๋ น์–ด ๋ถ„ํฌ๋ฅผ ์บก์ฒ˜ํ•˜๋Š” ํ‚ค์›Œ๋“œ๋กœ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. Wang ๋“ฑ(2022)์˜ ์ž‘์—… ํ’€๊ณผ Xu ๋“ฑ(2023)์˜ ์‚ฌํ›„ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, metadata๋Š” _use case_์™€ _skills_๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์ธก๋ฉด์„ ํฌ๊ด„ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. Use case๋Š” ์˜๋„ํ•œ ์ž‘์—…(_์˜ˆ: question answering ๋˜๋Š” ์ฐฝ์˜์  ๊ธ€์“ฐ๊ธฐ)์„ ์„ค๋ช…ํ•˜๋ฉฐ, Skills๋Š” ์ฃผ์–ด์ง„ instruction์— ์„ฑ๊ณต์ ์œผ๋กœ ์‘๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLM์ด ๊ฐ–์ถฐ์•ผ ํ•˜๋Š” ์ง€์‹(_์˜ˆ: algorithms ๋˜๋Š” communication)์ž…๋‹ˆ๋‹ค. Skills๋Š” ์ข…์ข… ๋‹ค์–‘ํ•œ use case์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ instruction์—๋Š” ํ•˜๋‚˜์˜ use case๊ฐ€ ์žˆ์œผ๋ฉฐ ์—ฌ๋Ÿฌ skills์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด metadata๋ฅผ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 7์˜ prompt ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก A.9). ๋ณด๋‹ค ์„ธ๋ถ„ํ™”๋œ instruction ์ˆ˜ํ–‰ ๋ฉ”ํŠธ๋ฆญ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๋” ํ’๋ถ€ํ•œ ์ •์˜๊ฐ€ ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ instruction ๋ถ„ํฌ์— ๊ฑธ์ณ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ ์šฉ ๊ฐ€๋Šฅํ•œ use case์™€ skills์„ ์šฐ์„ ์ ์œผ๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด metadata๋ฅผ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ช…๋ น ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ๋Š” ์ธ์ฝ”๋”๋กœ์„œ์˜ LLM**. ์ฃผ์–ด์ง„ ์‹œ๋“œ ๋ช…๋ น์–ด \mathcal{D}_{s}=\{I_{i}\}_{i=1}^{n}\)์„ ๋ช…๋ น์–ด _๋ฉ”ํƒ€๋ฐ์ดํ„ฐ_\mathcal{M}, ์ฆ‰, ๊ธฐ๋ณธ ํƒ€๊ฒŸ ๋ช…๋ น์–ด ๋ถ„ํฌ๋ฅผ ์บก์ฒ˜ํ•˜๋Š” ํ‚ค์›Œ๋“œ๋กœ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. Wang ๋“ฑ(2022)์˜ ์ž‘์—… ํ’€๊ณผ Xu ๋“ฑ(2023)์˜ ์‚ฌํ›„ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ์‚ฌ์šฉ ์‚ฌ๋ก€_์™€ _์Šคํ‚ฌ_์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์ธก๋ฉด์„ ํฌ๊ด„ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ ์‚ฌ๋ก€๋Š” ์˜๋„ํ•œ ์ž‘์—…(_์˜ˆ: ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋˜๋Š” ์ฐฝ์˜์  ๊ธ€์“ฐ๊ธฐ)์„ ์„ค๋ช…ํ•˜๋ฉฐ, ์Šคํ‚ฌ์€ ์ฃผ์–ด์ง„ ๋ช…๋ น์— ์„ฑ๊ณต์ ์œผ๋กœ ์‘๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLM์ด ๊ฐ–์ถฐ์•ผ ํ•˜๋Š” ์ง€์‹(_์˜ˆ: ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋˜๋Š” ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜)์ž…๋‹ˆ๋‹ค. ์Šคํ‚ฌ์€ ์ข…์ข… ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ ๋ช…๋ น์—๋Š” ํ•˜๋‚˜์˜ ์‚ฌ์šฉ ์‚ฌ๋ก€๊ฐ€ ์žˆ์œผ๋ฉฐ ์—ฌ๋Ÿฌ ์Šคํ‚ฌ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 7์˜ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก A.9). ๋ณด๋‹ค ์„ธ๋ถ„ํ™”๋œ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋ฉ”ํŠธ๋ฆญ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๋” ํ’๋ถ€ํ•œ ์ •์˜๊ฐ€ ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋ช…๋ น์–ด ๋ถ„ํฌ์— ๊ฑธ์ณ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€์™€ ์Šคํ‚ฌ์„ ์šฐ์„ ์ ์œผ๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Faced with both opportunities and challenges, it needs more attention on the research and development of LLMs. Inorder to provide a basic understanding of LLMs, this survey conducts a literature review of the recent advances in LLMs from four major aspects, including _pre-training_ (how to pre-train a capable LLM), _adaptation_ (how to effectively adapt pre-trained LLMs for better use), _utilization_ (how to use LLMs for solving various downstream tasks) and _capability evaluation_ (how to evaluate the abilities of LLMs and existing empirical findings). We thoroughly comb the literature and summarize the key findings, techniques, and methods of LLMs. For this survey, we also create a GitHub project website by collecting the supporting resources for LLMs, at the link [https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey). We are also aware of several related review articles on PLMs or LLMs [32, 36, 38, 39, 43, 48, 49, 50, 51, 52, 53, 54]. These papers either discuss PLMs or some specific (or general) aspects of LLMs. Compared with them, we focus on the techniques and methods to develop and use LLMs and provide a relatively comprehensive reference to important aspects of LLMs.
๊ธฐํšŒ์™€ ๋„์ „์— ์ง๋ฉดํ•œ LLM์˜ ์—ฐ๊ตฌ ๊ฐœ๋ฐœ์— ๋” ๋งŽ์€ attention์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” LLM์— ๋Œ€ํ•œ ๊ธฐ๋ณธ์ ์ธ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด _pre-training_ (์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•), _adaptation_ (์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•), _utilization_ (๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•), _capability evaluation_ (LLM์˜ ๋Šฅ๋ ฅ๊ณผ ๊ธฐ์กด ๊ฒฝํ—˜์  ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•) ๋“ฑ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ์ธก๋ฉด์—์„œ ์ตœ๊ทผ LLM์˜ ๋ฐœ์ „์— ๋Œ€ํ•œ ๋ฌธํ—Œ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ฌธํ—Œ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•˜๊ณ  LLM์˜ ์ฃผ์š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๊ธฐ์ˆ  ๋ฐ ๋ฐฉ๋ฒ•์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์ง€์› ๋ฆฌ์†Œ์Šค๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ GitHub ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, ๋งํฌ [https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey)์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ PLM ๋˜๋Š” LLM์— ๊ด€ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ๋ฆฌ๋ทฐ ๊ธฐ์‚ฌ [32, 36, 38, 39, 43, 48, 49, 50, 51, 52, 53, 54]๋„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ผ๋ฌธ์€ PLM ๋˜๋Š” LLM์˜ ํŠน์ •(๋˜๋Š” ์ผ๋ฐ˜์ ์ธ) ์ธก๋ฉด์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋น„๊ตํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” LLM์„ ๊ฐœ๋ฐœํ•˜๊ณ  ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ˆ ๊ณผ ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  LLM์˜ ์ค‘์š”ํ•œ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋น„๊ต์  ํฌ๊ด„์ ์ธ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๊ธฐํšŒ์™€ ๋„์ „์— ์ง๋ฉดํ•œ LLM์˜ ์—ฐ๊ตฌ ๊ฐœ๋ฐœ์— ๋” ๋งŽ์€ ๊ด€์‹ฌ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” LLM์— ๋Œ€ํ•œ ๊ธฐ๋ณธ์ ์ธ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ ํ•™์Šต(_์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•), ์ ์‘(_์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•), ํ™œ์šฉ(_๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•), ๋Šฅ๋ ฅ ํ‰๊ฐ€(_LLM์˜ ๋Šฅ๋ ฅ๊ณผ ๊ธฐ์กด ๊ฒฝํ—˜์  ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•) ๋“ฑ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ์ธก๋ฉด์—์„œ ์ตœ๊ทผ LLM์˜ ๋ฐœ์ „์— ๋Œ€ํ•œ ๋ฌธํ—Œ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ฌธํ—Œ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•˜๊ณ  LLM์˜ ์ฃผ์š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๊ธฐ์ˆ  ๋ฐ ๋ฐฉ๋ฒ•์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์ง€์› ๋ฆฌ์†Œ์Šค๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ GitHub ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, ๋งํฌ [https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey)์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ PLM ๋˜๋Š” LLM์— ๊ด€ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ๋ฆฌ๋ทฐ ๊ธฐ์‚ฌ [32, 36, 38, 39, 43, 48, 49, 50, 51, 52, 53, 54]๋„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ผ๋ฌธ์€ PLM ๋˜๋Š” LLM์˜ ํŠน์ •(๋˜๋Š” ์ผ๋ฐ˜์ ์ธ) ์ธก๋ฉด์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋น„๊ตํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” LLM์„ ๊ฐœ๋ฐœํ•˜๊ณ  ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ˆ ๊ณผ ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  LLM์˜ ์ค‘์š”ํ•œ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋น„๊ต์  ํฌ๊ด„์ ์ธ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**Remaining Issues.** In this part, we further discuss several important issues to apply LLMs to improve IR systems. First, though LLMs are capable of being as general-purpose task solvers, they are not directly well suited for existing IR systems: they require high overhead for inference [774, 782], have limitations in modeling long texts or document lists [778], and need special adaptation (_e.g._, instruction tuning) to perform the text ranking task [795]. Therefore, more systematic approaches to adapt LLMs for modern IR systems should be investigated, to leverage their benefits and meanwhile overcome these limitations. Secondly, the advent of LLMs sheds lights on the development of new information seeking ways (_e.g._, New Bing). It is meaningful to explore how to reshape the architecture and paradigm of IR by integrating the LLMs' capacities and the merits of existing IR systems [796]. Thirdly, existing work mainly focuses on text retrieval tasks, lacking a comprehensive consideration of multimodal information sources. As will be discussed in Section 8.1.4, multimodal large language models [797] are also widely studied, making it feasible to develop more powerful multimedia retrieval systems.
**ํ˜„์•ˆ**. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์„ ์ ์šฉํ•˜์—ฌ IR ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, LLM์€ ๋ฒ”์šฉ ํƒœ์Šคํฌ ์†”๋ฒ„๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์—๋Š” ์ง์ ‘ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ถ”๋ก ์„ ์œ„ํ•ด ๋†’์€ ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ณ [774, 782], ๊ธด ํ…์ŠคํŠธ๋‚˜ ๋ฌธ์„œ ๋ชฉ๋ก์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ[778], ํ…์ŠคํŠธ ์ˆœ์œ„ ์ง€์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ์ ์‘(_์˜ˆ: Instruction Tuning)์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค[795]. ๋”ฐ๋ผ์„œ LLM์˜ ์žฅ์ ์„ ํ™œ์šฉํ•˜๋Š” ๋™์‹œ์— ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์‹  IR ์‹œ์Šคํ…œ์— LLM์„ ์ ์šฉํ•˜๋Š” ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์—ฐ๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM์˜ ๋“ฑ์žฅ์œผ๋กœ ์ƒˆ๋กœ์šด ์ •๋ณด ํƒ์ƒ‰ ๋ฐฉ์‹(์˜ˆ: New Bing)์˜ ๊ฐœ๋ฐœ์— ์กฐ๋ช…์ด ๋น„์ถฐ์กŒ์Šต๋‹ˆ๋‹ค. LLM์˜ ์—ญ๋Ÿ‰๊ณผ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์˜ ์žฅ์ ์„ ํ†ตํ•ฉํ•˜์—ฌ IR์˜ ์•„ํ‚คํ…์ฒ˜์™€ ํŒจ๋Ÿฌ๋‹ค์ž„์„ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฏธ ์žˆ๋Š” ์ผ์ž…๋‹ˆ๋‹ค[796]. ์…‹์งธ, ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ •๋ณด ์†Œ์Šค์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. 8.1.4์ ˆ์—์„œ ๋…ผ์˜ํ•  ๊ฒƒ์ฒ˜๋Ÿผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ Large Language Models[797]๋„ ๋„๋ฆฌ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์–ด ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
**ํ˜„์•ˆ**. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์„ ์ ์šฉํ•˜์—ฌ IR ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, LLM์€ ๋ฒ”์šฉ ํƒœ์Šคํฌ ์†”๋ฒ„๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์—๋Š” ์ง์ ‘ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ถ”๋ก ์„ ์œ„ํ•ด ๋†’์€ ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ณ [774, 782], ๊ธด ํ…์ŠคํŠธ๋‚˜ ๋ฌธ์„œ ๋ชฉ๋ก์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ[778], ํ…์ŠคํŠธ ์ˆœ์œ„ ์ง€์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ์ ์‘(_์˜ˆ: ์ง€์‹œ ์กฐ์ •)์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค[795]. ๋”ฐ๋ผ์„œ LLM์˜ ์žฅ์ ์„ ํ™œ์šฉํ•˜๋Š” ๋™์‹œ์— ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์‹  IR ์‹œ์Šคํ…œ์— LLM์„ ์ ์šฉํ•˜๋Š” ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์—ฐ๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM์˜ ๋“ฑ์žฅ์œผ๋กœ ์ƒˆ๋กœ์šด ์ •๋ณด ํƒ์ƒ‰ ๋ฐฉ์‹(์˜ˆ: New Bing)์˜ ๊ฐœ๋ฐœ์— ์กฐ๋ช…์ด ๋น„์ถฐ์กŒ์Šต๋‹ˆ๋‹ค. LLM์˜ ์—ญ๋Ÿ‰๊ณผ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์˜ ์žฅ์ ์„ ํ†ตํ•ฉํ•˜์—ฌ IR์˜ ์•„ํ‚คํ…์ฒ˜์™€ ํŒจ๋Ÿฌ๋‹ค์ž„์„ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฏธ ์žˆ๋Š” ์ผ์ž…๋‹ˆ๋‹ค[796]. ์…‹์งธ, ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ •๋ณด ์†Œ์Šค์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. 8.1.4์ ˆ์—์„œ ๋…ผ์˜ํ•  ๊ฒƒ์ฒ˜๋Ÿผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ[797]๋„ ๋„๋ฆฌ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์–ด ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
**Gain from Model Scale.**: We observe that Yi-34B has substantial performance improvement compared to Yi-6B, though they utilized the same pretrain corpora. Larger model size leads to higher performance gain on Code and Math benchmarks, referring to Tab. 3, compared to benchmarks focusing on Commonsense Reasoning, Reading Comprehension, or Knowledge. **Data Quality.**: Smaller models of higher quality pretrain data, like Yi-34B or Qwen-14B, usually demonstrate better performance than models of larger size but (presumably) lower quality data, such as Falcon-180B (though the focus of Falcon-180B might be more on the scaling side, which is definitely of important value on its own). **Gap between GPT-4 and Open-source LLMs.**: Based on Tab. 2, we note that open-source LLMs still lag behind the performance of GPT-4 and GPT-3.5 on various benchmarks. Yet representative bilingual LLMs, e.g. Qwen-14B and Yi-34B, can match or even surpass the performance of GPT-4 on Chinese knowledge related benchmarks, including C-Eval [31], CMMLU [42], and Gaokao [90]. However, there is still a huge gap between GPT-4 and open-source models on reasoning-related benchmarks like BBH [72], code (HumanEval), and math (MATH).
**๋ชจ๋ธ ๊ทœ๋ชจ๋กœ๋ถ€ํ„ฐ์˜ ์ด๋“**: Yi-34B๋Š” ๋™์ผํ•œ pretrain corpora๋ฅผ ํ™œ์šฉํ–ˆ์ง€๋งŒ Yi-6B์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ํด์ˆ˜๋ก Commonsense Reasoning, Reading Comprehension ๋˜๋Š” Knowledge์— ์ดˆ์ ์„ ๋งž์ถ˜ ๋ฒค์น˜๋งˆํฌ์— ๋น„ํ•ด Code ๋ฐ Math ๋ฒค์น˜๋งˆํฌ์—์„œ ๋” ๋†’์€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(ํ‘œ 3 ์ฐธ์กฐ). **๋ฐ์ดํ„ฐ ํ’ˆ์งˆ**: Yi-34B ๋˜๋Š” Qwen-14B์™€ ๊ฐ™์ด ๋” ๋†’์€ ํ’ˆ์งˆ์˜ pretrain ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์†Œํ˜• ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ Falcon-180B์™€ ๊ฐ™์ด ๋” ํฐ ํฌ๊ธฐ์ด์ง€๋งŒ (์•„๋งˆ๋„) ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Falcon-180B์˜ ์ดˆ์ ์€ scaling ์ธก๋ฉด์— ๋” ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Š” ๊ทธ ์ž์ฒด๋กœ ์ค‘์š”ํ•œ ๊ฐ€์น˜๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค). **GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค LLM ๊ฐ„์˜ ๊ฒฉ์ฐจ**: ํ‘œ 2๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์˜คํ”ˆ ์†Œ์Šค LLM์€ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4 ๋ฐ GPT-3.5์˜ ์„ฑ๋Šฅ์— ์—ฌ์ „ํžˆ ๋’ค์ฒ˜์ง€๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€ํ‘œ์ ์ธ ์ด์ค‘ ์–ธ์–ด LLM์ธ Qwen-14B์™€ Yi-34B๋Š” C-Eval [31], CMMLU [42], Gaokao [90] ๋“ฑ ์ค‘๊ตญ์–ด ์ง€์‹ ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4์˜ ์„ฑ๋Šฅ์— ํ•„์ ํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ reasoning ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: BBH [72], code (HumanEval), math (MATH))์—์„œ๋Š” GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ๊ฐ„์— ์—ฌ์ „ํžˆ ํฐ ๊ฒฉ์ฐจ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
**๋ชจ๋ธ ๊ทœ๋ชจ๋กœ๋ถ€ํ„ฐ์˜ ์ด๋“**: Yi-34B๋Š” ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ๋ผ๋ฅผ ํ™œ์šฉํ–ˆ์ง€๋งŒ Yi-6B์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ํด์ˆ˜๋ก ๊ณตํ†ต ๊ฐ๊ฐ ์ถ”๋ก , ๋…ํ•ด ๋˜๋Š” ์ง€์‹์— ์ดˆ์ ์„ ๋งž์ถ˜ ๋ฒค์น˜๋งˆํฌ์— ๋น„ํ•ด ์ฝ”๋“œ ๋ฐ ์ˆ˜ํ•™ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋” ๋†’์€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(ํ‘œ 3 ์ฐธ์กฐ). **๋ฐ์ดํ„ฐ ํ’ˆ์งˆ**: Yi-34B ๋˜๋Š” Qwen-14B์™€ ๊ฐ™์ด ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์†Œํ˜• ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ Falcon-180B์™€ ๊ฐ™์ด ๋” ํฐ ํฌ๊ธฐ์ด์ง€๋งŒ (์•„๋งˆ๋„) ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Falcon-180B์˜ ์ดˆ์ ์€ ์Šค์ผ€์ผ๋ง ์ธก๋ฉด์— ๋” ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Š” ๊ทธ ์ž์ฒด๋กœ ์ค‘์š”ํ•œ ๊ฐ€์น˜๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค). **GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค LLM ๊ฐ„์˜ ๊ฒฉ์ฐจ**: ํ‘œ 2๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์˜คํ”ˆ ์†Œ์Šค LLM์€ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4 ๋ฐ GPT-3.5์˜ ์„ฑ๋Šฅ์— ์—ฌ์ „ํžˆ ๋’ค์ฒ˜์ง€๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€ํ‘œ์ ์ธ ์ด์ค‘ ์–ธ์–ด LLM์ธ Qwen-14B์™€ Yi-34B๋Š” C-Eval [31], CMMLU [42], Gaokao [90] ๋“ฑ ์ค‘๊ตญ์–ด ์ง€์‹ ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4์˜ ์„ฑ๋Šฅ์— ํ•„์ ํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ถ”๋ก  ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: BBH [72], ์ฝ”๋“œ(HumanEval), ์ˆ˜ํ•™(MATH))์—์„œ๋Š” GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ๊ฐ„์— ์—ฌ์ „ํžˆ ํฐ ๊ฒฉ์ฐจ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.