Self-Discover: Large Language Models Self-Compose Reasoning Structures
Abstract
We introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework is a self-discovery process where LLMs select multiple atomic reasoning modules such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2's performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT). Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference compute. Finally, we show that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.
Community
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Efficient Tool Use with Chain-of-Abstraction Reasoning (2024)
- Distilling Mathematical Reasoning Capabilities into Small Language Models (2024)
- TPD: Enhancing Student Language Model Reasoning via Principle Discovery and Guidance (2024)
- Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding (2024)
- Divide and Conquer for Large Language Models Reasoning (2024)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Introducing my new Blog Series: "AI Research Chronicle: Exploring the Latest in AI". Exploring the ever-expanding world of AI research is an exciting journey, full of promise and potential. With this new blog series, I aim to inspire and inform, unpacking the most interesting and visionary ideas in this rapidly evolving area. I covered this paper on my blog. - https://ajithp.com/2024/02/11/self-discover-large-language-models/
I created a custom GPT which automatically implements this Self-Discover 'Select - Adapt - Implement' approach, based on the paper; it utilizes the 39 reasoning modules provided in the paper also!
I'd love to get some feedback, I hope you all find it useful!
https://chat.openai.com/g/g-36cJS50di-self-discovering-gpt
I have had a little trouble getting it to reason the svg problem properly, if anyone has success or thinks they have an idea for improving the GPT instructions, I'd love to hear it!
I created a custom GPT which automatically implements this Self-Discover 'Select - Adapt - Implement' approach, based on the paper; it utilizes the 39 reasoning modules provided in the paper also!
I'd love to get some feedback, I hope you all find it useful!
https://chat.openai.com/g/g-36cJS50di-self-discovering-gptI have had a little trouble getting it to reason the svg problem properly, if anyone has success or thinks they have an idea for improving the GPT instructions, I'd love to hear it!
I got some interesting results using your prompt lol
@Flynnbo I don't have GPTPlus to try out CustomGPT.
So, I have open sourced Github repository to use the Self-Discover 'Select - Adapt - Implement' approach in your apps.
I looked at the one git hub shortly and asked myself if it coud be so simple.
https://github.com/meta-introspector/self-discover-prompt/issues/1 ran over some ideas with gemini on this if you are interested. My work on unimath and metacoq proof interpretation via the llm is how I would like to approach this. metacoq/unimath as one of the infintely many universal meta languages.
I implemented it using python. It works with openai models, Gemini Models and local GGUF models. You can even mix and match which models do the self-discover bit and which ones do the solving with the reasoning structure.
Would the code be available?
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper