Datasets:

slyq

/

papers-chunked

like 0

1. TAPE [22]: The motivation of TAPE is to leverage the 4.2 Text-levelEnhancementOgbn-arxiv Ogbn-products knowledge of LLMs to generate high-quality node fea- GCN MLP RevGAT SAGE SAGN MLP tures. Specifically,itusesLLMstogeneratepseudolabels For feature-level enhancement, LLMs in the pipeline must LM-phase GNN-phase GNN-phase Non-contextualizedShallowEmbeddingsInputfeatures Backbone LM-phaseRunningtime(s)Memory(GB) Runningtime(s) Memory(GB) TF-IDF 72.23 ±0.21 66.60 ±0.25 75.16 ±0.14 79.73 ±0.48 84.40 ±0.07 64.42 ±0.18 Word2Vec 71.74 ±0.29 55.50 ±0.23 73.78 ±0.19 81.33 ±0.79 84.12 ±0.18 69.27 ±0.54TF-IDFGCN N/A N/A 53 9.81 PLM/LLMEmbeddingswithoutFine-tuning(1024) RevGAT N/A N/A 873 7.32 Deberta-base 45.70 ±5.59 40.33 ±4.53 71.20 ±0.48 62.03 ±8.82 74.90 ±0.48 7.18 ±1.09Sentence-BERTGCN 239 1.30 48 7.11 LocalSentenceEmbeddingModels(384) RevGAT 239 1.30 674 4.37 Sentence-BERT(MiniLM) 73.10 ±0.25 71.62 ±0.10 76.94 ±0.11 82.51 ±0.53 84.79 ±0.23 72.73 ±0.34 text-ada-embedding-002 GCN 165 N/A 73 11.00 e5-large 73.74 ±0.12(1536)RevGAT 165 N/A 1038 8.3372.75 ±0.00 76.59 ±0.44 82.46 ±0.9185.47 ±0.2177.49 ±0.29 OnlineSentenceEmbeddingModelsDeberta-base GCN 13560 12.53 50 9.60 text-ada-embedding-002 72.76 ±0.23 72.17 ±0.00(768)RevGAT 13560 12.53 122 6.8276.64 ±0.2082.90 ±0.42 85.20 ±0.19 76.42 ±0.31 Fine-tunedPLMEmbeddings Fine-tunedDeberta-baseGLEM-GNN 74.65 ±0.12GCN 68071 18.22 N/A N/A72.90 ±0.11 75.80 ±0.39 82.15 ±0.16 84.01 ±0.0579.08 ±0.23 Others (768) RevGAT 68294 18.22 N/A N/A GIANT 73.29 ±0.10GIANTGCN N/A N/A 50 9.6073.06 ±0.11 75.90 ±0.1983.16 ±0.1986.67 ±0.0979.82 ±0.07 IterativeStructure(768) RevGAT N/A N/A 122 6.82 GLEM-GNN 75.93 ±0.19 N/A 76.97 ±0.19 83.16 ±0.09 87.36 ±0.07 N/A GLEM-LM 75.71 ±0.24 N/A 75.45 ±0.12 81.25 ±0.15 84.83 ±0.04 N/A and explanations. These explanations aim to make the 4.2.1 Experimental Setups logical relationship between the text features and cor- To evaluate these two strategies, we conduct experiments responding labels more cl

87.36 ±0.07 N/A GLEM-LM 75.71 ±0.24 N/A 75.45 ±0.12 81.25 ±0.15 84.83 ±0.04 N/A and explanations. These explanations aim to make the 4.2.1 Experimental Setups logical relationship between the text features and cor- To evaluate these two strategies, we conduct experiments responding labels more clear. For example, given the on two small datasets Cora and Pubmed considering the original attributes “mean-field approximation” and the cost to use the LLMs. For low labeling ratio and high la- ground truth label “probabilistic methods”, it will gener- beling ratio, we adopt the same setting as that in Table 1 ate a description such as “mean-field approximation is a and Table 2. For predictors, we adopt GCN, GAT, and widely adopted simplification technique for probabilistic MLP to study both the quality of textual embeddings be- models”, which makes the connection of these two at- fore and after aggregations. For LLMs, we adopt ChatGPT tributes much more clear. After generating pseudo labels with the latest version (gpt-3.5-turbo-0613). To better un- and explanations, they further adopt PLMs to be fine- derstand the effectiveness of TAPE, we separate it into TA, tuned on both the original text attributes and the expla- P, and E, where “TA” refers to “text attributes”, “P” refers nationsgeneratedbyLLMs,separately. Next,theygener- to “pseudo labels”, and “E” refers to “explanations”. For ate the corresponding text features and augmented text KEA, we try two approaches to inject the augmented tex- features based on the original text attributes and aug- tual attributes. The first approach is appending the aug- mented text attributes respectively, and finally ensemble mented textual attributes into the original attribute, which them together as the initial node features for GNNs. is denoted as “KEA-I”. Then the combined attributes are 2. Knowledge-Enhanced Augmentation : The motiva- encoded into features. The second approach is to encode tion behind Knowledge-Enhanced Augmentation (KEA) the augmented attributes and original attributes separately, is to enrich the text attributes by providing additional which is denoted as “KEA-S”. We report the results for information. KEA is inspired by knowledge-enhanced original, augmented, and ensembling features. Both TAPE PLMs such as ERNIE [61] and K-BERT [36] and aims and KEA adopt the cascading structures. After encoding to explicitly incorporate external knowledge. In KEA, the text attributes with LLMs, the generated embeddings we prompt the LLMs to generate a list of knowledge en

original, augmented, and ensembling features. Both TAPE PLMs such as ERNIE [61] and K-BERT [36] and aims and KEA adopt the cascading structures. After encoding to explicitly incorporate external knowledge. In KEA, the text attributes with LLMs, the generated embeddings we prompt the LLMs to generate a list of knowledge en- are adopted as the initial features for GNNs. We try two tities along with their text descriptions. For example, we approaches to encode the attributes, which are fine-tuned can generate a description for the abstract term “Hopf- PLMsandlocalsentenceembeddingmodels. Specifically,we Rinow theorem” as follows: “The Hopf-Rinow theorem adoptDeberta-baseande5-large. Toconductafaircompari- establishes that a Riemannian manifold, which is both son,wefirstdeterminethebettertextencoderbyevaluating complete and connected, is geodesically complete if and theiroverallperformance. Oncethetextencoderisselected, only if it is simply connected.” By providing such de- we proceed to compare the performance of the augmented scriptions, we establish a clearer connection between the attributes against the original attributes. theoremandthecategory“Riemanniangeometry”. Once A comprehensive evaluation of TAPE. We first gain weobtaintheentitylist,weencodeiteithertogetherwith a deeper understanding of TAPE through a comprehensive theoriginaltextattributeorseparately. Wetryencoding ablation study. The experimental results are shown in Ta- text attributes with fine-tuned PLMs and deep sentence ble5andTable6. Weshowtheapproachweadopttoencode embedding models. We also employ ensemble methods the text attributes in the bracket. In particular, we mainly to combine these embeddings. One potential advantage considerfine-tunedDeberta-base, whichisdenotedasPLM, of KEA is that it is loosely coupled with the prediction and e5-large, which is denoted as e5. performance of LLMs. In cases where LLMs generate in- Observation 7. The effectiveness of TAPE is mainly correct predictions, TAPE may potentially generate low- from the explanations E generated by LLMs. quality node features because the explanations provided Fromtheablationstudy,wecanseethatcomparedtopseudo byPLMsmayalsobeincorrect. However,withKEA,the labels P ,theexplanationspresentbetterstabilityacrossdif- augmented features may exhibit better stability since we ferent datasets. One main advantage of adopting explana- do not rely on explicit predictions from LLMs. tionsgeneratedbyLLMsisthattheseaugmentedattri

LMsmayalsobeincorrect. However,withKEA,the labels P ,theexplanationspresentbetterstabilityacrossdif- augmented features may exhibit better stability since we ferent datasets. One main advantage of adopting explana- do not rely on explicit predictions from LLMs. tionsgeneratedbyLLMsisthattheseaugmentedattributes present better performance in the low-labeling rate setting. From Table 5, we note that when choosing PLM as the en- coders, E performsmuchbetterthan TA inthelowlabeling rate setting. Compared to explanations, we find that the effectiveness of the P mainly depends on the zero-shot per- formanceofLLMs,whichmaypresentlargevariancesacross different datasets. In the following analysis, we use TA + E and neglect the pseudo labels generated by LLMs. Observation 8. Replacing fine-tuned PLMs with deep sentence embedding models can further im- prove the overall performance of TAPE. FromTable5andTable6,weobservethatadoptinge5-large as the LLMs to encode the text attributes can achieve good Figure 3: Illustrations for TAPE and KEA. TAPE leverages performanceacrossdifferentdatasetsanddifferentdatasplits. the knowledge of LLMs to generate explanations for their Specifically, the TA + E encoded with e5 can achieve top 3 predictions. For KEA, we prompt the LLMs to generate performanceinalmostallsettings. Inthefollowinganalysis, a list of technical terms with their descriptions. The main we adopt e5 to encode the original and enhanced attributes motivation is to augment the attribute information. TA + E . Table 5: A detailed ablation study of TAPE on Cora and Pubmed dataset in low labeling rate setting. For each combination of features and models, we useyellowto denote the best performance under a specific GNN/MLP model,greenthe second best one, andpinkthe third best one. Table 6: A detailed ablation

ion. TA + E . Table 5: A detailed ablation study of TAPE on Cora and Pubmed dataset in low labeling rate setting. For each combination of features and models, we useyellowto denote the best performance under a specific GNN/MLP model,greenthe second best one, andpinkthe third best one. Table 6: A detailed ablation study of TAPE on Cora and Pubmed dataset in the high labeling rate setting. For each combination of features and models, we useyellowto denote the best performance under a specific GNN/MLP model,green the second best one, andpinkthe third best one. 4.2.1.1 Effectiveness of KEA . 5. LLMSASTHEPREDICTORS We then show the results of KEA in Table 7 and Table 8. In the LLMs-as-Enhancers pipeline, the role of the LLMs For KEA-I ,weinjectthedescriptionofeachtechnicalterm remains somewhat limited since we only utilize their pre- directly into the original attribute. For KEA-S , we encode trained knowledge but overlook their reasoning capability. the generated description and original attribute separately. Drawing inspiration from the LLMs’ proficiency in handling Observation 9. The proposed knowledge enhance- complex tasks with implicit structures, such as logical rea- ment attributes KEA can enhance the performance soning[7]andrecommendation[14],wequestion: Isitpos- of the original attribute TA. sible for the LLM to independently perform predic- FromTable7andTable8,wefirstcomparetheperformance tive tasks on graph structures? By shifting our focus of features encoded by e5 and PLM. We see that the pro- to node attributes and overlooking the graph structures, we posed KEA is more fitted to the e5 encoder, and fine-tuned can perceive node classification as a text classification prob- PLM embeddings present poor performance on the low la- lem. In [60], the LLMs demonstrate significant promise, beling rate, thus we also select e5 as the encoder to further suggestingthattheycanproficientlyprocesstextattributes. compare the quality of attributes. From Table 9 we can see However, one key problem is that LLMs are not originally that the proposed KEA-I + TA and KEA-S + TA at- designed to process graph structures. Therefore, it can not tributes can consistently outperform the original attributes directly process structural information like GNNs. TA .

t LLMs are not originally that the proposed KEA-I + TA and KEA-S + TA at- designed to process graph structures. Therefore, it can not tributes can consistently outperform the original attributes directly process structural information like GNNs. TA . In this section, we aim to explore the potential of LLMs Observation 10. For different datasets, the most ef- as a predictor. In particular, we first check whether LLM fective enhancement methods may vary. can perform well without any structural information. Then, Moreover, we compare the performance of our proposed we further explore some prompts to incorporate structural KEA with TA + E , and the results are shown in Ta- information with natural languages. Finally, we show a case ble 9. We can see that on Cora , our methods can achieve study in Section 5.3 to explore its potential usage as an better performance while TA + E can achieve better per- annotator for graphs. formance on Pubmed . One potential explanation for this 5.1 HowCanLLMPerformonPopularGraph phenomenon is that TA + E relies more on the capabil- BenchmarkswithoutStructuralInforma- ity of LLMs. Although we have removed the pseudo labels tion? P , we find that the explanations still contain LLMs’ pre- dictions. As a result, the effectiveness of TA + E will be In this subsection, we treat the node classification problem influenced by LLMs’ performance on the dataset. As shown as a text classification problem by ignoring the structural in [22], the LLMs can achieve superior performance on the information. We adopt ChatGPT (gpt-3.5-turbo-0613) as Pubmed dataset but perform poorly on the Cora dataset. the LLMs to conduct all the experiments. We choose five Compared to TA + E , our proposed KEA only utilizes popular textual graph datasets with raw text attributes: the commonsense knowledge of the LLMs, which may have Cora [40], Citeseer [15], Pubmed [57], Ogbn-arxiv , and better stability across different datasets. Ogbn-products [23]. The details of these datasets can be foundinAppendixA.ConsideringthecoststoqueryLLMs’ Cora PubmedCora

bility across different datasets. Ogbn-products [23]. The details of these datasets can be foundinAppendixA.ConsideringthecoststoqueryLLMs’ Cora PubmedCora Pubmed GCN GAT MLP GCN GAT MLPGCN GAT MLP GCN GAT MLP TAPE 74.56 ±2.03 75.27 ±2.10 64.44 ±0.60TAPE 87.88 ±0.98 88.69 ±1.13 83.09 ±0.9192.22 ±1.3085.97 ±0.3193.35 ±1.5086.97 ±0.3395.05 ±0.2793.18 ±0.28 P 52.79 ±1.47 62.13 ±1.50 63.56 ±0.52 81.92 ±1.89P 64.90 ±1.39 80.11 ±4.01 70.31 ±1.91 85.73 ±0.59 91.60 ±0.62 93.65 ±0.3588.27 ±0.0193.27 ±0.15 TAPETAPE TA+E (e5)TA+E (e5)90.68 ±2.1283.38 ±0.4291.86 ±1.3684.00 ±0.0987.00 ±4.8375.73 ±0.5392.64 ±1.0087.44 ±0.4993.35 ±1.24 94.34 ±0.8686.71 ±0.9290.25 ±1.56 TA+E (PLM) 87.44 ±1.74 88.40 ±1.60 82.80 ±1.00 90.23 ±0.71TA+E (PLM) 78.02 ±0.56 64.08 ±12.36 55.72 ±11.98 80.70 ±1.73 79.66 ±3.08 76.42 ±2.1891.73 ±1.5895.40 ±0.32 E (PLM) 79.46 ±1.10 74.82 ±1.19 63.04 ±0.88 81.88 ±0.05 81.56 ±0.07 76.90 ±1.60E (PLM) 83.28 ±4.53 82.47 ±6.06 80.41 ±3.35 88.90 ±2.94 83.00 ±14.07 87.75 ±14.83 E (e5)E (e5) 89.39 ±2.6984.38 ±0.3690.13 ±2.5283.01 ±0.6084.05 ±4.03 89.68 ±0.78 90.61 ±1.61 91.09 ±0.8570.64 ±1.10 82.23 ±0.78 80.30 ±0.77 77.23 ±0.48 OriginalOriginal TA (PLM) 85.86 ±2.28 86.52 ±1.87 78.20 ±2.25TA (PLM) 59.23 ±1.16 57.38 ±2.01 30.98 ±0.68 62.12 ±0.07 61.57 ±0.07 53.65 ±0.2691.49 ±1.92 89.88 ±4.6394.65 ±0.13 attributesattributes TA (e5)TA (e5)90.53 ±2.3382.56 ±0.7389.10 ±3.2281.62 ±1.0986.19 ±4.38 89.65 ±0.85 89.55 ±1.16 91.39 ±0.4774.26 ±0.9382.63 ±1.13 79.67 ±0.80 80.38 ±1.94 Table 7: A detailed ablation study of KEA on Cora and Pubmed dataset in the low labeling rate setting. For each combination of features and models, we useyellowto denote the best performance under a specific GNN/MLP model,green the second best one, andpinkthe third best one. Table 8: A detailed ablation study of KEA on Cora and Pubmed dataset in the high labeling rate setting. For each combination of features and models, we useyellowto denote the best performance under a specific GNN/MLP model,green the second best one, andpinkthe third best one. Table 9: Comparison of the performance of TA, KEA-I, and KEA-S, and TA + E. The best performance is shown with an underline. Cora (low) means a low labeling rate setting, and Cora (high) denotes a high labeling rate setting. APIs, it’s not possible for us to test the whole dataset for gether with their labels for LLMs to better understand these graphs. Considering the rate limit imposed by Ope- thetask. Inadditiontothenode’scontent,thisapproach nAI 4, we randomly select 200 nodes from the test sets

high labeling rate setting. APIs, it’s not possible for us to test the whole dataset for gether with their labels for LLMs to better understand these graphs. Considering the rate limit imposed by Ope- thetask. Inadditiontothenode’scontent,thisapproach nAI 4, we randomly select 200 nodes from the test sets as integratesthecontentandlabelsofrandomlyselectedin- our test data. In order to ensure that these 200 nodes bet- context samples from the training set. In the section, we ter represent the performance of the entire set, we repeat all adopt random sampling to select few-shot prompts. experiments twice. Additionally, we employ zero-shot per- 3. Zero-shotpromptswithChain-of-Thoughts(CoT) : formance as a sanity check, comparing it with the results in CoT [70] presents its effectiveness in various reasoning TAPE [22] to ensure minimal discrepancies. tasks, which can greatly improve LLMs’ reasoning abil- We explore the following strategies: ities. In this study, we test whether CoT can improve 1. Zero-shot prompts : This approach solely involves the LLMs’ capability on node classification tasks. On the attribute of a given node. basis of zero-shot prompts, we guide the LLMs to gen- 2. Few-shot prompts : On the basis of zero-shot prompts, erate the thought process by using the prompt ”think it few-shotprompts provide in-context learningsamples to- step by step”. 4. Few-shot prompts with CoT : Inspired by [82], which 4 https://platform.openai.com/docs/guides/ demonstrates that incorporating the CoT process gen- rate-limits/overview erated by LLMs can further improve LLMs’ reasoning Cora (low) Pubmed (low) GCN GAT MLP GCN GAT MLPCora PubmedCora Pubmed TA 82.56 ± 0.73 81.62 ± 1.09 74.26 ± 0.93 82.63 ± 1.13 79.67 ± 0.80 80.38 ± 1.94GCN GAT MLP GCN GAT MLPGCN GAT MLP GCN GAT MLP OriginalOriginalKEA-I + TA 83.20 ± 0.56 83.38 ± 0.63 74.34 ± 0.97 83.30 ± 1.75 81.16 ± 0.87 80.74 ± 2.44TA(PLM) 59.23 ±1.16 57.38 ±2.01 30.98 ±0.68 62.12 ±0.07 61.57 ±0.

Pubmed TA 82.56 ± 0.73 81.62 ± 1.09 74.26 ± 0.93 82.63 ± 1.13 79.67 ± 0.80 80.38 ± 1.94GCN GAT MLP GCN GAT MLPGCN GAT MLP GCN GAT MLP OriginalOriginalKEA-I + TA 83.20 ± 0.56 83.38 ± 0.63 74.34 ± 0.97 83.30 ± 1.75 81.16 ± 0.87 80.74 ± 2.44TA(PLM) 59.23 ±1.16 57.38 ±2.01 30.98 ±0.68 62.12 ±0.07 61.57 ±0.07 53.65 ±0.26TA (PLM) 85.86 ±2.28 86.52 ±1.87 78.20 ±2.2591.49 ±1.92 89.88 ±4.6394.65 ±0.13 attributes TA(e5) 82.56 ±0.73 81.62 ±1.09 74.26 ±0.93 82.63 ±1.13 79.67 ±0.80 80.38 ±1.94 Attributes TA (e5) 90.53 ±2.33 89.10 ±3.22 86.19 ±4.38 89.65 ±0.85 89.55 ±1.16 91.39 ±0.47 KEA-S + TA 84.63 ± 0.58 85.02 ± 0.40 76.11 ± 2.66 82.93 ± 2.38 81.34 ± 1.51 80.74 ± 2.44 TA+E 83.38 ± 0.42 84.00 ± 0.09 75.73 ± 0.53KEA-I+TA(e5)KEA-I+TA (e5)83.20 ±0.5691.12 ±1.7690.24 ±2.9383.38 ±0.6387.88 ±4.44 90.19 ±0.83 90.60 ±1.22 92.12 ±0.7474.34 ±0.9787.44±83.30 ±1.750.4986.7181.16 ±0.87±0.9290.2580.74 ±2.44±1.56 KEA-I+TA(PLM) 53.21 ±11.54 55.38 ±4.64 31.80 ±3.63 57.13 ±8.20 58.66 ±4.27 52.28 ±4.47KEA-I+TA (PLM) 87.07 ±1.04 87.66 ±0.86 79.12 ±2.77Cora (high) Pubmed (high)92.32 ±0.6492.29 ±1.4394.85 ±0.20 KEA-I(e5) 81.35 ±0.77 82.04 ±0.72 70.64 ±1.10 81.98 ±0.91KEA-I (e5)91.09 ±1.78 90.13 ±2.7686.78 ±4.12 89.56 ±0.82 90.25 ±1.34 91.92 ±0.8081.04 ±1.39 79.73 ±1.63 KEA KEA-I(PLM) 36.68 ±18.63 37.69 ±12.79 30.46 ±0.60 56.22 ±7.17 59.33 ±1.69 52.79 ±0.51KEA-I (PLM) 86.08 ±2.35 85.23 ±3.15 77.97 ±2.87GCN GAT MLP GCN GAT MLP91.73 ±0.5891.93 ±1.7694.76 ±0.33 KEA KEA-S+TA(e5)KEA-S+TA (e5)84.63 ±0.5891.09 ±1.7892.30 ±1.6985.02 ±0.4088.95 ±4.96 90.40 ±0.9276.11 ±2.6682.93 ±2.3890.82 ±1.30 91.78 ±0.5681.34 ±1.5180.74 ±2.44 TA 90.53 ± 2.33 89.10 ± 3.22 86.19 ± 4.38 89.65 ± 0.85 89.55 ± 1.16 91.39 ± 0.47KEA-S+TA(PLM) 51.36 ±16.13 52.85 ±7.00 34.56 ±5.09 59.47 ±6.09 51.93 ±3.27 51.11 ±2.63KEA-S+TA (PLM) 83.98 ±5.13 87.33 ±1.68 80.04 ±1.32 86.11 ±5.68 89.04 ±5.82 94.35 ±0.48 KEA-I + TA KEA-S(e5)KEA-S (e5) 89.39 ±2.6991.12±1.76 90.24 ± 2.93 87.88 ± 4.44 90.19 ± 0.83 90.60 ± 1.22 92.12 ± 0.7484.38 ±0.3690.13 ±2.52 84.05 ±4.03 89.68 ±0.78 90.61 ±1.61 91.09 ±0.8583.01 ±0.60 70.64 ±1.10 82.23 ±0.78 80.30 ±0.77 77.23 ±0.48 KEA-S + TA 91.09 ± 1.78KEA-S(PLM) 28.97 ±18.24 43.88 ±10.31 30.36 ±0.58 61.22 ±0.94 54.93 ±1.55 47.94 ±0.89KEA-S (PLM) 83.35 ±7.30 85.67 ±2.00 76.76 ±1.82 79.68 ±19.57 69.90 ±19.75 85.91 ±6.4792.30±1.6988.95±4.96 90.40 ± 0.92 90.82 ± 1.30 91.78 ± 0.56 TA+E 90.68 ± 2.12 91.86 ± 1.36 87.00 ± 4.83 92.64 ± 1.00 93.35 ± 1.24 94.34 ± 0.86 capabilities. Building upon the few-shot prompts, this soning capability [70]. However, we find that it’s not effec- approach enables the LLMs to generate a step-by-step tive for the node classification task. This phenomenon can thoughtprocessforthein-contextsamples. Subsequently,

4.83 92.64 ± 1.00 93.35 ± 1.24 94.34 ± 0.86 capabilities. Building upon the few-shot prompts, this soning capability [70]. However, we find that it’s not effec- approach enables the LLMs to generate a step-by-step tive for the node classification task. This phenomenon can thoughtprocessforthein-contextsamples. Subsequently, bepotentiallyexplainedby Observation12 . Incontrastto thegeneratedCoTprocessesareinsertedintotheprompt mathematical reasoning, where a single answer is typically as auxiliary information. expected,multiplereasonablechainsofthoughtcanexistfor Output Parsing. In addition, we need a parser to ex- node classification. An example is shown in Table 12. This tract the output from LLMs. We devise a straightforward phenomenon poses a challenge for LLMs as they may strug- approach to retrieve the predictions from the outputs. Ini- gle to match the ground truth labels due to the presence of tially, we instruct the LLMs to generate the results in a multiple reasonable labels. formatted output like “a python list”. Then, we can use Observation 14. For prompts that are very similar the symbols “[” and “]” to locate the expected outputs. It in semantics, there may be huge differences in their shouldbenotedthatthisdesignaimstoextracttheinforma- effects. tion more easily but has little influence on the performance. Inaddition,weobservethatTAPE[22]implementsaunique We observe that sometimes LLMs will output contents that prompt on the Ogbn-arxiv dataset, yielding impressive re- are slightly different from the expected format, for exam- sults via zero-shot prompts. The primary distinction be- ple, output the expected format “Information Retrieval” to tween their prompts and ours lies in the label design. Given “Information Extraction”. In such cases, we compute the that all papers originate from the computer science sub- edit distance between the extracted output and the cate- category of Arxiv, they employ the brief term ”arxiv cs gory names and select the one with the smallest distance. subcategories” as a substitute for these 40 categories. Re- This method proves effective when the input context is rela- markably, this minor alteration contributes to a substan- tively short. If this strategy encounters errors, we resort to tial enhancement in performance. To delve deeper into this extracting the first mentioned categories in the output texts phenomenon, we experiment with three disparate label de- as the predictions. If there’s no match, then the model’s signs: (1) Strategy 1: the original Arxiv identifier, such as prediction

egy encounters errors, we resort to tial enhancement in performance. To delve deeper into this extracting the first mentioned categories in the output texts phenomenon, we experiment with three disparate label de- as the predictions. If there’s no match, then the model’s signs: (1) Strategy 1: the original Arxiv identifier, such as prediction for the node is incorrect. ”arxivcs.CV”; (2)Strategy2: naturallanguagedescriptors, ToreducethevarianceofLLMs’predictions,wesetthetem- like ”computer vision”; and (3) Strategy 3: the specialized peratureto0. Forfew-shotcases, wefindthatprovidingtoo prompt, utilizing ”arxiv cs subcategory” to denote all cat- much context will cause LLMs to generate outputs that are egories. Unexpectedly, we discover that Strategy 3 signifi- notcompatiblewiththeexpectedformats. Therefore,weset cantly outperforms the other two (refer to Table 13). a maximum number of samples to ensure that LLMs gener- GiventhatLLMsundergopre-trainingonextensivetextcor- ateoutputswithvalidformats. Inthisstudy, wechoosethis pora, it’s likely that these corpora include papers from the number to 2 and adopt accuracy as the performance metric. Arxiv database. That specific prompt could potentially en- 5.1.1 Observations hancethe“activation”ofthesemodels’correspondingmem- ory. However, the reason for the excellent results achieved Observation 11. LLMs present preliminary effec- by this kind of prompt might not stem from the simple data tiveness on some datasets. memorization of the LLM [25]. When applying to papers AccordingtotheresultsinTable10, itisevidentthatLLMs after 2023 that are not included in the pre-training corpus demonstrateremarkablezero-shotperformanceon Pubmed . of the LLMs, this prompt also achieves similar effectiveness. When it comes to Ogbn-products , LLMs can achieve per- This phenomenon reminds us that when using ChatGPT, formance levels comparable to fine-tuned PLMs. However, sometimes providing more information in the prompt (such there is a noticeable performance gap between LLMs and ascategoryinformationfromthe Ogbn-arxiv dataset)may GNNs on Cora and Pubmed datasets. To gain a deeper actually lead to a decrease in performance. understanding of this observation, it is essential to analyze the output of LLMs. 5.2 Incorporating Structural Information in Observation 12. Wrong predictions made by LLMs thePrompts are sometimes also reasonable. After investigating the output of LLMs, we find th

actually lead to a decrease in performance. understanding of this observation, it is essential to analyze the output of LLMs. 5.2 Incorporating Structural Information in Observation 12. Wrong predictions made by LLMs thePrompts are sometimes also reasonable. After investigating the output of LLMs, we find that a part As we note, LLMs can already present superior zero-shot of the wrong predictions made by LLMs are very reason- performance on some datasets without providing any struc- able. An example is shown in Table 11. In this example, tural information. However, there is still a large perfor- we can see that besides the ground truth label ”Reinforce- mance gap between LLMs and GNNs in Cora , Citeseer , ment Learning”, ”Neural Networks” is also a reasonable la- and Ogbn-arxiv . Then a question naturally raises that bel, which also appears in the texts. We find that this is a whether we can further increase LLMs’ performance by in- common problem for Cora , Citeseer , and Ogbn-arxiv . corporating structural information? To answer this prob- For Ogbn-arxiv , there are usually multiple labels for one lem, we first need to identify how to denote the structural paperonthewebsite. However,inthe Ogbn-arxiv dataset, informationintheprompt. LLMssuchasChatGPTarenot only one of them is chosen as the ground truth. This leads originallydesignedforgraphstructures,sotheycannotpro- to a misalignment between LLMs’ commonsense knowledge cess adjacency matrices like GNNs. In this part, we study and the annotation bias inherent in these datasets. More- several ways to convey structural information and test their over, we find that introducing few-shot samples presents lit- effectiveness on the Cora dataset. tle help to mitigate the annotation bias. Specifically, we first consider inputting the whole graph into Observation 13. Chain-of-thoughts do not bring in the LLMs. Using Cora dataset as an example, we try to performance gain. use prompts like “node 1: ⟨paper content ⟩” to represent Forreasoningtasksinthegeneraldomain,chain-of-thoughts attributes, and prompts like “node 1 cites node 2” to rep- isbelievedtobeaneffectiveapproachtoincreaseLLM’srea- resent the edge. However, we find that this approach is not Table 10: Performance of LLMs on real-world text attributed graphs without structural information, we also include the result of GCN (or SAGE for Ogbn-products ) together with Sentence-BERT features. For Cora , Citeseer , Pubmed , we show the results of the low labeling rate sett

resent the edge. However, we find that this approach is not Table 10: Performance of LLMs on real-world text attributed graphs without structural information, we also include the result of GCN (or SAGE for Ogbn-products ) together with Sentence-BERT features. For Cora , Citeseer , Pubmed , we show the results of the low labeling rate setting. Table11: AwrongbutreasonablepredictionmadebyLLMs GNNs typically have 2 layers, indicating that the 2-hop neighbor information is the most useful in most cases. Con- Paper: The Neural Network House: An overview; Typi- sideringtheinputcontextlimitofLLMs,weempiricallyfind cal home comfort systems utilize only rudimentary forms that each time we can summarize the attribute information of energy management and conservation. The most sophis- of 5 neighbors. In this paper, we sample neighbors once and ticated technology in common use today is an automatic only summarize those selected neighbors. In practice, we setback thermostat. Tremendous potential remains for im- can sample multiple times and summarize each of them to proving the efficiency of electric and gas usage... obtain more fine-grained neighborhood information. Ground Truth: Reinforcement Learning Observation 15. Neighborhood summarization is LLM’s Prediction: Neural Networks likely to achieve performance gain. FromTable14,wenotethatincorporatingneighborhoodin- formation in either zero-shot or few-shot approaches yields performance gains compared to the zero-shot prompt with- out structural information except on the Pubmed dataset. Table 12: An example that LLMs generate CoT processes By following the ”homophily” assumption [87; 39], which not matching with ground truth labels suggests that neighboring nodes tend to share the same la- bels, the inclusion of neighboring information can poten- Paper: The Neural Network House: An overview.: Typ- tially alleviate annotation bias. For instance, let’s consider

suggests that neighboring nodes tend to share the same la- bels, the inclusion of neighboring information can poten- Paper: The Neural Network House: An overview.: Typ- tially alleviate annotation bias. For instance, let’s consider ical home comfort systems utilize only rudimentary forms apaperfromArxivcoveringgeneraltopicsliketransformers. of energy management and conservation. The most sophis- Merely analyzing the content of this paper makes it difficult ticated technology in common use today is an automatic to determine which category the author would choose, as setback thermostat. Tremendous potential remains for im- categories such as ”Artificial Intelligence,” ”Machine Learn- proving the efficiency of electric and gas usage... ing,”and”ComputerVision”areallplausibleoptions. How- Generated Chain-of-thoughts: The paper discusses the ever, by examining its citation relationships, we can better use of neural networks for intelligent control and mentions infer the author’s bias. If the paper cites numerous sources the utilization of neural network reinforcement learning and from the ”Computer Vision” domain, it is likely that the prediction techniques. Therefore, the most likely category author is also a researcher in that field, thereby favoring for this paper is ’Neural Networks’. the selection of this category. Consequently, structural in- Ground Truth: Reinforcement Learning formation provides implicit supervision to assist LLMs in LLM’s Prediction: Neural Networks capturing the inherent annotation bias in the dataset. How- ever, fromthe Pubmed dataset, weobservethatincorporat- ing neighborhood information results in clear performance drop, which necessitates a deep analysis below. Observation 16. LLMs with structure prompts may Table 13: Performance of LLMs on OGB-Arxiv dataset, suffer from heterophilous neighboring nodes. with three different label designs. From Table 14, we observe that LLMs perform worse on

Observation 16. LLMs with structure prompts may Table 13: Performance of LLMs on OGB-Arxiv dataset, suffer from heterophilous neighboring nodes. with three different label designs. From Table 14, we observe that LLMs perform worse on Pubmed after incorporating the structural information. To Strategy 1 Strategy 2 Strategy 3 gainadeeperunderstanding, wefocusonthosenodeswhere Ogbn-arxiv 48.5 51.8 74.5 zero-shot prompts without structural information can lead to correct prediction but prompts with 2-hop information can’t. An example of this kind of node is shown in Table 15. feasible since LLMs usually present a small input context After analyzing the 2-hop neighbors of this node, we find length restriction. As a result, we consider an “ego-graph” that 15 out of 19 2-hop neighboring nodes have different view, which refers to the subgraphs induced from the center labels against this node. This case is usually denoted as nodes. In this way, we can narrow the number of nodes to ”heterophily” [87], which is a phenomenon in graph theory be considered. where nodes in a graph tend to connect with nodes that are Specifically, we first organize the neighbors of the current dissimilar to them. In this case, we find that both GNNs nodes as a list of dictionaries consisting of attributes and la- and LLMs with a structure-aware prompt make wrong pre- bels of the neighboring nodes for training nodes. Then, the dictions. However, LLMs ignoring structural information LLMs summarize the neighborhood information. It should get correct predictions, which indicates that LLMs with be noted that we only consider 2-hop neighbors because a structure-aware prompt may also suffer from the ”het- Cora Citeseer Pubmed Ogbn-arxiv Ogbn-products Zero-shot 67.00 ± 1.41 65.50 ± 3.53 90.75 ± 5.30 51.75 ± 3.89 70.75 ± 2.48Few-shot 67.75 ± 3.53 66.00 ± 5.66 85.50 ± 2.80 50.25 ± 1.06 77.75 ± 1.06 Zero-shot with COT 64.00 ± 0.71 66.50 ± 2.82 86.25 ± 3.29 50.50 ± 1.41 71.25 ± 1.06 Few-shot with COT 64.00 ± 1.

Cora Citeseer Pubmed Ogbn-arxiv Ogbn-products Zero-shot 67.00 ± 1.41 65.50 ± 3.53 90.75 ± 5.30 51.75 ± 3.89 70.75 ± 2.48Few-shot 67.75 ± 3.53 66.00 ± 5.66 85.50 ± 2.80 50.25 ± 1.06 77.75 ± 1.06 Zero-shot with COT 64.00 ± 0.71 66.50 ± 2.82 86.25 ± 3.29 50.50 ± 1.41 71.25 ± 1.06 Few-shot with COT 64.00 ± 1.41 60.50 ± 4.94 85.50 ± 4.94 47.25 ± 2.47 73.25 ± 1.77 GCN/SAGE 82.20 ± 0.49 71.19 ± 1.10 81.01 ± 1.32 73.10 ± 0.25 82.51 ± 0.53 Table 14: Performance of LLMs on real-world text attributed graphs with summarized neighborhood information. For Cora , Citeseer , Pubmed , we show the results of the low labeling rate setting. We also include the result of GCN (or SAGE for Ogbn-products ) together with Sentence-BERT features. erophily” problem. significant performance improvement compared to others. Consequently, the primary challenge can be summarized as Table 15: GNNs and LLMs with structure-aware prompts follows: how can we effectively select both the critical nodes are both wrong within the graph and the reliable nodes in the context of LLMs? Paper: Title: C-reactiveproteinandincidentcardiovascular Taking into account the complexity of these two challenges, events among men with diabetes. we don’t intend to comprehensively address them in this Abstract: OBJECTIVE: Several large prospective studies paper. Instead, we present a preliminary study to evaluate have shown that baseline levels of C-reactive protein (CRP) the performance of a simple strategy: randomly selecting a ... subset of nodes for annotation. It is worth noting that ad- Neighbor Summary: This paper focuses on different aspects vanced selection strategies such as active learning [72] could of type2diabetes mellitus. Itexploresthelevelsofvarious beadoptedtoimprovethefinalperformance. Weleavesuch markerssuchastumornecrosisfactor-alpha,interleukin-2... exploration as future work. Regarding the annotation bud- Ground truth: ”Diabetes Mellitus Type 1” get, we adopt a ”low labeling rate” setting, wherein we ran- Structure-ignorant prompts: ”Diabetes Mellitus domly select a total of 20 nodes multiplied by the number Type 1”

exploration as future work. Regarding the annotation bud- Ground truth: ”Diabetes Mellitus Type 1” get, we adopt a ”low labeling rate” setting, wherein we ran- Structure-ignorant prompts: ”Diabetes Mellitus domly select a total of 20 nodes multiplied by the number Type 1” of classes. For the selected nodes, we adopt 75% of them Structure-aware prompt: ”Diabetes Mellitus Type as training nodes and the rest as validation nodes. Con- 2” sequently, we annotate a total of 140 nodes in the Cora GNN: ”Diabetes Mellitus Type 2” dataset and 60 nodes in the Pubmed dataset. In this part, we use GCN as the GNN model and adopt the embeddings generated by the Sentence-BERT model. The results are shown in Table 16. We can observe that training GCN on the pseudo labels can lead to satisfying performance. Par- ticularly, it can match the performance of GCN trained on 5.3 CaseStudy: LLMsasthePseudoAnnota- ground truth labels with 10 shots per class. As a refer- tors ence, around 67% of the pseudo labels for Cora can match From Table 10, we show that LLMs can be good zero-shot ground truth labels, while around 93% of the pseudo labels predictors onseveralreal-worldgraphs,whichprovidesthe for Pubmed are ground truth labels. possibility to conduct zero-shot inference on datasets with- Table 16: Performance of GCN trained on either pseudo outlabels. DespitetheeffectivenessofLLMs,itstillpresents labels generated by LLMs, or ground truth labels two problems: (1) The price of using LLMs’ API is not cheap,andconductinginferenceonalltestingnodesforlarge Cora Pubmed graphsincurshighcosts; (2)Whetheritisalocallydeployed Using pseudo labels open-source LLM or a closed source LLM accessed through 20 shots × #class 64.95 ± 0.98 71.70 ± 1.06 anAP

Cora Pubmed graphsincurshighcosts; (2)Whetheritisalocallydeployed Using pseudo labels open-source LLM or a closed source LLM accessed through 20 shots × #class 64.95 ± 0.98 71.70 ± 1.06 anAPI,theinferencewiththeseLLMsaremuchslowerthan Using ground truth GNNs, since the former has high computational resource re- 3 shots per class 52.63 ± 1.46 59.35 ± 2.67 quirements, while the latter has rate limits. One potential 5 shots per class 58.97 ± 1.41 65.98 ± 0.74 solution to these challenges is leveraging the knowledge of 10 shots per class 69.87 ± 2.27 71.51 ± 0.77 LLMs to train smaller models like GNNs, which inspires a potential application of LLMs to be used as annotators. Basedonthepreliminaryexperimentaloutcomes,LLMsdis- Observation 17. The quality of pseudo labels is key play encouraging results on certain datasets, thus highlight- to downstream performance. ingtheirpotentialforgeneratinghigh-qualitypseudo-labels. Although we don’t place significant emphasis on the selec- However, the use of LLMs as an annotator introduces a tionofnodestobelabeled,thepreliminaryresultsshowthat new challenge. A key consideration lies in deciding the there is relatively little variance among different random se- nodes that should be annotated. Unlike the self-labeling lections. Comparing this to the impact of pseudo labels, we in GNNs[8; 34; 32], where confidence-based or information- observe that the quality of pseudo labels can make a sig- basedmetricsareemployedtoestimatethequalityofpseudo- nificant difference. When higher quality pseudo labels are labels. Itremainsadifficulttasktodeterminetheconfidence used, GNNs perform much better on Pubmed compared to of pseudo-labels generated by LLMs. Additionally, differ- Cora . This result highlights the importance of developing ent nodes within a graph have distinct impacts on other an approach to select confident nodes for LLMs. nodes [72]. Annotating certain nodes can result in a more Observation 18. Getting the confidence by simply Cora Citeseer Pubmed Ogbn-arxiv Ogbn-products Zero-shot 67.00 ±1.41 65.50 ±3.53 90.75 ±5.30 51.75 ±3.89 70.75 ±2.48 Zero-Shotwith2-hopinfo 71.75 ±0.35 62.00 ±1.41 88.00

nodes [72]. Annotating certain nodes can result in a more Observation 18. Getting the confidence by simply Cora Citeseer Pubmed Ogbn-arxiv Ogbn-products Zero-shot 67.00 ±1.41 65.50 ±3.53 90.75 ±5.30 51.75 ±3.89 70.75 ±2.48 Zero-Shotwith2-hopinfo 71.75 ±0.35 62.00 ±1.41 88.00 ±1.41 55.00 ±2.83 75.25 ±3.53Few-shot 67.75 ±3.53 66.00 ±5.66 85.50 ±2.80 50.25 ±1.06 77.75 ±1.06 Few-Shotwith2-hopinfo 74.00 ±4.24 67.00 ±4.94 79.25 ±6.71 52.25 ±3.18 76.00 ±2.82 GCN/SAGE 82.20 ±0.49 71.19 ±1.10 81.01 ±1.32 73.10 ±0.25 82.51 ±0.53prompting the LLMs may not work since they are Observation 19. LLMs-as-Predictors demonstrate too “confident”. robustness when facing OOD data. Basedonpreviousobservations,wechecksomesimplestrate- From Table 18, we find that LLMs-as-Predictors present gies to achieve the confidence level of LLMs’ outputs. Ini- promisingrobustnessagainstOODdata. Itshouldbenoted tially,weattempttoprompttheLLMsdirectlyfortheircon- that we only try a simple structure-ignorant prompt, and fidence level. However, we discover that most of the time, we may further improve the OOD performance of LLMs by LLMs simply output a value of 1, rendering it meaningless. selectingproperin-contextsamplesandincorporatingstruc- Examples are shown in Table 17. tural information. In a nutshell, LLMs present great poten- tial to enhance the OOD generalization capability of graph Table 17: Prompts used to generate neighbor summary models. Instruction 6. RELATEDWORK Output the confidence level in the range of 0 to 1 and the Following our proposed two pipelines, i.e., LLMs as the En- most 1 possible category of this paper as a python dict, like hancers and LLMs as the Predictors, we review existing ”prediction”: ”XX”, ”confidence”: ”XX” works in this section. 6.1 LLMsastheEnhancers Intherecentsurgeofresearch,increasingattentionhasbeen Another potential solution is to utilize LLMs that support paid on the intersection of LLMs and GNNs in the realm of prediction logits, such as text-davinci-003. However, we ob- TAGs [83; 6; 78; 77; 49; 22; 86; 24; 33; 10]. Compared to serve that the probability of the outputs from these models shallow embeddings, LLMs can provide a richer repository is consistently close to 1, rendering the output not helpful.

paid on the intersection of LLMs and GNNs in the realm of prediction logits, such as text-davinci-003. However, we ob- TAGs [83; 6; 78; 77; 49; 22; 86; 24; 33; 10]. Compared to serve that the probability of the outputs from these models shallow embeddings, LLMs can provide a richer repository is consistently close to 1, rendering the output not helpful. ofcommonsenseknowledge,whichcouldpotentiallyenhance the performance of downstream tasks [51]. 5.4 Case Study: Applying LLMs to handle Several studies employ PLMs as text encoders, transform- out-of-distributiondata ing text attributes into node features, which can thus be classified as feature-level enhancement . The integration Out-of-distribution(OOD)learningaddressesscenarioswhere structuresvaryamongtheseworks: someadoptasimplecas- training and test data are drawn from different distribu- cading structure [49; 6; 78; 37], while others opt for an iter- tions. Given the ubiquity of distribution shifts in graph ativestructure[83; 74; 77]. Forthoseutilizingthecascading data [29], OOD generalization on graphs has emerged as a structure,preliminaryinvestigationshavebeenconductedto crucial research direction in recent years. A recent bench- determine how the quality of text embeddings affects down- mark, GOOD [17], reveals that existing GNN-based mod- stream classification performance [49]. GIANT [6] attempts els struggle with robustness when confronted with distri- to incorporate structural information into the pre-training butional shifts. In contrast, LLMs have demonstrated com- stage of PLMs, achieving improved performance albeit with mendablerobustnessontextualdatainthepresenceofOOD additionaltrainingoverhead. SimTEG[10]suggeststhatus- scenarios [67]. Node classification on the TAG, when disre- ing embeddings obtained through efficiently fine-tuned pa- garding graph structures, can also be considered as a text rameters to replace the original embeddings of pre-trained classification task. Therefore, in this section, we initiate language models can solve the problem of overfitting dur- a preliminary exploration into the application of LLMs for ing fine-tuning, thereby further enhancing the performance OOD scenarios on graphs. of the cascading structure. OneForAll [33] further adopts ExperimentalSetups. WeadopttheGOOD-Arxivdataset sentence embedding model to unify the feature space, and from the GOOD benchmark [17] considering its text at- propose a unified model for divers

erformance OOD scenarios on graphs. of the cascading structure. OneForAll [33] further adopts ExperimentalSetups. WeadopttheGOOD-Arxivdataset sentence embedding model to unify the feature space, and from the GOOD benchmark [17] considering its text at- propose a unified model for diverse tasks across multiple tribute availability. Specifically, we adopt all four types datasets. Thiscascadingstructurehasalsobeensuccessfully of the OOD shift: “Concept-degree”, “Covariate-degree”, applied to tasks such as fact verification [37] and question “Concept-time”,and“Covariate-time”fromtheGOOD.The answering [78]. However, despite its simplicity, recent stud- final results are shown in Table 18. We adopt the prompt ies [83] have identified potential drawbacks of the cascading from TAPE [22] since it achieves better performance on the structure. Specifically, it establishes a tenuous connection Ogbn-arxiv dataset. For comparison, we take the best between the text attribute and the graph. The embeddings baseline models from the GOOD benchmark. generated by the PLMs do not take graph structures into account, and the parameters of the PLMs remain constant Table 18: OOD performance comparison. “Val” means during the GNN training process. Alternatively, in the iter- the results on the IID validation sets. “Test” indicates ativestructure, Graphformers[74]facilitatestheco-training the results of the OOD test sets. We can see that LLMs- of PLMs and GNNs using each other’s generated embed- as-Predictors consistently outperform the best GNN-based dings. GLEM [83] takes this a step further by considering OODbaselines. Moreover,thegapbetweenIIDperformance pseudolabelsgeneratedbybothPLMsandGNNsandincor- and OOD performance is small. poratingthemintotheoptimizationprocess. DRAGON[77] successfullyextendstheiterativestructuretotheknowledge Val Test Best baseline (test) graph domain. concept degree 73.01 72.79 63.00 ComparedtothesestudiesfocusingonPLMs, arecentstud- covariate degree 70.23 68.21 59.08 y[22]considerstheusageofembedding-invisibleLLMssuch concept time 72.66 71.98 67.45 asChatGPT[45]forrepresentationlearningonTAGs,which covariate time 74.28 74.37 71.34 aims to adopt LLMs to enhance the text attributes and thuscanbecategorizedinto

geofembedding-invisibleLLMssuch concept time 72.66 71.98 67.45 asChatGPT[45]forrepresentationlearningonTAGs,which covariate time 74.28 74.37 71.34 aims to adopt LLMs to enhance the text attributes and thuscanbecategorizedinto text-level enhancement . Thisworkintroducesapromptdesignedtogenerateexplanations contrastivelearningtoalignthegraphandtextfeaturespaces. for the predictions made by LLMs. These generated expla- It also introduces dual-stage instruction tuning, where the nations are subsequently encoded into augmented features first stage adopts self-supervised instruction tuning to make by PLMs. Through the ensemble of these augmented fea- LLMsbetterunderstandgraph-structuredinformation. The tures with the original features, the proposed methodology second stage adopts task-specific fine-tuning to allow LLMs demonstrates its efficacy and accomplishes state-of-the-art achieve task-specific knowledge and then make predictions. performance on the Ogbn-arxiv leaderboard [23]. Never- GraphLLM [3] and DGTL [50] apply this pipeline to graph theless, the study offers limited analytical insights into the reasoning tasks and graph representation learning. underlying reasons for the success of this approach. Addi- tionally,wehaveidentifiedapotentialconcernregardingthe 7. CONCLUSIONS,LIMITATIONS,ANDFU- prompts utilized in the referenced study. Another work pertaining to the integration of LLMs and TUREDIRECTIONS GNNs is the Graph-Toolformer [80]. Drawing inspirations In this section, we summarize our key findings, present the from Toolformer [56], this study utilizes LLMs as an inter- limitations of this study and discuss the potential directions face to bridge the natural language commands and GNNs. of leveraging LLMs in graph machine learning. This approach doesn’t change the features and training of GNNs, which is out of our scope. 7.1 KeyFindings In this paper, we propose two potential pipelines: LLMs-as- 6.2 LLMsasthePredictors Enhancers and LLMs-as-Predictors that incorporate LLMs While LLMs-as-Enhancers have proven to be effective, the to handle the text-attributed graphs. Our rigorous empiri- pipeline still requires GNNs for final predictions. In a sig- cal studies reveal several interesting findings which provide nificant shift from this approach, recent studies [18; 65] new insights for future studies. We highlight some key find- have begun exploring a unique pipeline that solely relies

to handle the text-attributed graphs. Our rigorous empiri- pipeline still requires GNNs for final predictions. In a sig- cal studies reveal several interesting findings which provide nificant shift from this approach, recent studies [18; 65] new insights for future studies. We highlight some key find- have begun exploring a unique pipeline that solely relies ings below and more can be found from Observation 1 to on LLMs for final predictions. These works fall under the Observation 19. category of LLMs-as-Predictors . The first series of work Finding 1. For LLMs-as-Enhancers , deep sentence focus on applying closed-source LLMs without tuning the embedding models present effectiveness in terms of parameters. GPT4Graph [18] evaluates the potential of performance and efficiency. We empirically find that LLMs in executing knowledge graph (KG) reasoning and whenweadoptdeepsentenceembeddingmodelsasenhancers node classification tasks. Their findings indicate that these at the feature level, they present good performance under models can deliver competitive results for short-range KG different dataset split settings, and also scalability. This reasoning but struggle with long-range KG reasoning and indicates that they are good candidates to enhance text at- node classification tasks. However, its presentation is pretty tributes at the feature level. vagueandtheydon’tgivethedetailedformatoftheprompt Finding 2. For LLMs-as-Enhancers ,thecombination they use. Considering the publicity of the Arxiv data, the of LLMs’ augmentations and ensembling demonst- dataleakageprobleminevaluationisfurtherstudiedin[25]. rates its effectiveness. As demonstrated in Section 4.2, NLGraph [65] introduces a synthetic benchmark to assess when LLMs are utilized as enhancers at the text level, we graph structure reasoning capabilities. The study primar- observe performance improvements by ensembling the aug- ily concentrates on traditional graph reasoning tasks such mentedattributeswiththeoriginalattributesacrossdatasets as shortest path, maximum flow, and bipartite matching, and data splits. This suggests a promising approach to while only offering limited analysis on node classification enhance the performance of attribute-related tasks. The tasks. This does not align with our central focus, primar- proposed pipeline involves augmenting the attributes with ily on graph learning, with a specific emphasis on node LLMs and subsequently ensembling the original attributes classification tasks. GraphText [84] further tries to apply with the augmented ones. LLMs to a broader range of non-text-attributed graphs by

mar- proposed pipeline involves augmenting the attributes with ily on graph learning, with a specific emphasis on node LLMs and subsequently ensembling the original attributes classification tasks. GraphText [84] further tries to apply with the augmented ones. LLMs to a broader range of non-text-attributed graphs by Finding 3. For LLMs-as-Predictors , LLMs present converting the original features into clustering centers or preliminary effectiveness but also indicate potential pseudo labels. LLM4Dyg [81] further evaluates LLMs’ ca- evaluation problem. In Section 5, we conduct prelimi- pabilityfortemporalgraph-relatedtasks. LLMGNN[4]and nary experiments on applying LLMs as predictors, utilizing GPT4GNAS [66] apply LLMs-as-predictors as annotators both textual attributes and edge relationships. The results and agents for neural architecture search, respectively. demonstrate that LLMs present effectiveness in processing As these closed-source LLMs only accept text-type inputs, textualattributesandachievinggoodzero-shotperformance the first type of methods requires transforming graphs into on certain datasets. Moreover, our analysis reveals two po- certain form of natural language, either directly using node tential problems within the existing evaluation framework: attributes or describing the graph structure using natural (1) There are instances where LLMs’ inaccurate predictions language. Meanwhile, due to the input length limitations canalsobeconsideredreasonable, particularlyinthecaseof of LLMs, this transformation process often results in the citation datasets where multiple labels may be appropriate. lossofaconsiderableamountofinformationfromthegraph. (2) We find a potential test data leakage problem on Ogbn- Therefore,thesecondtypeofworkinvolvesfine-tuningLLMs arxiv , which underscores the need for a careful reconsider- toenablethemtounderstandgraphinformationrepresented ation of how to appropriately evaluate the performance of as embeddings. InstructGLM [79] combines textual instruc- LLMs on real-world datasets. tions with node features in embedding form, enabling LLMs tounderstandnodefeaturesthroughinstructiontuning. Sub- 7.2 Limitations sequently, it predicts the type of nodes based on the given A deeper understanding of the effectiveness of text instructions. GraphGPT[62]furtherintroducescross-modal embeddings. Despite the effectiveness of deep sentenceembedding models, our understanding of why their embed- tion” 5 (2) the ground truth labels may present ambiguity, dingsoutperformPLMs’onnodeclassificationtasksremains

understanding of the effectiveness of text instructions. GraphGPT[62]furtherintroducescross-modal embeddings. Despite the effectiveness of deep sentenceembedding models, our understanding of why their embed- tion” 5 (2) the ground truth labels may present ambiguity, dingsoutperformPLMs’onnodeclassificationtasksremains and the performance calculated based on them may not re- limited. Furthermore, we observe a performance gap be- flect LLMs’ genuine capability. For the first problem, one tween deep sentence embedding models and GLEM on the possible mitigation is to use the latest dataset which is not Ogbn-products dataset, which may be related to the do- included in the training corpus of LLMs. However, that mains of the dataset. Moreover, as shown in Observation 4, meansweneedtokeepcollectingdataandannotatingthem, GNNs demonstrate different levels of effectiveness on differ- which seems not an effective solution. For the second prob- enttextembeddings. However, wegivelimitedexplanations lem, one possible solution is to reconsider the ground truth for this phenomenon. To gain a deeper understanding, we design. For instance, for the categorization of academic pa- need to have a look at the original feature space and the pers, we may adopt a multi-label setting and select all ap- feature space after aggregation. This phenomenon may po- plicable categories as the ground truth. However, for more tentiallyberelatedtotheanistrophyinlanguagemodelem- general tasks, it remains a challenge to design more rea- beddings [12]. More in-depth analysis is required to better sonable ground truths. Generally speaking, it’s a valuable understand these phenomena. future direction to rethink how to properly evaluate LLMs. Costs of LLM augmentations. In the work, we study AligningthefeaturespaceofgraphmodelsandLLMs. TAPE and KEA to enhance the textual attributes at the Currently, a major obstacle hindering the wider application text level. Although these methods have proven to be ef- of LLMs in the field of graph learning is the discrepancy be- fective, they require querying LLMs’ APIs at least N times tween the feature space of LLMs and that of graphs. This for a graph with N nodes. Given the cost associated with discrepancy makes it difficult for LLMs to effectively under- LLMs, this poses a significant expense when dealing with stand information in the graph domain. There are mainly large-scale datasets. Consequently, we have not presented two approaches to address this issue in current work. The results for the Ogbn-arxiv and Ogbn-products datasets. first approach is to

effectively under- LLMs, this poses a significant expense when dealing with stand information in the graph domain. There are mainly large-scale datasets. Consequently, we have not presented two approaches to address this issue in current work. The results for the Ogbn-arxiv and Ogbn-products datasets. first approach is to translate the information on the graph Text-formatted hand-crafted prompts to represent into natural language that LLMs can understand. The sec- graphs. In Section 5, we limit our study to the use of “nat- ond approach involves directly inputting the graph infor- ural language” prompts for graph representation. However, mation in the form of embeddings and then using instruc- variousotherformatsexistforrepresentinggraphsinnatural tion tuning to enable LLMs to understand this information. language such as XML, YAML, GML, and more [55]. More- However, both methods have their evident limitations. For over, wemainlydesignthesepromptsinahand-craftedway, the first method, the translation process can result in in- which is mainly based on trial and error. It’s thus worth- formation loss, and the inherent input length limitation of while to consider exploring more prompt formats and how LLMs also prevents users from inputting large-scale graphs. to come up with automatic prompts. For the second method, the introduction of tuning signifi- cantly increases computational overhead. Is there a better 7.3 FutureDirections way to align LLMs with graphs? A recent work targeting Extending the current pipelines to more tasks and multimodality [47] has shown new possibilities. It demon- more types of graphs. In this study, our primary fo- stratesthatwithfixedLLMparameters, onlyalineartrans- cus is on investigating the node classification task for text- formation layer is needed to convert information from the attributedgraphs. Nevertheless,itremainsunexploredwhe- visual domain into content that can be effectively processed ther these two pipelines can be extended to other graph- byLLMs,andsuchanarchitecturealsoholdsgreatpotential learning tasks or other types of graphs. Certain tasks neces- in the field of graph machine learning. sitatetheutilizationoflong-rangeinformation[11],andrep- resenting such information within LLMs’ limited input con- 8. REFERENCES text poses a significant challenge. Furthermore, we demon- [1]R. Anil, A. M. Dai, O. Firat, M. Johnson, D. Lep- strate that LLMs exhibit promising initial results in graphs ikhin, A. Passos, S. Shakeri, E. Taropa, P. Bailey, co

n[11],andrep- resenting such information within LLMs’ limited input con- 8. REFERENCES text poses a significant challenge. Furthermore, we demon- [1]R. Anil, A. M. Dai, O. Firat, M. Johnson, D. Lep- strate that LLMs exhibit promising initial results in graphs ikhin, A. Passos, S. Shakeri, E. Taropa, P. Bailey, containingabundanttextualinformation,particularlyinnat- Z. Chen, et al. Palm 2 technical report. arXiv preprint urallanguage. However,theexplorationoftheireffectiveex- arXiv:2305.10403 , 2023. tension to other types of graphs with non-natural language information, such as molecular graph [13; 30], still needs [2]S.Bubeck,V.Chandrasekaran,R.Eldan,J.A.Gehrke, further exploration. E. Horvitz, E. Kamar, P. Lee, Y. T. Lee, Y.-F. Li, Using LLMs more efficiently. Despite the effectiveness S. M. Lundberg, H. Nori, H. Palangi, M. T. Ribeiro, ofLLMs, theinherentoperationalefficiencyandoperational and Y. Zhang. Sparks of artificial general intelligence: cost of these models still pose significant challenges. Taking Early experiments with gpt-4. ArXiv , abs/2303.12712, ChatGPT, which is accessed through an API, as an exam- 2023. ple,thecurrentbillingmodelincurshighcostsforprocessing [3]Z. Chai, T. Zhang, L. Wu, K. Han, X. Hu, X. Huang, large-scalegraphs. Asforlocallydeployedopen-sourcelarge and Y. Yang. Graphllm: Boosting graph reason- models, even just using them for inference requires substan- ing ability of large language model. arXiv preprint tial hardware resources, not to mention training the models arXiv:2310.05845 , 2023. with parameter updates. Therefore, developing more effi- cient strategies to utilize LLMs is currently a challenge. [4]Z. Chen, H. Mao, H. Wen, H. Han, W. Jin, H. Zhang, EvaluatingLLMs’capabilityforgraphlearningtasks. H. Liu, and J. Tang. Label-free node classification In this paper, we briefly talk about the potential pitfalls of on graphs with large language models (llms). arXiv the current evaluation framework. There are mainly two preprint arXiv:2310.04668 , 2023. problems: (1)thetestdatamayalreadyappearinthetrain- ing corpus of LLMs, which is referred to as ”contamina- 5 https://hitz-zentroa.github.io/lm-contamination/ [5]W.-L. Chiang, X. Liu, S. Si, Y. Li, S. Bengio, and C.-J. [16]J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, Hsieh. Cluster-gcn: An efficient algorithm for training and G. E. Dahl. Neural message passing for quantum deep

rred to as ”contamina- 5 https://hitz-zentroa.github.io/lm-contamination/ [5]W.-L. Chiang, X. Liu, S. Si, Y. Li, S. Bengio, and C.-J. [16]J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, Hsieh. Cluster-gcn: An efficient algorithm for training and G. E. Dahl. Neural message passing for quantum deep and large graph convolutional networks. Proceed- chemistry. ArXiv , abs/1704.01212, 2017. ings of the 25th ACM SIGKDD International Confer- [17]S.Gui, X.Li, L.Wang, andS.Ji.GOOD:Agraphout- ence on Knowledge Discovery & Data Mining , 2019. of-distribution benchmark. In Thirty-sixth Conference [6]E. Chien, W.-C. Chang, C.-J. Hsieh, H.-F. Yu, on Neural Information Processing Systems Datasets J.Zhang,O.Milenkovic,andI.S.Dhillon.Nodefeature and Benchmarks Track , 2022. extraction by self-supervised multi-scale neighborhood prediction. In ICLR 2022 , 2022. [18]J. Guo, L. Du, and H. Liu. Gpt4graph: Can large lan- guage models understand graph structured data? an [7]A. Creswell, M. Shanahan, and I. Higgins. Selection- empirical evaluation and benchmarking. arXiv preprint inference: Exploiting large language models for inter- arXiv:2305.15066 , 2023. pretable logical reasoning. In The Eleventh Interna- [19]W. L. Hamilton, R. Ying, and J. Leskovec. Inductive tional Conference on Learning Representations , 2023. representation learning on large graphs. In NIPS , 2017. [8]E. Dai, C. Aggarwal, and S. Wang. Nrgnn: Learning a [20]Z. S. Harris. Distributional structure. Word , 10(2- label noise resistant graph neural network on sparsely 3):146–162, 1954. and noisily labeled graphs. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & [21]P. He, X. Liu, J. Gao, and W. Chen. Deberta: Data Mining , KDD ’21, page 227–236, New York, NY, Decoding-enhanced bert with disentangled attention. USA, 2021. Association for Computing Machinery. arXiv preprint arXiv:2006.03654 , 2020. [9]J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. [22]X. He, X. Bresson, T. Laurent, and B. Hooi. Explana- BERT: Pre-training of deep bidirectional transformers tionsasfeatures: Llm-basedfeaturesfortext-attributed for language understanding

arXiv preprint arXiv:2006.03654 , 2020. [9]J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. [22]X. He, X. Bresson, T. Laurent, and B. Hooi. Explana- BERT: Pre-training of deep bidirectional transformers tionsasfeatures: Llm-basedfeaturesfortext-attributed for language understanding. In Proceedings of the 2019 graphs. arXiv preprint arXiv:2305.19523 , 2023. Conference of the North American Chapter of the As- sociation for Computational Linguistics: Human Lan- [23]W. Hu, M. Fey, M. Zitnik, Y. Dong, H. Ren, B. Liu, guageTechnologies, Volume1(LongandShortPapers) , M. Catasta, and J. Leskovec. Open graph benchmark: pages 4171–4186, Minneapolis, Minnesota, June 2019. Datasets for machine learning on graphs. Advances in Association for Computational Linguistics. neuralinformationprocessingsystems ,33:22118–22133, [10]K. Duan, Q. Liu, T.-S. Chua, S. Yan, W. T. Ooi, 2020. Q. Xie, and J. He. Simteg: A frustratingly simple ap- [24]Z. Hu, Y. Dong, K. Wang, K.-W. Chang, and Y. Sun. proach improves textual graph learning. arXiv preprint Gpt-gnn: Generative pre-training of graph neural net- arXiv:2308.02565 , 2023. works. Proceedings of the 26th ACM SIGKDD Inter- [11]V. P. Dwivedi, L. Ramp´aˇsek, M. Galkin, A. Parviz, national Conference on Knowledge Discovery & Data G. Wolf, A. T. Luu, and D. Beaini. Long range graph Mining , 2020. benchmark. In Thirty-sixth Conference on Neural In- [25]J. Huang, X. Zhang, Q. Mei, and J. Ma. Can llms ef- formation Processing Systems Datasets and Bench- fectively leverage graph structural information: When marks Track , 2022. and why. arXiv preprint arXiv:2309.16595 , 2023. [12]K.Ethayarajh.Howcontextualarecontextualizedword [26]Y. Ji, Y. Gong, Y. Peng, C. Ni, P. Sun, D. Pan, B. Ma, representations? Comparing the geometry of BERT, and X. Li. Exploring chatgpt’s ability to rank content: ELMo, and GPT-2 embeddings. In Proceedings of A preliminary study on consistency with human pref- the 2019 Conference on Empirical Methods in Natu- erences. ArXiv , abs/2303.07610, 2023. ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP- [27]T. N. Kipf and M. Welling. Semi-supervised classifi- IJCNLP)

A preliminary study on consistency with human pref- the 2019 Conference on Empirical Methods in Natu- erences. ArXiv , abs/2303.07610, 2023. ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP- [27]T. N. Kipf and M. Welling. Semi-supervised classifi- IJCNLP) , pages 55–65, Hong Kong, China, Nov. 2019. cation with graph convolutional networks. In Interna- Association for Computational Linguistics. tional Conference on Learning Representations , 2017. [13]M. Fey and J. E. Lenssen. Fast graph repre- [28]G. Li, M. M¨uller, B. Ghanem, and V. Koltun. Train- sentation learning with pytorch geometric. ArXiv , ing graph neural networks with 1000 layers. In Inter- abs/1903.02428, 2019. national conference on machine learning , pages 6437– 6449. PMLR, 2021. [14]Y. Gao, T. Sheng, Y. Xiang, Y. Xiong, H. Wang, and J. Zhang. Chat-rec: Towards interactive and ex- [29]H. Li, X. Wang, Z. Zhang, and W. Zhu. Out-of- plainablellms-augmentedrecommendersystem. ArXiv , distribution generalization on graphs: A survey. arXiv abs/2303.14524, 2023. preprint arXiv:2202.07987 , 2022. [15]C. L. Giles, K. D. Bollacker, and S. Lawrence. Citeseer: [30]J. Li, Y. Liu, W. Fan, X. Wei, H. Liu, J. Tang, and An automatic citation indexing system. In Proceedings Q. Li. Empowering molecule discovery for molecule- of the Third ACM Conference on Digital Libraries , DL captiontranslationwithlargelanguagemodels: Achat- ´98, pages 89–98, New York, NY, USA, 1998. ACM. gpt perspective. ArXiv , abs/2306.06615, 2023.[31]Q. Li, X. Li, L. Chen, and D. Wu. Distilling knowl- [44]A.Neelakantan,T.Xu,R.Puri,A.Radford,J.M.Han, edge on text graph for social media attribute inference. J. Tworek, Q. Yuan, N. A. Tezak, J. W. Kim, C. Hal- In Proceedings of the 45th International ACM SIGIR lacy, J. Heidecke, P. Shyam, B. Power, T. E. Nekoul, Conference on Research and Development in Informa- G.Sastry, G.Krueger, D.P.Schnurr, F.P.Such, K.S.- tion Retrieval , SIGIR ’22, page 2024–2028, New York, K.Hsu,M.Thompson,T.Khan,T.Sherbakov,J.Jang, NY,USA,2022.AssociationforComputingMachinery. P. Welinder, and L. Weng. Text and

onference on Research and Development in Informa- G.Sastry, G.Krueger, D.P.Schnurr, F.P.Such, K.S.- tion Retrieval , SIGIR ’22, page 2024–2028, New York, K.Hsu,M.Thompson,T.Khan,T.Sherbakov,J.Jang, NY,USA,2022.AssociationforComputingMachinery. P. Welinder, and L. Weng. Text and code embeddings [32]Y.Li, J.Yin, andL.Chen.Informativepseudo-labeling by contrastive pre-training. ArXiv , abs/2201.10005, forgraphneuralnetworkswithfewlabels. Data Mining 2022. and Knowledge Discovery , 37(1):228–254, 2023. [45]OpenAI. Introducing chatgpt, 2022. [33]H. Liu, J. Feng, L. Kong, N. Liang, D. Tao, Y. Chen, [46]OpenAI. Gpt-4 technical report. ArXiv , and M. Zhang. One for all: Towards training one abs/2303.08774, 2023. graph model for all classification tasks. arXiv preprint arXiv:2310.00149 , 2023. [47]Z.Pang,Z.Xie,Y.Man,andY.-X.Wang.Frozentrans- [34]H.Liu, B.Hu, X.Wang, C.Shi, Z.Zhang, andJ.Zhou. formers in language models are effective visual encoder Confidence may cheat: Self-training on graph neu- layers. arXiv preprint arXiv:2310.12973 , 2023. ral networks under distribution shift. In Proceedings [48]F. Petroni, T. Rockt¨aschel, P. Lewis, A. Bakhtin, of the ACM Web Conference 2022 , WWW ’22, page Y. Wu, A. H. Miller, and S. Riedel. Language models 1248–1258, New York, NY, USA, 2022. Association for as knowledge bases? ArXiv , abs/1909.01066, 2019. Computing Machinery. [49]S. Purchase, A. Zhao, and R. D. Mullins. Revis- [35]J. Liu, C. Liu, R. Lv, K. Zhou, and Y. Zhang. Is chat- iting embeddings for graph neural networks. ArXiv , gpt a good recommender? a preliminary study. arXiv abs/2209.09338, 2022. preprint arXiv:2304.10149 , 2023. [36]W. Liu, P. Zhou, Z. Zhao, Z. Wang, Q. Ju, H. Deng, [50]Y. Qin, X. Wang, Z. Zhang, and W. Zhu. Disen- andP.Wang.K-bert: Enablinglanguagerepresentation tangled representation learning with large language with knowledge graph. In AAAI Conference on Artifi- models for text-attributed graphs. arXiv preprint cial Intelligence , 2019. arXiv:2310.18152 , 2023. [37]Z. Liu, C. Xiong, M. Sun, and Z. Liu. Fine-grained [51]X.Qiu, T.Sun, Y.Xu,

with knowledge graph. In AAAI Conference on Artifi- models for text-attributed graphs. arXiv preprint cial Intelligence , 2019. arXiv:2310.18152 , 2023. [37]Z. Liu, C. Xiong, M. Sun, and Z. Liu. Fine-grained [51]X.Qiu, T.Sun, Y.Xu, Y.Shao, N.Dai, andX.Huang. factverificationwithkernelgraphattentionnetwork.In Pre-trained models for natural language processing: A Proceedings of the 58th Annual Meeting of the Associ- survey. Science China Technological Sciences , 63:1872 ation for Computational Linguistics , pages 7342–7351, – 1897, 2020. Online, July 2020. Association for Computational Lin- [52]A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, guistics. and I. Sutskever. Language models are unsupervised [38]Y. Ma and J. Tang. Deep Learning on Graphs . Cam- multitask learners. 2019. bridge University Press, 2021. [53]C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, [39]H. Mao, Z. Chen, W. Jin, H. Han, Y. Ma, T. Zhao, M. Matena, Y. Zhou, W. Li, and P. J. Liu. Explor- N.Shah,andJ.Tang.Demystifyingstructuraldisparity ing the limits of transfer learning with a unified text- in graph neural networks: Can one size fit all? arXiv to-text transformer. Journal of Machine Learning Re- preprint arXiv:2306.01323 , 2023. search , 21(140):1–67, 2020. [40]A. McCallum, K. Nigam, J. D. M. Rennie, and K. Sey- [54]N.ReimersandI.Gurevych.Sentence-BERT:Sentence more. Automating the construction of internet portals embeddings using Siamese BERT-networks. In Pro- with machine learning. Information Retrieval , 3:127– ceedings of the 2019 Conference on Empirical Methods 163, 2000. in Natural Language Processing and the 9th Interna- tional Joint Conference on Natural Language Process- [41]A. Miaschi and F. Dell’Orletta. Contextual and non- ing (EMNLP-IJCNLP) , pages 3982–3992, Hong Kong, contextual word embeddings: an in-depth linguistic in- China, Nov. 2019. Association for Computational Lin- vestigation.In Proceedingsofthe5thWorkshoponRep- guistics. resentation Learning for NLP , pages 110–119, Online, [55]M. Roughan

P) , pages 3982–3992, Hong Kong, contextual word embeddings: an in-depth linguistic in- China, Nov. 2019. Association for Computational Lin- vestigation.In Proceedingsofthe5thWorkshoponRep- guistics. resentation Learning for NLP , pages 110–119, Online, [55]M. Roughan and S. J. Tuke. Unravelling graph- July 2020. Association for Computational Linguistics. exchange file formats. ArXiv , abs/1503.02781, 2015. [42]T. Mikolov, K. Chen, G. Corrado, and J. Dean. Effi- cientestimationofwordrepresentationsinvectorspace. [56]T. Schick, J. Dwivedi-Yu, R. Dess`ı, R. Raileanu, arXiv preprint arXiv:1301.3781 , 2013. M. Lomeli, L. Zettlemoyer, N. Cancedda, and [43]N. Muennighoff, N. Tazi, L. Magne, and N. Reimers. T. Scialom. Toolformer: Language models can MTEB: Massive text embedding benchmark. In Pro- teach themselves to use tools. arXiv preprint ceedings of the 17th Conference of the European Chap- arXiv:2302.04761 , 2023. ter of the Association for Computational Linguistics , [57]P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Galligher, pages2014–2037, Dubrovnik, Croatia, May2023.Asso- and T. Eliassi-Rad. Collective classification in network ciation for Computational Linguistics. data. AI Magazine , 29(3):93, Sep. 2008.[58]C. Sun, H. Gu, and J. Hu. Scalable and adaptive [71]T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. De- graph neural networks with self-label-enhanced train- langue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Fun- ing. arXiv preprint arXiv:2104.09376 , 2021. towicz, and J. Brew. Huggingface’s transformers: State-of-the-art natural language processing. ArXiv , [59]T.Sun,Y.Shao,H.Qian,X.Huang,andX.Qiu.Black- abs/1910.03771, 2019. boxtuningforlanguage-model-as-a-service.In Interna- tional Conference on Machine Learning , pages 20841– [72]Y. Wu, Y. Xu, A. Singh, Y. Yang, and A. W. 20855. PMLR, 2022. Dubrawski. Active learning for graph neural networks via node feature propagation. ArXiv , abs/1910.07567, [60]X. Sun, X. Li, J. Li, F. Wu, S. Guo, T. Zhang, and 2019. G. Wang. Text classification via large language models. ArXiv , abs/2305.08377, 202

Dubrawski. Active learning for graph neural networks via node feature propagation. ArXiv , abs/1910.07567, [60]X. Sun, X. Li, J. Li, F. Wu, S. Guo, T. Zhang, and 2019. G. Wang. Text classification via large language models. ArXiv , abs/2305.08377, 2023. [73]F. Xia, K. Sun, S. Yu, A. Aziz, L. Wan, S. Pan, and H. Liu. Graph learning: A survey. IEEE Transactions [61]Y. Sun, S. Wang, Y. Li, S. Feng, X. Chen, H. Zhang, on Artificial Intelligence , 2:109–127, 2021. X.Tian,D.Zhu,H.Tian,andH.Wu.Ernie: Enhanced [74]J. Yang, Z. Liu, S. Xiao, C. Li, D. Lian, S. Agrawal, representation through knowledge integration. ArXiv , abs/1904.09223, 2019. A. S, G. Sun, and X. Xie. Graphformers: GNN- nested transformers for representation learning on tex- [62]J. Tang, Y. Yang, W. Wei, L. Shi, L. Su, S. Cheng, tual graph. In A. Beygelzimer, Y. Dauphin, P. Liang, D. Yin, and C. Huang. Graphgpt: Graph instruc- and J. W. Vaughan, editors, Advances in Neural Infor- tion tuning for large language models. arXiv preprint mation Processing Systems , 2021. arXiv:2310.13023 , 2023. [75]Z. Yang, W. W. Cohen, and R. Salakhutdinov. Revis- [63]H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. iting semi-supervised learning with graph embeddings. Lachaux, T. Lacroix, B. Rozi`ere, N. Goyal, E. Ham- ArXiv , abs/1603.08861, 2016. bro, F. Azhar, et al. Llama: Open and efficient founda- [76]L. Yao, C. Mao, and Y. Luo. Graph convolutional net- tionlanguagemodels. arXivpreprintarXiv:2302.13971 , works for text classification. ArXiv , abs/1809.05679, 2023. 2018. [64]P. Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, [77]M. Yasunaga, A. Bosselut, H. Ren, X. Zhang, C. D. P. Li`o, and Y. Bengio. Graph attention networks. In Manning, P. Liang, and J. Leskovec. Deep bidirectional International Conference on Learning Representations , language-knowledgegraphpretraining.In Neural Infor- 2018. mation Processing Systems (NeurIPS) , 2022. [65]H. Wang, S. Feng, T.

Manning, P. Liang, and J. Leskovec. Deep bidirectional International Conference on Learning Representations , language-knowledgegraphpretraining.In Neural Infor- 2018. mation Processing Systems (NeurIPS) , 2022. [65]H. Wang, S. Feng, T. He, Z. Tan, X. Han, [78]M.Yasunaga,J.Leskovec,andP.Liang.Linkbert: Pre- and Y. Tsvetkov. Can language models solve graph training language models with document links. In Pro- problems in natural language? arXiv preprint ceedings of the 60th Annual Meeting of the Associa- arXiv:2305.10037 , 2023. tion for Computational Linguistics (Volume 1: Long [66]H. Wang, Y. Gao, X. Zheng, P. Zhang, H. Chen, and Papers) , pages 8003–8016, 2022. J. Bu. Graph neural architecture search with gpt-4. [79]R. Ye, C. Zhang, R. Wang, S. Xu, and Y. Zhang. arXiv preprint arXiv:2310.01436 , 2023. Natural language is all a graph needs. arXiv preprint [67]J. Wang, X. Hu, W. Hou, H. Chen, R. Zheng, arXiv:2308.07134 , 2023. Y. Wang, L. Yang, H. Huang, W. Ye, X. Geng, [80]J. Zhang. Graph-toolformer: To empower llms with et al. On the robustness of chatgpt: An adversar- graph reasoning ability via prompt augmented by chat- ial and out-of-distribution perspective. arXiv preprint gpt. arXiv preprint arXiv:2304.11116 , 2023. arXiv:2302.12095 , 2023. [68]L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, [81]Z. Zhang, X. Wang, Z. Zhang, H. Li, Y. Qin, S. Wu, D. Jiang, R. Majumder, and F. Wei. Text embeddings and W. Zhu. Llm4dyg: Can large language models by weakly-supervised contrastive pre-training. arXiv solve problems on dynamic graphs? arXiv preprint preprint arXiv:2212.03533 , 2022. arXiv:2310.17110 , 2023. [69]M. Wang, L. Yu, D. Zheng, Q. Gan, Y. Gai, Z. Ye, [82]Z.Zhang,A.Zhang,M.Li,andA.J.Smola.Automatic M. Li, J. Zhou, Q. Huang, C. Ma, Z. Huang, Q. Guo, chain of thought prompting in large language models. H. Zhang, H. Lin, J. J. Zhao, J. Li, A. Smola, ArXiv , abs/2210.03493, 2022. and Z. Zhang. Deep graph library: Towards effi- [83]J.Zhao,M.Qu,C.Li,H.Yan,Q.Liu,R.Li,X.Xie,and cient and scalable deep learning on graphs. ArXiv , J

f thought prompting in large language models. H. Zhang, H. Lin, J. J. Zhao, J. Li, A. Smola, ArXiv , abs/2210.03493, 2022. and Z. Zhang. Deep graph library: Towards effi- [83]J.Zhao,M.Qu,C.Li,H.Yan,Q.Liu,R.Li,X.Xie,and cient and scalable deep learning on graphs. ArXiv , J. Tang. Learning on large-scale text-attributed graphs abs/1909.01315, 2019. via variational inference. In The Eleventh International [70]J. Wei, X. Wang, D. Schuurmans, M. Bosma, E. Chi, Conference on Learning Representations , 2023. Q. Le, and D. Zhou. Chain of thought prompting elic- [84]J.Zhao,L.Zhuo,Y.Shen,M.Qu,K.Liu,M.Bronstein, its reasoning in large language models. arXiv preprint Z. Zhu, and J. Tang. Graphtext: Graph reasoning in arXiv:2201.11903 , 2022. text space. arXiv preprint arXiv:2310.01089 , 2023. [85]W.X.Zhao,K.Zhou,J.Li,T.Tang,X.Wang,Y.Hou, Ogbn-arxiv and Ogbn-products [23] These dataset are Y. Min, B. Zhang, J. Zhang, Z. Dong, Y. Du, C. Yang, selected from the popular OGB benchmark [23], and de- Y. Chen, Z. Chen, J. Jiang, R. Ren, Y. Li, X. Tang, scriptions for these datasets can be found in https://ogb. Z. Liu, P. Liu, J. Nie, and J. rong Wen. A survey of stanford.edu/docs/nodeprop . large language models. ArXiv , abs/2303.18223, 2023. [86]J. Zhu, Y. Cui, Y. Liu, H. Sun, X. Li, M. Pelger, B. EXPERIMENTSETUPS L. Zhang, T. Yan, R. Zhang, and H. Zhao. Textgnn: B.1 ComputingEnvironment Improving text encoder via graph neural network in WeimplementallthebaselinemodelswithPyG[13],DGL[69], sponsored search. Proceedings of the Web Conference and transformers [71] modules. The experiments were con- 2021 , 2021. ducted in a GPU server with eight NVIDIA RTX A5000 GPUs, each with 24GB VRAM. [87]J. Zhu, Y. Yan, L. Zhao, M. Heimann, L. Akoglu, and D. Koutra. Beyond homophily in graph neural net- B.2 Hyperparameters works: Current limitations and effective designs. In For RevGAT, GraphSage, and SAGN models, we directly H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and adopttheb

Akoglu, and D. Koutra. Beyond homophily in graph neural net- B.2 Hyperparameters works: Current limitations and effective designs. In For RevGAT, GraphSage, and SAGN models, we directly H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and adoptthebesthyperparametersfromtheOGBleaderboard 7. H. Lin, editors, Advances in Neural Information Pro- For Deberta-base on Cora and Pubmed , we follow the hy- cessing Systems , volume 33, pages 7793–7804. Curran perparameter setting of TAPE [22]. In terms of GLEM, for Associates, Inc., 2020. the LM part, we follow the hyperparameter setting in their reporsitory 8. For GCN, GAT, MLP, we use the following APPENDIX hyperparameter search range. (a) Hidden dimension: {8, 16, 32, 64, 128, 256 }. A. DATASETS (b) Number of layers: {1, 2, 3 } In this work, we mainly use the following five real-world (c) Normalization: {None, BatchNorm }; graph datasets. Their statistics are shown in Table 19. (d) Learning rate: {1e-2, 5e-2, 5e-3, 1e-3 } (e) Weight Decay: {1e-5, 5e-5, 5e-4, 0 } Table 19: Statistics of the graph datasets. (f) Dropout: {0., 0.1, 0.5, 0.8 } (g) Number of heads for GAT: {1, 4, 8 } C. DEMONSTRATIONSOFTAPE ExamplesforPubmed Afteranalyzingthe Pubmed data- set, we find an interesting phenomenon that sometimes the label of the paper just appears in the raw text attributes. AnexampleisshowninTable20. Thispropertyof Pubmed may be

label of the paper just appears in the raw text attributes. AnexampleisshowninTable20. Thispropertyof Pubmed may be related to the superior zero-shot performance of A.1 DatasetDescription LLMs on this dataset. This can also potentially explain In this part, we give a brief introduction to each graph why GCN and GAT are outperformed by MLP in the high dataset. It should be noted that it’s cumbersome to get the labeling ratio. When the link between node attributes and raw text attributes for some datasets, and we will elaborate node labels can be easily found and adequate to determine them below. The structural information and label informa- the categories, incorporating neighbors coming from other tion of these datasets can be achieved from Pyg 6. We will categories will introduce noise. also release the pre-processed versions of these datasets to Table 20: An illustrative example for Pubmed assist future related studies. Cora[40] Cora isapapercitationdatasetwiththefollow- ing seven categories: [’Rule Learning’, ’Neural Networks’, Title: Predictive power of sequential measures of albumin- ’Case Based’, ’Genetic Algorithms’, ’Theory’, ’Reinforce- uria for progression to ESRD or death in Pima Indians with ment Learning’, ’Probabilistic Methods’]. The raw text at- type 2 diabetes . tributes can be obtained from https://people.cs.umass. ... (content omitted here) edu/ ~ mccallum/data.html . Ground truth label: Diabetes Mellitus Type 2 Citeseer[15] Citeseer isapapercitationdatasetwiththe following seven categories: [”Agents”, ”ML”, ”IR”, ”DB”, ”HCI”, ”AI”]. Note that we find that the TAG versopm onlycontainsthetextattributesfor3186nodes. Asaresult, we take the graph consisted of these 3186 nodes with 4277 edges. Pubmed [57] Pubmed is a paper citation dataset consist- ing scientific journals collected from the PubMed database with the following three categories: [’Diabetes Mellitus, Ex- perimental’, ’Diabetes Mellitus Type 1’, ’Diabetes Mellitus Type 2’]. 6 https://pytorch-geometric.readthedocs.io/en/ 7 https://github.com/snap-stanford/ogb latest/modules/data.html 8 htt

with the following three categories: [’Diabetes Mellitus, Ex- perimental’, ’Diabetes Mellitus Type 1’, ’Diabetes Mellitus Type 2’]. 6 https://pytorch-geometric.readthedocs.io/en/ 7 https://github.com/snap-stanford/ogb latest/modules/data.html 8 https://github.com/AndyJZhao/GLEM Dataset #Nodes #Edges Task Metric Cora [40] 2,708 5,429 7-classclassif. Accuracy Citeseer * [15] 3,186 4,277 6-classclassif. Accuracy Pubmed [57] 19,717 44,338 3-classclassif. Accuracy Ogbn-arxiv [23] 169,343 1,166,243 40-classclassif. Accuracy Ogbn-products [23] 2,449,029 61,859,140 47-classclassif. Accuracy

SELF-DISCOVER: LargeLanguageModelsSelf-ComposeReasoningStructures Pei Zhou1 Jay Pujara1 Xiang Ren1 Xinyun Chen2 Heng-Tze Cheng2 Quoc V. Le2 Ed H. Chi2 Denny Zhou2 Swaroop Mishra2 Huaixiu Steven Zheng2 Abstract son. Forexample, few-shot andzero-shot chain-of-thought We introduce SELF-DISCOVER, a general frame- (CoT)(Nyeetal.,2021;Weietal.,2022;Kojimaetal.,2022; workforLLMstoself-discoverthetask-intrinsic Yasunaga et al.,2023) resembles how humans solve prob- reasoningstructurestotacklecomplexreasoning lemsstep-by-step,decomposition-basedprompting(Zhou problems that are challenging for typical prompt- et al.,2022a;Drozdov et al.,2022;Patel et al.,2022;Hao ing methods. Core to the framework is a self- et al.,2023;Khot et al.,2022) is inspired by how humans discovery process where LLMs select multiple breakdown a complex problem into a series of smaller atomic reasoning modules such as critical think- subproblems, and then solve those subproblems one by ingandstep-by-stepthinking,andcompose them one (Polya,2004), and step-back prompting (Zheng et al., into an explicit reasoning structure for LLMs to 2023) is motivated by how humans reflect on task nature follow during decoding. SELF-DISCOVER sub- toderivegeneralprinciples. However,afundamentallimi- stantially improves GPT-4 and PaLM 2’s per- tation is that each technique itself serves as an atomic rea- formance onchallenging reasoningbenchmarks soningmodulemaking animplicitpriorassumptionofthe suchasBigBench-Hard,groundedagentreason- process on how to tackle a given task. Instead, we argue ing, and MATH, by as much as 32% compared that each task has a unique intrinsic structure underlying toChain ofThought (CoT).Furthermore, SELF- thereasoning processinvolvedin solving itefficiently. For DISCOVER outperformsinference-intensivemeth- instance,least-to-mostprompting(Zhouetal.,2022a;Droz- ods suchas CoT-Self-Consistency by morethan dov et al.,2022) has shownto be much more effective than 20%,whilerequiring10-40xfewerinferencecom- CoT (Wei et al.,2022) at solving tasks such as symbolic pute. Finally, we show that the self-discovered manipulationand compositionalgeneralization, duetothe reasoning structures are universally applicable decomposition structure of the tasks. acrossmodelfamilies: from PaLM 2-LtoGPT-4, Thispaperaimsatself-discoveringtheunderlyingreasoning and from GPT-4 to Llama2, and share commonal- structureuniquetoeachtask,whilebeinghighlyefficientin ities with human reasoning patterns. termsof computation. Ourapproach, SELF-DISCOVER,is inspiredbyhowhumansinternallydeviseareasoningpro- gramforproblem-solving(Newelletal.,1958;Rasmussen, 1. Introduction 1983), as illustrated in Figure2. From a set of atomic reasoning modules described in natural language such as LargeLanguageModels(LLM)(Brownetal.,2020;Chowd- “breakdownintosubtasks”and“ criticalthinking”,anLLM, hery et al.,2022;OpenAI,2023b;Anil et al.,2023) pow- and task examples without labels, SELF-DISCOVER com- eredby transformers(Vaswanietal.,2017)have produced

1983), as illustrated in Figure2. From a set of atomic reasoning modules described in natural language such as LargeLanguageModels(LLM)(Brownetal.,2020;Chowd- “breakdownintosubtasks”and“ criticalthinking”,anLLM, hery et al.,2022;OpenAI,2023b;Anil et al.,2023) pow- and task examples without labels, SELF-DISCOVER com- eredby transformers(Vaswanietal.,2017)have produced poses a coherent reasoning structure intrinsic to the task impressivebreakthroughsingeneratingcoherenttexts(Ope- (Stage 1) and then solves instances of the task using the nAI,2022), andfollowinginstructions (Zhongetal.,2021; discovered structure (Stage 2). Stage 1 operates at the task- Mishra et al.,2022c;Wei et al.,2021;Chung et al.,2022; level and uses three actions to guide the LLM to generate Ouyang et al.,2022). In pursuit of the goal to enhance a reasoning structure for the task. At Stage 2, during the LLMs’ capability to reason and solve complex problems, finaldecoding,theLLM simplyfollows the self-discovered various prompting methods have been proposed, drawing structure to arrive at the final answer. inspirations from cognitive theories of how humans rea- Solving problems using SELF-DISCOVER brings several 1University of Southern California 2Google DeepMind. Cor- benefits compared to other methods for LLM reasoning. respondence to: Pei Zhou <peiz@usc.edu>, Swaroop Mishra First, the discovered reasoning structure is grounded in <swaroopmishra@google.com>,HuaixiuStevenZheng<steven- atomicreasoning modulesbenefiting fromthestrengths of zheng@google.com>. multiple reasoning modulesin contrast to applyinga priori Preprint. modulesuchasCoT.Second, SELF-DISCOVER isefficient 1 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure1. SELF-DISCOVER guides LLMs to self-discover and compose atomic reasoning modules into a reasoning structure to solve challengingtasks. ThroughtestingonchallengingreasoningbenchmarksincudingBigBench-Hard(BBH),agentreasoning(T4D),and MATH, we find thatSELF-DISCOVER outperforms Direct Answering on 23/25 and CoT on 21/25 tasks in zero-shot setting using PaLM 2-L. Full BBH results are in AppendixCTable3. incomputationasitonlyrequires3moreinferencestepson computationerrors(e.g. math). Wealsotakeacloserlook the task-level, while being more performant than inference- at the self-discovered reasoning structures, and show the heavyensembleapproachessuchasself-consistency(Wang universality of them by transferability study from PaLM et al.,2022). Lastly, the discovered reasoning structure 2-L to GPT-4, and from GPT-4 to Llama-2-70B. We hope is intrinsic to the task, and conveys LLMs’ insights about toencouragemorefutureworkonstructuredreasoningfor the task in a more interpretable way than the optimized solving challenging problems using LLMs. prompts (Zhou et al.,2022b;Yang et al.,2023). We test SELF-DISCOVER on 25 challenging reasoning 2. Self-Discovering Reasoning Structures for tasks including Big Bench-Hard (BBH) (Suzgun et al., Problem-Solving 2022), Thinking for Doing (T4D) (Zhou et al.,2023) and We take inspiration from how humans use prior knowledge MATH (Hendrycks et al.,2021). SELF-DISCOVER outper- and skills to devise a reasoning program to solve prob- formsCoTon21/25taskwithperformancegainsupto42% lems (Newell et al.,1958;Rasmussen,1983). When we (Figure1),highlighti

tructures for tasks including Big Bench-Hard (BBH) (Suzgun et al., Problem-Solving 2022), Thinking for Doing (T4D) (Zhou et al.,2023) and We take inspiration from how humans use prior knowledge MATH (Hendrycks et al.,2021). SELF-DISCOVER outper- and skills to devise a reasoning program to solve prob- formsCoTon21/25taskwithperformancegainsupto42% lems (Newell et al.,1958;Rasmussen,1983). When we (Figure1),highlightingtheadvantageoftheself-discovered face a new problem, we often first search internally what reasoning structure composed from the atomic reasoning knowledge and skills from our prior experience might be modulesagainstasingleaprioriCoTmodule. Furthermore, helpful to solve it. Then we will attempt to apply relevant we demonstrate that SELF-DISCOVER achieves superior knowledge andskills tothis task. And finallywe willcon- performanceagainstinference-heavymethodssuchasCoT nect multiple individual skills and knowledge to solve the + Self-Consistency and majority voting of every module problem. Wedesign SELF-DISCOVER toenactthesesteps while requiring 10-40x fewer inference compute (Figure5). into two stages as illustrated in Figure2. Finally, we compare SELF-DISCOVER with prompts op- timized (OPRO) using a training set (Yang et al.,2023) Given a task and a set of reasoning module descriptions (Figure9). We find that SELF-DISCOVER still performs on representinghigh-levelproblem-solvingheuristicssuchas parorbetterthanOPROwhiletheself-discoveredreasoning “Usecriticalthinking”and“ Let’sthinkstepbystep”,Stage1 structure are much more interpretable. of SELF-DISCOVER aimsto uncover the intrinsicreasoning Weconduct a set ofanalysis to understand the effectiveness structurefor solvingthistaskvia meta-reasoning. Specifi- of SELF-DISCOVER. BybreakingdownBBHtasksinto4 cally, we uses three meta-prompts to guide LLMs to select, differentcategories,wefindthatSELF-DISCOVERperforms adapt,andimplementanactionablereasoningstructurewith best on tasks requiring world knowledge and has a mod- no labels or training required. We format the structure in erate performance boost on algorithmic tasks compared to key-value pairs similar to JSON due to interpretability and CoT (Figure4). This is further confirmed by the error anal- findingsonfollowingJSONboostsreasoningandgeneration ysis on MATH, where 74.7% model failures comes from quality(Zhouetal.,2023;OpenAI,2023a). Thestructureof 2 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure2.Illustration of using SELF-DISCOVER for problem-solving. Given a generative LM, task, and seed reasoning module descriptions, we guide LMs to generate a reasoning structure in key-value format to solve the task. Finally, models can follow the self-discovered structures to solve the every instance from the task by filling in the values in JSON step-by-step. the meta-prompts andfull promptsare shownin Appendix. taskathand. Forexample,from“breaktheproblemintosub- Stage1operatesontask-level,meaningweonlyneedtorun problems”to “ calculateeacharithmeticoperationin order” SELF-DISCOVER once foreach task. Then, in Stage2, we for arithmetic problems. Given selected reasoning module can simply usethe discovered reasoningstructure tosolve subsetD S fromthepreviousstep,ADAPTrephraseseach every instance of the given task by instructing models to oftheselectedmoduletobemorespecifictothetask. Sim- followtheprovidedstructurebyfillingeachkeyandar

problems”to “ calculateeacharithmeticoperationin order” SELF-DISCOVER once foreach task. Then, in Stage2, we for arithmetic problems. Given selected reasoning module can simply usethe discovered reasoningstructure tosolve subsetD S fromthepreviousstep,ADAPTrephraseseach every instance of the given task by instructing models to oftheselectedmoduletobemorespecifictothetask. Sim- followtheprovidedstructurebyfillingeachkeyandarrive ilarly to SELECT, this stage uses a meta-prompt pA and at a final answer. a generative modelM to generate the adapted reasoning module descriptionsD A : 2.1. Stage 1: Self-Discover Task-Specific Structures D A = M (pA ∥ D S ∥ ti). (2) The first stage consists of three actions: 1) SELECT, where IMPLEMENT Finally, given the adapted reasoning mod- relevantreasoningmodulesfortask-solvingarechosenfrom uledescriptionsD A , SELF-DISCOVER operationalizesthe theset ofreasoningmodule descriptions;2) ADAPT, where reasoning modules into an implemented reasoning struc- descriptionsofselectedreasoningmodulesarerephrasedto tureD I with specified instruction on what to generate for be more specific to the task at hand; and 3) IMPLEMENT, each step. In additionto a meta promptpI, IMPLEMENT where the adapted reasoningdescriptions areimplemented also provides a demonstration of a human-written reason- into a structured actionable plan so that the task can be ing structureS human on another task to better convert the solved by following the structure. natural language descriptions into a reasoning structure: SELECT First, not every reasoning module is helpful for D I = M (pA ∥ S human ∥ D A ∥ ti). (3) every task, so the first stage of SELF-DISCOVER guides modelto selectmodulesthat areusefulbased ontaskexam- 2.2. Stage 2: Tackle Tasks Using Discovered Structures ples. Forexample,“reflectivethinking”mighthelpsearch Afterthe threestages, wehave animplemented reasoning forfirst-principletheoriesonscienceproblems,while“cre- structureD I uniquely adapted for the task we need tosolve ativethinking”helpsongeneratinganovelcontinuationto T . Then we can simply append the reasoning structure to a story. Given raw set of reasoning module descriptions all instances of the task and prompt models to follow the D suchas“criticalthinking”,and“ breaktheprobleminto reasoning structure to generate an answerA : sub-problems” (full set in AppendixA), and a few task ex- ampleswithoutlabelsti ∈ T , SELF-DISCOVER firstselects A = M (D S ∥ t),∀t∈ T. (4) asubsetofreasoningmodulesD S thatareusefulforsolving More details of prompts are included in AppendixA. the tasks by using a modelM and a meta-promptpS : D S = M (pS ∥ D ∥ ti). (1) 3. Experiment Setup ADAPT Sinceeachreasoningmoduleprovidesageneral 3.1. Tasks descriptionofhowtosolveproblems,thenextstepof SELF- We focus on diverse reasoning benchmarks that are still DISCOVER aims at tailoring each selected module to the challenging for LLMs: BIG-Bench Hard (BBH) (Suzgun 3 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure3.Illustrationofthreeactionsof SELF-DISCOVER. WeuseLMstocomposeacoherentreasoni

descriptionofhowtosolveproblems,thenextstepof SELF- We focus on diverse reasoning benchmarks that are still DISCOVER aims at tailoring each selected module to the challenging for LLMs: BIG-Bench Hard (BBH) (Suzgun 3 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure3.Illustrationofthreeactionsof SELF-DISCOVER. WeuseLMstocomposeacoherentreasoningstructurebyselectingrelevant modules, adapting to task-specific descriptions, and implement a reasoning structure in JSON. etal.,2022)contains23carefully-selectedchallengingtasks • DirectPrompting,wheremodeldirectlygeneratesthe fromBIG-Bench(Srivastavaetal.,2023). BBHtaskscover answer without intermediate reasoning steps. adiverserangeofreasoningproblemsspanningthefollow- • CoT (Wei et al.,2022;Kojima et al.,2022), where ing 4 categories according to their authors: 1) Algorithmic models are prompted to generate a reasoning process and Multi-Step Arithmetic Reasoning, 2) Natural Language leading to the final answer. Understanding, 3) Use of World Knowledge, and 4) Mul- tilingual Knowledge and Reasoning. We also test on a • Plan-and-Solve(Wangetal.,2023),wheremodelsare grounded social agent reasoning task called Thinking for prompted to first generate a plan and then solve the Doing (T4D) where models must leverage mental state rea- problem. SELF-DISCOVER differs by grounding the soning to determine actions to perform (Zhou et al.,2023), reasoningstructureinatomicreasoningmodules,and where GPT-4 with CoT only reaches around 50%. Finally, promptingthedecodingtofollowtheexplicitkey-value we subsample 200 examples from the MATH (Hendrycks reasoning structure. etal.,2021)testset,andgenerateinstance-levelreasoning structures via a one-shot demonstration to adapt to the com- Next, we also consider other baselines that make use of plexityofMATHtasks. Forevaluations,weuseaccuracyto the raw seed reasoning modules (RM) we pass to SELF- measure the model performance on BBH, T4D and MATH DISCOVER. Wecomparewiththefollowingmethods’per- (details can be found in AppendixB). formance and the inference call efficiency on a subset of 3.2. Models tasks. We use several state-of-the-art LLMs: GPT-4 (gpt-4-turbo- • CoT-Self-Consistency (Wang et al.,2022), we sample preview)(OpenAI,2023b),GPT-3.5-turbo(ChatGPT)(Ope- multipleoutputsfromLLMwithCoTandaggregatean- nAI, 2022) 1, instruction-tuned PaLM 2-L (Anil et al.,swerstogetthefinalanswer. Wecomparethismethod 2023) 2, and an open-source LLM Llama2-70B (Touvron onasubsetoftasksduetothecostofrepetitivequeries. et al.,2023). • Majority voting of each RM: we prompt models to 3.3. Baselines solvethetasksbyappendingeachRMandusemajority voting of all answers to get the final answer. We exam- Wecompare SELF-DISCOVER withother zero-shotprompt- inewhetherintegrating multiple RMsintoacoherent ing methods for LLM reasoning: reasoning structure is advantageous to applying each 1accessed in October-December 2023 RMtosolvethetaskandusemajorityvotingtoensem- 2ForMATH,weuseaPaLM2-Lmodelwithastrongerinstruc- blethempost-hoc,whichcostsmuchmoreinference tion

- Wecompare SELF-DISCOVER withother zero-shotprompt- inewhetherintegrating multiple RMsintoacoherent ing methods for LLM reasoning: reasoning structure is advantageous to applying each 1accessed in October-December 2023 RMtosolvethetaskandusemajorityvotingtoensem- 2ForMATH,weuseaPaLM2-Lmodelwithastrongerinstruc- blethempost-hoc,whichcostsmuchmoreinference tion tuning to enable betterinstruction following of more complex computation. reasoning structures. • BestofeachRM:thismethodassumesthatwehaveac- cesstooraclelabelsandusesthehighestaccuracyfrom 4 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Table1.Self-Discover significantly improves LLM reasoning across a diverse set of25 complex tasks: BBH, T4D and MATH. CoT: zero-shot Chain of Thought (Kojima et al.,2022). PS: plan- and-solve prompting (Wang et al.,2023). Method BBH T4D MATH PaLM 2-L 56% 30% 45% PaLM 2-L + CoT 60% 40% 42% PaLM 2-L + PS 61% 42% 49% PaLM 2-L + Self-Discover 67% 69% 50.5% GPT-4 58% 51% 70.5% Figure4.Breakdown of SELF-DISCOVER performance im- GPT-4 + CoT 75% 52% 71% provement on 4 categories on PaLM 2-L. SELF-DISCOVER per- GPT-4 + PS 73% 53% 70% forms the best on tasks requiring world knowledge. GPT-4 + Self-Discover 81% 85% 73% applyingeachRM.Wecomparewiththistoexamine over direct answering and CoT of PaLM 2-L are shown whether SELF-DISCOVER competeswithmethodsthat in Figure1, where we find SELF-DISCOVER outperforms depend on perfect prior knowledge of which RM to them on over 20/24 tasks. For a per-task performance for use on a new task. all 23 BBH tasks, please refer to AppendixC. On the grounded social agent task T4D, SELF- Furthermore, for analysis on universality of reasoning struc- DISCOVER reaches over ≥ 27% (32% ) absolute tures, we comparewith a prompt-optimization method that improvement over all baselines on PaLM 2-L (GPT-4). requireatrainingsettoimproveprompts: LLMsasoptimiz- SELF-DISCOVER achieves 69% and 85% accuracy on ers(OPRO)(Yangetal.,2023). Weaimto showthatwhen PaLM2-L andGPT-4, significantlyoutperforming previous weapplystructuresorpromptsoptimizedfromonemodel, SoTApromptingmethodsuchasForeseeandReflect(FaR) thereasoningstructurescanretainmoreperformancegains which employs an expert-designed reasoning structure. than the wordings of prompts. In contrast, SELF-DISCOVER generates the reasoning structure automatically from a set of atomic reasoning modules without human interventions. 4. Results ForMATH,weobserveamoderategainof1%-7%(2%-3%) We answer the following questions through experimental on PaLM 2-L (GPT-4) from SELF-DISCOVER compared results: 1) Doesdiscoveringreasoningstructuresimprove to the baselines. Upon error analysis (see AppendixDfor LLM reas

ly from a set of atomic reasoning modules without human interventions. 4. Results ForMATH,weobserveamoderategainof1%-7%(2%-3%) We answer the following questions through experimental on PaLM 2-L (GPT-4) from SELF-DISCOVER compared results: 1) Doesdiscoveringreasoningstructuresimprove to the baselines. Upon error analysis (see AppendixDfor LLM reasoning capabilities? (4.1) 2) Which categories details), we find that the reasoning structures generated by of problems do SELF-DISCOVER perform the best? (4.2) PaLM 2-L from SELF-DISCOVER are correct 87.5% of the and 3) Can SELF-DISCOVER boost LLM performance ef- time: human experts can follow the reasoning structures ficiently? (4.3) Finally,we willshowqualitativeexamples to solve the tasks perfectly. The majority of the failures of self-discovered structures, LLM output following the (74.7%)comesfromerrorsinexecutingthecomputations, structures, and compare with LLM output following other consistent with prior findings (Zheng et al.,2023). prompting methods for reasoning (4.4). 4.2. Which Types of Problems Do 4.1. Does SELF-DISCOVER Improve LLM Reasoning? SELF-DISCOVER Help the Most? Overall,SELF-DISCOVERimprovesPaLM2-LandGPT- SELF-DISCOVER performs best on tasks that require 4’s reasoning across diverse set of reasoning tasks. Ta- diverse world knowledge. Figure4presents the aver- ble1showstheoverallresultsoncomplexreasoningtasks age improvement in terms of delta in accuracy of SELF- of BBH, T4D and MATH using PaLM 2-L and GPT-4. DISCOVER over direct answer and CoT on 4 categories WecompareSelf-Discoverwithbaselinesincludingdirect of reasoning tasks we test. We adopt the categoriza- prompting, CoT, and Plan-and-Solve (PS). tion fromSuzgun et al.(2022). We find that SELF- Onaggregated23tasksofBBH, SELF-DISCOVER achieves DISCOVER improves over these two baselines on all cate- 7%and6%absoluteimprovementonPaLM2-LoverChain- gories,butespeciallyontasksthatrequireworldknowledge of-Thoughtand Plan-and-Solve, respectively. Similargains suchassportsunderstanding,movierecommendation,and (6%and8%)areobservedwhenSELF-DISCOVERisapplied ruin names. to GPT-4. Breakdown results of each task’s improvement Thesetasksdemandmodelstoreasonusingfactandgeneral 5 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure5.Comparisonofaccuracywithnumberofinferencecallsrequiredperinstance. ForCoT-Self-Consistency,wesample10 times. Best of each RM method requires gold labels (*). SELF-DISCOVER requires only 1 inference call per instance (plus 3 more meta-prompts on the task-level), same as Direct and CoT while reaching better performance compared with 40x more call required methods (majority voting of each RM) on GPT-4. We acknowledge that SELF-DISCOVER input and output are longer than CoT and Directprompting,increasingcost. However,asthenumberofinstancesincreases,theefficiencyofSELF-DISCOVER intermsofinference per instance is highly desirable. commonsenseknowledge. Weinterpret SELF-DISCOVER’s advantagesonthesetasksasstrengthfromintegratingmul- tiplereasoningmodulesfromvariousperspectivesasonly applyingCoTmightmisskeyknowledgeinthereasoning process. We observe that the gain on the Algorithmic cate- goryismoderate,consistentwiththefindingsfromSec.4.1 on MATH. 4.3. How Efficient is SELF-DISCOVER? SELF-DISCOVER achievesbetterperformancewhilere- quiring 10-40x fewer i

es,theefficiencyofSELF-DISCOVER intermsofinference per instance is highly desirable. commonsenseknowledge. Weinterpret SELF-DISCOVER’s advantagesonthesetasksasstrengthfromintegratingmul- tiplereasoningmodulesfromvariousperspectivesasonly applyingCoTmightmisskeyknowledgeinthereasoning process. We observe that the gain on the Algorithmic cate- goryismoderate,consistentwiththefindingsfromSec.4.1 on MATH. 4.3. How Efficient is SELF-DISCOVER? SELF-DISCOVER achievesbetterperformancewhilere- quiring 10-40x fewer inference computer compared to self-consistency or majority voting. Here we examine a subset of 2 tasks from BBH and present a more thor- ough comparison of methods including those requiring Figure6.Examplesofself-discoveredstructuresonBBH tasks many inference calls that are too costly to run on all 24 usingPaLM2-L.Weobservetraitsofatomicreasoningmodules tasks. Figure5shows average accuracy and number of such as “step-by-step thinking”, “ reflect on task nature”, and an in- inference calls required per instance for each method us- terestingcreativethinkingcasewheremodelsdevise analgorithm ing GPT-4. Accuracy wise (y-axis), we find that SELF- using stack to solve parenthesis parsing task. DISCOVER outperforms other baselines even those that re- quire repeated inference calls such as CoT-self-consistency multiplereasoningmodules,andprovidesinsightsonhow and majority voting of applying each RM. Efficiency wise to solve the tasks. Furthermore, example of comparing (x-axis), SELF-DISCOVER only requires one call per in- reasoning processesfrom CoT, Plan-and-Solve, and SELF- stanceandthreemoreinferencecallsonthetask-level,CoT- DISCOVER is shown in Figure7. We find that CoT and self-consistency requires 10 times more since we have to Plan-and-Solve makes incorrect assertions early and arrives sample 10 times for each instance, and methods using each at a wrong answer while following structure from SELF- RMrequires40timesmoreasweuse40RMs. Insummary, DISCOVER leads the model to generate logical conclusions SELF-DISCOVERpresentsitselfastrongreasoningboosting (“path is closed as the beginning and ending coordinates method that is efficient to deploy on large-scale. are the same”) and arrive at the correct answer. 4.4. Qualitative Examples 5. Deep DivingInto Self-DiscoveredReasoning Weshowexamplesofmodel-discoveredstructuresfordiffer- Structures entreasoningtasksinFigure6fromPaLM2-L.Weobserve After experimental results showing the effectiveness and thateachstructureisuniquelyadaptedtothetask,integrates efficiency of SELF-DISCOVER on a range of reasoning 6 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure7.ComparisonofgeneratedreasoningprocessfromCoT,Plan-and-Solve, and SELF-DISCOVER onBBH-geometricshape task. BothCoTandPlan-and-Solveincorrectlyassertsthatthepathdoesnotformaregularshapeasitisnotaclosedpath(highlightedin red) and arrive at a wrong answer. The reasoningstructure (in blueCourier font) from SELF-DISCOVER first breaks down each line segmentandanalyzethecoordinatescarefully,thenleverageslogicalreasoningtoconcludethatitformsaclosedshapeasthepathendsat the same coordinate (highlighted in purple and orange), and selects the correct answer through final reasoning. tasks, thissection further analyzesare allactions of SELF- DISCOVER needed and what other benefits can self- discovered structures bring? In Sec.5.1, we show thatit is critical tothe model’sperformance touse thereasoning structures discovered through the three steps of SELECT, ADAPT and IMPLEMENT. In Sec.5.2, we demonstrate theuniversalityoftheself-discoveredreasoningstructures

hapeasthepathendsat the same coordinate (highlighted in purple and orange), and selects the correct answer through final reasoning. tasks, thissection further analyzesare allactions of SELF- DISCOVER needed and what other benefits can self- discovered structures bring? In Sec.5.1, we show thatit is critical tothe model’sperformance touse thereasoning structures discovered through the three steps of SELECT, ADAPT and IMPLEMENT. In Sec.5.2, we demonstrate theuniversalityoftheself-discoveredreasoningstructures by (1) applying the structures discovered by PaLM 2-L to GPT-4,(2)applyingthestructuresdiscoveredbyGPT-4to Llama-2-70B.Wefurthershowthecommonalitiesbetween the reasoning structures and human reasoning patterns in Figure8.Ablation study on three SELF-DISCOVER actions on AppendixE. 4reasoning tasks: all threeactionsare beneficialfor task-solving. 5.1. Importance of SELF-DISCOVER Actions We conduct ablation study on the three actions: SELECT, 5.2. Towards Universality of Discovered Reasoning ADAPT,andIMPLEMENTtoanalyzetheeffectsof SELF- Structures DISCOVER actions. Figure8showresultsusingGPT-4on4 Applying PaLM 2-L Discovered Structures to GPT-4 reasoningtaskswhenweapplySELECT (-S)orapplySE- We first use a PaLM 2-L model to discover the reasoning LECTandADAPT(-SA)orapplyallthreeactions. Wefind structuresof4reasoningtasks. Then,weapplytheresulting thatwitheachstage,model’szero-shotreasoningcapability reasoning structures to the decoding of GPT-4 as grounding. improveconsistently across tasks, indicatingthat all three We compare our approach to OPRO (Yang et al.,2023) actions are beneficial. In particular, after all three actions whichdiscoveredzero-shot-promptsthroughoptimizations. SAI,thereasoningstructuresareadaptedtobetaskspecific, We apply OPRO prompts optimized using PaLM 2-L on and bring the most gain to solving the reasoning tasks. each task to GPT-4 on the same reasoning tasks. Figure9 shows that SELF-DISCOVER outperforms OPROon 3 out of4tasksdespitethatOPROused20%datatooptimizethe 7 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures prompting methods has some strengths and weaknesses in terms of their successful applicationdomain. Our work SELF-DISCOVER presentsthemissingpieceintheprompt- ing literature, as SELF-DISCOVER provides a way to self- compose over various prompting methods via the proposed self-discovery mechanism. Composing over prompting methodsin SELF-DISCOVER isanalogoustotheprogram- ming literature where a program is written using various basic buildingblocks such asfor loop, if/elsecondition etc. 6.2. Reasoning and Planning Figure9.Transferrabilitytestsofoptimizedprompts(OPRO) With the development of various reasoning and plan- andcomposedstructures(SELF-DISCOVER). Theresultsshown ning benchmarks such as GSM8

where a program is written using various basic buildingblocks such asfor loop, if/elsecondition etc. 6.2. Reasoning and Planning Figure9.Transferrabilitytestsofoptimizedprompts(OPRO) With the development of various reasoning and plan- andcomposedstructures(SELF-DISCOVER). Theresultsshown ning benchmarks such as GSM8K (Cobbe et al.,2021), are from GPT-4 using the prompts and structures optimized or Math(Hendrycksetal.),BigBench(Srivastavaetal.,2023) composedusingPaLM2-L.Wefindthatself-discoveredreasoning etc.,variousmethodshavebeenproposedtoimprovemodel structure transfers more robustly than optimized prompts. performance. Oftenthesemethodsinducespecificreason- prompt. Incontrast, SELF-DISCOVER isdoneinazero-shot ing structures mimicking the reasoning structure of the un- manner, demonstrating the efficiency of our method and derlying task associated with the dataset. For example, universality of the discovered reasoning structures. chain of thought (Wei et al.,2022) and scratchpad (Nye ApplyingGPT-4DiscoveredStructurestoLlama2and et al.,2021) induce generation of explanations associated ChatGPT Motivated by transferrability performance with a reasoning question. Similarly other methods induces acrossLLMs,wefurtherinvestigatecanself-discoveredrea- specific reasoning structures such as question summariza- soning structures from LLMs boost reasoning for smaller tion (Kuznia et al.,2022), question decomposition (Patel LMs that are challenging to come up with structures them- et al.,2022), program generation (Mishra et al.,2022a; selves3. We use GPT-4 to discover the task-intrinsic rea- Chenetal.,2022;Gaoetal.,2023b),etc. However,inareal soning structures, and then apply those structures to the world user traffic, queries can be diverse covering various decodingofopen-sourcedLlama2-70BaswellasGPT-3.5- reasoning structures. Our work SELF-DISCOVER allows turbo (ChatGPT) on two subsets of tasks from BBH. We modelstocombinemultiplereasoningapproachesbyself- findthatusingself-discoveredstructuresonLlama2(52%) composingintoastructurewithouttheneedtoaccesstask outperforms CoT (42%) on disambiguation QA zero-shot labels. There have been some related work that explores andonGPT-3.5-turbo(56%)outperformsCoT(51%)onge- LLM combining skills in-context such as SkiC (Chen et al., ometrywith3-shotdemonstrationfromstructuredreasoning 2023),devisingastrategy(Gaoet al.,2023a),andplanning process. with iterative quering (Liu et al.,2023). However, they requirehumanannotatingskillsandreasoningplanswhile SELF-DISCOVERleveragesascalablesolutionwiththehelp 6. Related Work of LLM’s meta-task reasoning capabilities. 6.1. Prompting Methods 7. Conclusion Recent advancements in the area of LLMs have given rise We introduce SELF-DISCOVER, an efficient and performant to a plethora of few-shot (Brown et al.,2020) and instruc- framework for models to self-discover a reasoning structure tion (Mishra et al.,2022c;Wei et al.,2021;Ouyang et al., for any task from a seed set of general problem-solving 2022) prompting techniques, including Chain-of-Thought skills. We observe drastic improvements on challengin

ancements in the area of LLMs have given rise We introduce SELF-DISCOVER, an efficient and performant to a plethora of few-shot (Brown et al.,2020) and instruc- framework for models to self-discover a reasoning structure tion (Mishra et al.,2022c;Wei et al.,2021;Ouyang et al., for any task from a seed set of general problem-solving 2022) prompting techniques, including Chain-of-Thought skills. We observe drastic improvements on challenging prompting (CoT) (Nye et al.,2021;Wei et al.,2022), Least- reasoning benchmarks from multiple LLMs up to 30%. Ab- to-most prompting (Zhou et al.,2022a;Drozdov et al., lations study of SELF-DISCOVER demonstrates that the 2022), Decomposed prompting (Khot et al.,2022), Re- composedreasoningstructuresareuniversallytransferable framing (Mishra et al.,2022b), Help Me Think Prompt- betweenLLMs. Forward looking,we areexcited toexplore ing (Mishra & Nouri,2023), Stepback Prompting (Zheng more on LLM structured reasoning to push the boundary et al.,2023) and search-based approaches like Tree-of- of problem-solving and discover potentials for Human-AI Thought(ToT)(Yaoetal.,2023a),Graph-of-Thought(Besta collaboration. et al.,2023;Yao et al.,2023b), Branch-solve-merge (Saha et al.,2023) and RAP (Hao et al.,2023). Each of the 3We triedzero-shot meta prompting Llama2 but observedlow- quality structure outputs. 8 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Acknowledgement Gao,C.,Jiang,H.,Cai,D.,Shi,S.,andLam,W. Strategyllm: We thankAndrew Dai and AdamsYu ofGoogle DeepMind Largelanguage modelsas strategygenerators, executors, for their insightful feedback on this paper. optimizers, and evaluators for problem solving. arXiv preprint arXiv:2311.08803, 2023a. References Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Anil, R., Dai, A. M., Firat, O., Johnson, M., Lepikhin, Y., Callan, J., and Neubig, G. Pal: Program-aided lan- D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, guagemodels. InInternationalConferenceonMachine Z., et al. Palm 2 technical report. arXiv preprint Learning, pp. 10764–10799. PMLR, 2023b. arXiv:2305.10403, 2023. Hao, S., Gu, Y., Ma, H.,Hong, J. J., Wang, Z., Wang, D. Z., Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gi- andHu, Z. Reasoningwith languagemodel isplanning aninazzi, L., Gajda, J., Lehmann, T., Podstawski, M., with world model. arXiv preprint arXiv:2305.14992, Niewiadomski, H.,Nyczyk, P., etal. Graph ofthoughts: 2023. Solving elaborate problems with large language models. Hendrycks, D., Burns,C., Kadavath, S.,Arora, A., Basart, arXiv preprint arXiv:2308.09687, 2023. S., Tang, E., Song, D., and Steinhardt, J. Measuring Brown,T.,Mann, B., Ryder, N.,Subbiah, M., Kaplan, J.D., mathematicalproblemsolvingwiththemathdataset.Sort, Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., 2(4):0–6. Askell, A., et al. Languagemodels are few-shot learners. Advancesinneuralinformationprocessingsystems,33: Hendrycks, D., Burns,C., Kadavath, S.,Arora, A., Basart, 1877–1901, 2020. S.,Tang,E.,Song,D.,andSteinhardt,J. Measuringmath- Chen, J., Pan, X., Yu, D., Song, K., Wang, X., Yu, D., ematical problem solving wit

Sort, Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., 2(4):0–6. Askell, A., et al. Languagemodels are few-shot learners. Advancesinneuralinformationprocessingsystems,33: Hendrycks, D., Burns,C., Kadavath, S.,Arora, A., Basart, 1877–1901, 2020. S.,Tang,E.,Song,D.,andSteinhardt,J. Measuringmath- Chen, J., Pan, X., Yu, D., Song, K., Wang, X., Yu, D., ematical problem solving with the math dataset, 2021. and Chen, J. Skills-in-context prompting: Unlocking Khot, T., Trivedi, H., Finlayson, M., Fu, Y., Richardson, K., compositionalityinlargelanguagemodels.arXivpreprint Clark, P., and Sabharwal, A. Decomposed prompting: arXiv:2308.00304, 2023. A modular approach for solving complex tasks. In The Chen, W., Ma, X., Wang, X., and Cohen, W. W. Program Eleventh International Conference on Learning Repre- of thoughts prompting: Disentangling computation from sentations, 2022. reasoningfornumericalreasoningtasks. arXivpreprint Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., and Iwasawa, arXiv:2211.12588, 2022. Y. Large language models are zero-shot reasoners. Ad- Chowdhery,A., Narang,S., Devlin,J., Bosma,M., Mishra, vances in neural information processing systems, 35: G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., 22199–22213, 2022. Gehrmann, S., et al. Palm: Scaling language modeling Kuznia, K., Mishra, S., Parmar, M., and Baral, C. Less is with pathways. arXiv preprint arXiv:2204.02311, 2022. more: Summaryoflonginstructionsisbetterforprogram Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tay, Y., synthesis. In Proceedings of the 2022 Conference on Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, EmpiricalMethods inNaturalLanguage Processing, pp. S.,etal. Scalinginstruction-finetunedlanguagemodels. 4532–4552, 2022. arXiv preprint arXiv:2210.11416, 2022. Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Liu, T., Guo, Q., Yang, Y., Hu, X., Zhang, Y., Qiu, X., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, and Zhang, Z. Plan, verify and switch: Integrated R., et al. Training verifiers to solvemath word problems. reasoning with diverse x-of-thoughts. arXiv preprint arXiv preprint arXiv:2110.14168, 2021. arXiv:2310.14628, 2023. Drozdov, A., Schärli, N., Akyürek, E., Scales, N., Song, Mishra, S. and Nouri, E. HELP ME THINK: A simple X., Chen, X., Bousquet, O., and Zhou, D. Composi- promptingstrategyfornon-expertstocreatecustomized tionalsemanticparsingwithlargelanguagemodels.arXiv content with models. In Rogers, A., Boyd-Graber, J., preprint arXiv:2209.15003, 2022. and Okazaki, N. (eds.), Findings of the Association for ComputationalLinguistics: ACL2023,pp.11834–11890, Fernando, C., Banarse, D., Michalewski, H., Osindero, Toronto, Canada, July 2023. Association for Computa- S.,andRocktäschel,T. Promptbreeder: Self-referential tional Linguistics. doi: 10.18653/v1/2023.findings-acl. self-improvement via prompt evolution. arXiv preprint 751. URL https://aclanthology.org/2023. arXiv:2309.16797, 2023. findings-acl.751. 9 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Mishra, S., Finlayson, M., Lu, P., Tang, L., Wellec

erential tional Linguistics. doi: 10.18653/v1/2023.findings-acl. self-improvement via prompt evolution. arXiv preprint 751. URL https://aclanthology.org/2023. arXiv:2309.16797, 2023. findings-acl.751. 9 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Mishra, S., Finlayson, M., Lu, P., Tang, L., Welleck, S., Saha, S., Levy, O., Celikyilmaz, A., Bansal, M., Weston, Baral, C., Rajpurohit, T., Tafjord, O., Sabharwal, A., J., and Li, X. Branch-solve-merge improves large lan- Clark,P.,etal. Lila: Aunifiedbenchmarkformathemati- guage model evaluation and generation. arXiv preprint cal reasoning. InProceedings of the 2022 Conference on arXiv:2310.15123, 2023. EmpiricalMethods inNaturalLanguage Processing, pp. Srivastava,A.,Rastogi,A.,Rao,A.,Shoeb,A.A.M.,Abid, 5807–5832, 2022a. A., Fisch, A., Brown, A. R., Santoro, A., Gupta, A., Mishra,S.,Khashabi,D.,Baral,C.,Choi,Y.,andHajishirzi, Garriga-Alonso, A., et al. Beyond the imitation game: H. Reframing instructional prompts to gptk’s language. Quantifyingandextrapolatingthecapabilitiesoflanguage InFindingsoftheAssociationforComputationalLinguis- models. Transactions on Machine Learning Research, tics: ACL 2022, pp. 589–612, 2022b. 2023. Mishra,S.,Khashabi,D.,Baral,C.,andHajishirzi,H.Cross- Suzgun, M., Scales, N., Schärli, N., Gehrmann, S., Tay, task generalization via natural language crowdsourcing Y., Chung, H. W., Chowdhery, A., Le, Q. V., Chi, instructions. InProceedingsof the60thAnnual Meeting E. H., Zhou, D., et al. Challenging big-bench tasks and oftheAssociationforComputationalLinguistics(Volume whether chain-of-thought can solve them. arXiv preprint 1: Long Papers), pp. 3470–3487, 2022c. arXiv:2210.09261, 2022. Newell, A., Shaw, J. C., and Simon, H. A. Elements of a Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, theory of humanproblem solving. Psychological review, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., 65(3):151, 1958. Bhosale, S., et al. Llama 2: Open foundation and fine- tuned chat models. arXiv preprint arXiv:2307.09288, Nye,M.,Andreassen,A.J.,Gur-Ari,G.,Michalewski,H., 2023. Austin,J.,Bieber,D.,Dohan,D.,Lewkowycz,A.,Bosma, Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones, M., Luan, D., et al. Show your work: Scratchpads for L.,Gomez,A.N.,Kaiser,L.u.,andPolosukhin,I. Atten- intermediate computation with language models. arXiv tion is all you need. In Advances in Neural Information preprint arXiv:2112.00114, 2021. ProcessingSystems,volume 30.CurranAssociates,Inc., OpenAI. Chatgpt: Optimizing language models for dia- 2017. URL https://proceedings.neurips. logue, 2022. URL https://openai.com/blog/ cc/paper_files/paper/2017/file/ chatgpt/. 3f5ee243547dee91fbd053c1c4a845aa-Paper. pdf. OpenAI. Json generation mode, 2023a. URL Wang,L.,Xu, W.,Lan,Y.,Hu, Z.,Lan,Y.,Lee, R.K.-W., https://platform.openai.com/docs/ and Lim, E.-P. Plan-and-solve prompting: Improving guides/text-generation/json-mode.

file/ chatgpt/. 3f5ee243547dee91fbd053c1c4a845aa-Paper. pdf. OpenAI. Json generation mode, 2023a. URL Wang,L.,Xu, W.,Lan,Y.,Hu, Z.,Lan,Y.,Lee, R.K.-W., https://platform.openai.com/docs/ and Lim, E.-P. Plan-and-solve prompting: Improving guides/text-generation/json-mode. zero-shot chain-of-thought reasoning by large language OpenAI, R. Gpt-4 technical report. arXiv, pp. 2303–08774, models. arXiv preprint arXiv:2305.04091, 2023. 2023b. Wang, X., Wei, J., Schuurmans, D., Le, Q. V., Chi, Ouyang,L.,Wu,J.,Jiang,X.,Almeida,D.,Wainwright,C., E. H., Narang, S., Chowdhery, A., and Zhou, D. Self- Mishkin,P.,Zhang,C.,Agarwal,S.,Slama,K.,Ray,A., consistencyimproveschainofthoughtreasoninginlan- et al. Training language models to follow instructions guage models. In The Eleventh International Conference withhumanfeedback. AdvancesinNeuralInformation on Learning Representations, 2022. Processing Systems, 35:27730–27744, 2022. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A. W., Lester, Patel, P., Mishra, S., Parmar, M., and Baral, C. Is a ques- B., Du, N., Dai, A. M., and Le, Q. V. Finetuned lan- tiondecompositionunitallweneed? InProceedingsof guage models are zero-shot learners. In International the 2022 Conference on Empirical Methods in Natural Conference on Learning Representations, 2021. Language Processing, pp. 4553–4569, 2022. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. Chain-of-thought Polya, G. How to solve it: A new aspect of mathematical prompting elicits reasoning in large language models. method, volume 85. Princeton university press, 2004. Advances in Neural Information Processing Systems, 35: Rasmussen, J. Skills, rules, and knowledge; signals, signs, 24824–24837, 2022. and symbols, and other distinctions in human perfor- Yang,C.,Wang,X.,Lu,Y.,Liu,H.,Le,Q.V.,Zhou,D.,and mance models. IEEE transactions on systems, man, and Chen, X. Large language models as optimizers. arXiv cybernetics, (3):257–266, 1983. preprint arXiv:2309.03409, 2023. 10 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Yao,S.,Yu,D.,Zhao,J.,Shafran,I.,Griffiths,T.L.,Cao,Y., and Narasimhan, K. Tree of thoughts: Deliberate prob- lem solving with large language models. arXiv preprint arXiv:2305.10601, 2023a. Yao, Y., Li, Z., and Zhao, H. Beyond chain-of-thought, effective graph-of-thought reasoning in large language models. arXiv preprint arXiv:2305.16582, 2023b. Yasunaga, M., Chen, X., Li, Y., Pasupat, P., Leskovec, J., Liang,P.,Chi,E.H.,andZhou,D.Largelanguagemodels asanalogicalreasoners. arXivpreprintarXiv:2310.01714, 2023. Zheng,H.S.,Mishra,S.,Chen,X.,Cheng,H.-T.,Chi,E.H., Le, Q. V., and Zhou, D. Take a step back: Evoking reasoningviaabstractioninlargelanguagemodels. arXiv preprint arXiv:2310.06117, 2023. Zhong, R., Lee, K., Zhang, Z., and Klein, D. Adapt- ing language models for zero-shot learning by meta- tuning on dataset and promptcollections. arXiv preprint arXiv:2104.04670, 2021. Zhou,D.,Schärli,N.,Hou,L.,Wei,J.,Scales,N.,Wang,X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q. V., et al. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conferen

reasoningviaabstractioninlargelanguagemodels. arXiv preprint arXiv:2310.06117, 2023. Zhong, R., Lee, K., Zhang, Z., and Klein, D. Adapt- ing language models for zero-shot learning by meta- tuning on dataset and promptcollections. arXiv preprint arXiv:2104.04670, 2021. Zhou,D.,Schärli,N.,Hou,L.,Wei,J.,Scales,N.,Wang,X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q. V., et al. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conference on Learning Representations, 2022a. Zhou, P., Madaan, A., Potharaju, S. P., Gupta, A., McKee, K. R., Holtzman, A., Pujara, J., Ren, X., Mishra, S., Nematzadeh, A., et al. How far are large language mod- els from agents with theory-of-mind? arXiv preprint arXiv:2310.03051, 2023. Zhou, Y., Muresanu, A. I., Han, Z., Paster, K., Pitis, S., Chan,H., andBa, J. Large languagemodels are human- level prompt engineers. In The Eleventh International Conference on Learning Representations, 2022b. 11 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures A. Self-Discover Prompt Details Table2shows all 39 reasoning modules we use for SELF-DISCOVER, adopted fromFernando et al.(2023), that contain cognitive heuristics of problem-solving. Figure10contains the structure of the three actions of SELF-DISCOVER during Stage 1, where it discovers an intrinsic reasoning structure on the task-level. For Stage 2,where we usethe self-discoveredstructure to solve thetask instances, westart with theprompt: “Follow the step-by-stepreasoningplaninJSONtocorrectlysolvethetask. Fillinthevaluesfollowingthekeysbyreasoningspecifically about thetaskgiven. Do notsimply rephrase thekeys.”,followed bythe reasoningstructure, andfinally thetask instance. Figure10.Meta-Promptsforthe three actionsof SELF-DISCOVER. Eachmeta-prompt consistsofan instructioninthe beginning and the end, reasoning module descriptions, and task examples without labels. For IMPLEMENT, to show model an example of a reasoning structure (plan), we present a human-written structure in JSON for another task. B. Evaluation Details We use accuracy and exact matching as with other methods tested on BBH, T4D and MATH. To properly evaluate the generated answers from LLMs, we prompt the models to end the answer with “Thus, the final answer is [X]”, where X is eitherone answer option suchas “A” ora string such as“ valid”. During evaluation, we manually examine each task’s outputs from LLMs and designheuristics to extract the final answers. For MATH dataset, we find that it ischallenging to extracttheanswers accurately. Asaresult, wesubsample200testexamplesfrom MATH, andmanuallysanitycheckand annotate the extracted answers for all methods tested in our paper. C. BBH Per Task Performance Per-task performance on BBH (23 tasks in total) are shown in Table3. D. Error Analysis We performanerror analysisof SELF-DISCOVER onthe MATH datasetof200 samplestounderstandthe failuremodes. We manually annotate whether the generated reasoning structure is correct or not together with whether the correctness of modelpredictionusing SELF-DISCOVER. Areasoningstructureisdefinedascorrectifahumanexpertcansolvethetaskby simply following the reasoning structure. 12 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Table2.All 39 reasoning modules consisting of high-level cognitive heuristics for problem-solving. We adopt them fromFernando et al. (2023). ReasoningModules 1HowcouldIdeviseanexperimenttohelpsolvethatproblem? 2Makealistofideasforsolvingthisproblem,andapplythemonebyonetotheproblemtoseeifanyprogresscanbemade. 3HowcouldImeasureprogressonthisproblem? 4HowcanIsimplifytheproblemsothatitiseasiertosolve? 5Whatarethekeyassumptionsunderlyingthisproblem? 6Whatarethepotentialrisksanddrawbacksofeachsolution? 7Whatarethea

asoning modules consisting of high-level cognitive heuristics for problem-solving. We adopt them fromFernando et al. (2023). ReasoningModules 1HowcouldIdeviseanexperimenttohelpsolvethatproblem? 2Makealistofideasforsolvingthisproblem,andapplythemonebyonetotheproblemtoseeifanyprogresscanbemade. 3HowcouldImeasureprogressonthisproblem? 4HowcanIsimplifytheproblemsothatitiseasiertosolve? 5Whatarethekeyassumptionsunderlyingthisproblem? 6Whatarethepotentialrisksanddrawbacksofeachsolution? 7Whatarethealternativeperspectivesorviewpointsonthisproblem? 8Whatarethelong-termimplicationsofthisproblemanditssolutions? 9HowcanIbreakdownthisproblemintosmaller,moremanageableparts? 10CriticalThinking: Thisstyleinvolvesanalyzingtheproblemfromdifferentperspectives,questioningassumptions,andevaluating theevidenceorinformationavailable. Itfocusesonlogicalreasoning,evidence-baseddecision-making,andidentifying potentialbiasesorflawsinthinking. 11Trycreativethinking,generateinnovativeandout-of-the-boxideastosolvetheproblem. Exploreunconventionalsolutions, thinkingbeyondtraditionalboundaries,andencouragingimaginationandoriginality. 12Seekinputandcollaborationfromotherstosolvetheproblem. Emphasizeteamwork,opencommunication,andleveragingthe diverseperspectivesandexpertiseofagrouptocomeupwitheffectivesolutions. 13Usesystemsthinking: Considertheproblemaspartofalargersystemandunderstandingtheinterconnectednessofvariouselements. Focusesonidentifyingtheunderlyingcauses,feedbackloops,andinterdependenciesthatinfluencetheproblem,anddevelopingholistic solutionsthataddressthesystemasawhole. 14UseRiskAnalysis: Evaluatepotentialrisks,uncertainties,andtradeoffsassociatedwithdifferentsolutionsorapproachestoa problem. Emphasizeassessingthepotentialconsequencesandlikelihoodofsuccessorfailure,andmakinginformeddecisionsbased13 onabalancedanalysisofrisksandbenefits. 15UseReflectiveThinking: Stepbackfromtheproblem,takethetimeforintrospectionandself-reflection. Examinepersonalbiases, assumptions,andmentalmodelsthatmayinfluenceproblem-solving,andbeingopentolearningfrompastexperiencestoimprove futureapproaches. 16Whatisthecoreissueorproblemthatneedstobeaddressed? 17Whataretheunderlyingcausesorfactorscontributingtotheproblem? 18Arethereanypotentialsolutionsorstrategiesthathavebeentriedbefore? Ifyes,whatweretheoutcomesandlessonslearned? 19Whatarethepotentialobstaclesorchallengesthatmightariseinsolvingthisproblem? 20Arethereanyrelevantdataorinformationthatcanprovideinsightsintotheproblem? Ifyes,whatdatasourcesareavailable, andhowcantheybeanalyzed? 21Arethereanystakeholdersorindividualswhoaredirectlyaffectedbytheproblem? Whataretheirperspectivesandneeds? 22Whatresources(financial,human,technological,etc.) areneededtotackletheproblemeffectively? 23Howcanprogressorsuccessinsolvingtheproblembemeasuredorevaluated? 24Whatindicatorsormetricscanbeused? 25Istheproblematechnicalorpracticalonethatrequiresaspecificexpertiseorskillset? Orisitmoreofaconceptualor theoreticalproblem? 26Doestheprobleminvolveaphysicalconstraint,suchaslimitedresources,infrastructure,orspace? 27Istheproblemrelatedtohumanbehavior,suchasasocial,cultural,orpsychologicalissue? 28Doestheprobleminvolvedecision-makingorplanning,wherechoicesneedtobemadeunderuncertaintyorwithcompeting objectives? 29Istheproblemananalyticalonethatrequiresdataanalysis,modeling,oroptimizationtechniques? 30Istheproblemadesignchallengethatrequirescreativesolutionsandinnovation? 31Doestheproblemrequireaddressingsystemicorstructuralissuesratherthanjustindividualinstances? 32Istheproblemtime-sensitiveorurgent,requiringimmediateattentionandaction? 33Whatkindsofsolutiontypicallyareproducedforthiskindofproblemspecification? 34Giventheproblemspecificationandthecurrentbestsolution,haveaguessaboutotherpossiblesolutions. 35Let’simaginethecurrentbestsolutionistotallywrong,whatotherwaysaretheretothinkabouttheproblemspecification? 36Whatisthebestwaytomodifythiscurrentbestsolution,givenwhatyouknowaboutthesekindsofproblemspecification? 37Ignoringthecurrentbestsolution,createanentirelynewsolutiontotheproblem. 38Let’sthinkstepbystep. 39Let’smakea

entionandaction? 33Whatkindsofsolutiontypicallyareproducedforthiskindofproblemspecification? 34Giventheproblemspecificationandthecurrentbestsolution,haveaguessaboutotherpossiblesolutions. 35Let’simaginethecurrentbestsolutionistotallywrong,whatotherwaysaretheretothinkabouttheproblemspecification? 36Whatisthebestwaytomodifythiscurrentbestsolution,givenwhatyouknowaboutthesekindsofproblemspecification? 37Ignoringthecurrentbestsolution,createanentirelynewsolutiontotheproblem. 38Let’sthinkstepbystep. 39Let’smakeastepbystepplanandimplementitwithgoodnotionandexplanation. SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Table3. Big Bench-Hard (Suzgun et al.,2022) per-task performance of GPT-4 and PaLM 2-L with S ELF-DISCOVER. Out of 200 examples, we find that 87.5% (175) examples have correct reasoning structures. 12.5% (25) examples have incorrectreasoningstructuresleadingtopredictionerrors. Table4shows4suchexampleswheretheLLMmisunderstands the task, or makes an error in one of the steps or adds unnecessary steps in the reasoning structure. Next, we analyze the errors made by the model in SELF-DISCOVER: out of 99 examples where the model prediction is wrong,wrongreasoningstructuresaccountforonly25.3%oftheerrors. Theremaining74.7%errorsareduetoerrorsin the intermediate calculations suchas math computations. Table5shows 3 examples of sucherrors. Thisinsight indicates that futureimprovements shouldaim at improvingthe step-wise calculationaccuracy of LLMs,such as usingtools or code generation. E. Further Anaysis Model-Discovered Reasoning Structures vs. Human Reasoning Patterns We investigate whether LLM-discovered reasoningstructuressharesomecommonalitieswithhumanreasoningpatterns. Wegivehumans3taskinstanceswithout labels and an example reasoning structure (same as SELF-DISCOVER meta-reasoning stage) and ask them to write a reasoning structure for a task before solving it. Figure11shows comparison of human and LLM-composed reasoning structuresontheBBH-navigationtask. Weobservesimilarstructuressuchasmental-notingaftereachmovement. From promisingfindingsofLLMself-discoveredstructuresboostandsharetraitsofhumanmeta-reasoning,wehopetoencourage more future work to study humna-AI collaboration for complex problem-solving. BigBench-HardTask Human(Avg.) Human(Max) GPT-4 GPT-4 GPT-4 PaLM2-L PaLM2-L PaLM2-L Direct +CoT +Self-Discover Direct +CoT +Self-Discover boolean_expressions 79 100 73 83 85 71 84 84 causal_judgement 70 100 67 75 80 46 59 61 date_understanding 77 100 74 80 81 73 78 78 disambiguation_qa 67 93 60 70 80 54 50 57 dyck_languages 48 100 69 73 77 94 95 9814 formal_fallacies 91 100 60 60 80 60 63 69 geometric_shapes 54 100

93 60 70 80 54 50 57 dyck_languages 48 100 69 73 77 94 95 9814 formal_fallacies 91 100 60 60 80 60 63 69 geometric_shapes 54 100 30 56 60 33 34 39 hyperbaton 75 100 68 69 76 80 75 82 logical_deduction_seven_objects 40 89 60 70 70 45 39 50 movie_recommendation 61 90 70 70 86 83 54 66 multistep_arithmetic_two 10 25 10 92 70 4 50 47 navigate 82 100 70 90 90 38 63 67 object_counting 86 100 90 100 100 27 44 70 penguins_in_a_table 78 100 80 100 90 70 67 75 reasoning_about_colored_objects 75 100 77 80 79 36 79 75 ruin_names 78 100 90 80 97 79 58 90 salient_translation_error_detection 37 80 40 50 70 56 48 60 snarks 77 100 73 89 97 58 62 86 sports_understanding 71 100 54 61 90 44 47 89 temporal_sequences 91 100 96 99 100 99 97 99 tracking_shuffled_objects_seven_objects 65 100 24 80 68 22 58 36 web_of_lies 81 100 15 80 71 54 42 67 word_sorting 63 100 65 90 85 12 4 15 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Table4.Examples of wrong reasoning structures for MATH. The first error in the reasoning structure is highlighted in red. Prompt Reasoning Structure Error Howmanynumbersbetween 1. Findthenumberofmultiplesof3between1and Need to sub- 1 and 2005 are integer multi- 2005. tract the num- ples of 3 or 4 but not 12? 2. Find the number of multiples of 4 between 1 ber of mul- and 2005.

Error Howmanynumbersbetween 1. Findthenumberofmultiplesof3between1and Need to sub- 1 and 2005 are integer multi- 2005. tract the num- ples of 3 or 4 but not 12? 2. Find the number of multiples of 4 between 1 ber of mul- and 2005. tiples of 12 3. Find the number of multiples of 12 between 1 twice instead and 2005. of once. 4. Add the number of multiples of 3 and 4. 5. Subtractthenumberofmultiplesof12(toavoid double counting). 6. The result is the number of integers between 1 and 2005 that are multiples of 3 or 4 but not 12. How many numbers are in 1. Find the number of pairs in the list. LLM misun- the list 6,7,10,11,14,15,..., 2. Find the number of numbers in each pair. derstands the 94,95,98? 3. Multiplythenumber ofpairsby thenumberof task: the rea- numbers in each pair to find the total number of soning struc- numbers in the list. ture is irrele- vant. A restaurant offers three 1. Define the variables: There are 365 desserts, and exactly twice - M: number of main courses nights in the as many appetizers as main - A: number of appetizers (twice the number of year 2003. courses. Adinnerconsistsof main courses) an appetizer, a main course, - D: number of desserts (fixed at 3) and a dessert. What is the - T: total number of possible dinner combinations leastnumberofmaincourses 2. Calculate the total number of possible dinner thattherestaurantshouldof- combinations: fer so that a customer could - T = M * A * D have a different dinner each 3. Set the total number of possible dinner combi- night in the year 2003? nations equal to the number of nights in the year 2003: - T = 2003 4. Substitute the expression for A in terms of M: - T = M * (2 * M) * 3 5. Solve the equation for M: - 6M 2 = 2003 6. Find the smallest integer value of M that satis- fies the equation. Determine the number of 1. Count the total number of letters in the word Step 5 is un- ways to arrange the letters of NINE. necessary and

- 6M 2 = 2003 6. Find the smallest integer value of M that satis- fies the equation. Determine the number of 1. Count the total number of letters in the word Step 5 is un- ways to arrange the letters of NINE. necessary and the word NINE. 2. Identify thenumberoftimeseachletterappears leads to error in the word. in the final re- 3. Calculate the factorial of the total number of sult. letters. 4. Dividethe factorialby thefactorial of the num- ber of times each letter appears. 5. Multiply theresultsfrom step4for each unique letter.} 15 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Table5.Examples of wrong calculations for MATH. The first error in the intermediate computations is highlighted in red. Prompt Self-Discover Prediction Error How many ways are there "1. Calculatethetotalnumberofwaystoarrange There are 7ro- to arrange 6 people around 6 people in 7 seats without considering rotations tations instead acirculartablewith7seats? as the same.": 7 * 6 * 5 * 4 * 3 * 2 = 5040 of 6 that are (Two seatingsareconsidered - There are 7 choices for the first person. consideredthe the same if one is a rotation - There are 6 choices for the second person. same. of the other.) - There are 5 choices for the third person. - There are 4 choices for the fourth person. - There are 3 choices for the fifth person. - There are 2 choices for the sixth person. "2. Dividetheresultfrom step1by thenumberof rotationsthatare consideredthesame.":5040 /6 = 840 -Thereare6rotationsthatareconsideredthesame. A customer ordered 15 "1. Findthenumberofwaystopackagethechoco- Thereis0way piecesofgourmetchocolate. lates using only boxes of 1 piece.": 1, to package 15 The order can be packaged "2. Findthenumberofwaystopackagethechoco- chocolates us- in small boxes that contain lates using only boxes of 2 pieces.": 1, ingonlyboxes 1, 2 or 4 pieces of chocolate. "3. Findthenumberofwaystopackagethechoco- of 2 pieces. Any box that is used must lates using only boxes of 4 pieces.": 0, be full. Howmany different "4. Findthenumberofwaystopackagethechoco- combinations of boxes can lates using a combination of boxes of 1 and 2 beusedforthecustomer’s15 pieces.": 7, chocolate pie

ngonlyboxes 1, 2 or 4 pieces of chocolate. "3. Findthenumberofwaystopackagethechoco- of 2 pieces. Any box that is used must lates using only boxes of 4 pieces.": 0, be full. Howmany different "4. Findthenumberofwaystopackagethechoco- combinations of boxes can lates using a combination of boxes of 1 and 2 beusedforthecustomer’s15 pieces.": 7, chocolate pieces? One such "5. Findthenumberofwaystopackagethechoco- combination to be included lates using a combination of boxes of 1 and 4 is touse seven2-piece boxes pieces.": 3, and one 1-piece box. "6. Findthenumberofwaystopackagethechoco- lates using a combination of boxes of 2 and 4 pieces.": 2, "7. Findthenumberofwaystopackagethechoco- lates using a combination of boxes of 1, 2, and 4 pieces.": 4, "8. Sum up the total number of ways from the sevencasesabove.": 1+1+0+7+3+2+4=18 How many primes arein the "1. IdentifytherowofPascal’sTrianglethatstarts 6 is not a row of Pascal’s Triangle that witha 1followedbya 6.": "1, 6,15, 20,15, 6,1", prime. startswitha 1 followedbya "2. Listall the numbers in that row.": [1, 6, 15, 20, 6 ? 15, 6, 1], "3. Check each number in the list for primality.": [False, True, False, False, False, True, False], "4. Countthe numberofprimes foundinthe list.": 2 16 SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures Figure11.Casestudyofhuman-writtenstructuresharescommonalitieswithLLM-discoveredreasoningstructure. Weobserve similar reasoning patterns–both structures contain step-wise analysis of each instruction. 17

UnderstandingLLMs:AComprehensiveOverviewfromTraining toInference YihengLiu a,HaoHe a,TianleHan a,XuZhang a,MengyuanLiu a,JiamingTian a, YutongZhang b,JiaqiWang c,XiaohuiGao d,TianyangZhong d,YiPan e,ShaochenXu e, ZihaoWu e,ZhengliangLiu e,XinZhang b,ShuZhang c,XintaoHu d,TuoZhang d, NingQiang a,TianmingLiu eandBaoGe a aSchoolofPhysicsandInformationTechnology,ShaanxiNormalUniversity,Xi’an,710119,Shaanxi,China bInstituteofMedicalResearch,NorthwesternPolytechnicalUniversity,Xi’an,710072,Shaanxi,China cSchoolofComputerScience,NorthwesternPolytechnicalUniversity,Xi’an,710072,Shaanxi,China dSchoolofAutomation,NorthwesternPolytechnicalUniversity,Xi’an,710072,Shaanxi,China eSchoolofComputing,TheUniversityofGeorgia,Athens,30602,USA ARTICLE INFO ABSTRACT Keywords: TheintroductionofChatGPThasledtoasignificantincreaseintheutilizationofLarge LargeLanguageModels LanguageModels(LLMs)foraddressingdownstreamtasks.There’sanincreasingfocuson Training cost-efficienttraininganddeploymentwithinthiscontext.Low-costtraininganddeploymentof Inference LLMsrepresentthefuturedevelopmenttrend.Thispaperreviewstheevolutionoflargelanguage Survey modeltrainingtechniquesandinferencedeploymenttechnologiesalignedwiththisemerging trend.Thediscussionontrainingincludesvariousaspects,includingdatapreprocessing,training architecture,pre-trainingtasks,paralleltraining,andrelevantcontentrelatedtomodelfine- tuning.Ontheinferenceside,thepapercoverstopicssuchasmodelcompression,parallel computation,memoryscheduling,andstructuraloptimization.ItalsoexploresLLMs’utilization andprovidesinsightsintotheirfuturedevelopment. 1. Introduction Languagemodeling(LM)isafundamentalapproachforachievingcognitiveintelligenceinthefieldofnatural languageprocessing(NLP),anditsprogresshasbeennotableinrecentyears[1;2;3].Itassumesacentralrole inunderstanding,generating,andmanipulatinghumanlanguage,servingasthecornerstoneforadiverserangeof NLPapplications[4],includingmachinetranslation,chatbots,sentimentanalysis,andtextsummarization.With theevolutionofdeeplearning,theearlystatisticallanguagemodels(SLM)havegraduallytransformedintoneural languagemodels(NLM)basedonneuralnetworks.Thisshiftischaracterizedbytheadoptionofwordembeddings, representingwordsasdistributedvectors.Notably,thesewordembeddingshaveconsistentlyexcelledinpracticalNLP tasks,profoundlyshapingthefield’sprogress.Pre-trainedlanguagemodels(PLM)representasubsequentphasein theevolutionoflanguagemodelsfollowingNLM.EarlyattemptsatPLMsincludedELMo[5],whichwasbuiltona BidirectionalLSTMarchitecture.However,withtheadventofthetransformerarchitecture[6],characterizedbyparallel self-attentionmechanisms,thepre-trainingandfine-tuninglearningparadigmhaspropelledPLMtoprominenceas theprevailingapproach.Thesemodelsaretypicallytrainedviaself-supervisiononextensivedatasets,cementingtheir statusastheprimarymethodologyinthefield. TheTransformerarchitectureisexceptionallywell-suitedforscalingupmodels,andresearchanalysishasrevealed thatincreasingthemodel’sscaleortrainingdatasizecansignificantlyenhanceitsperformance.Manystudieshave pushedtheboundariesofmodelperformancebycontinuouslyexpandingthescaleofPLM[7;8;9;10].Asmodels growlarger,aremarkablephenomenonknownas"emergence"occurs,whereintheyexhibitastonishingperformance [8].Thesemodelsarecapableofgeneratinghigh-qualitytextandpossessrobustlearningandreasoningabilities.They caneventacklefew-shotlearningtasksthroughin-contextlearning(ICL)[8].Thisremarkablecapabilityenablestheir seamlessapplicationtoawiderangeofdownstreamtasksacrossdiversedomains[11;12;13;14]. ∗Correspond

nystudieshave pushedtheboundariesofmodelperformancebycontinuouslyexpandingthescaleofPLM[7;8;9;10].Asmodels growlarger,aremarkablephenomenonknownas"emergence"occurs,whereintheyexhibitastonishingperformance [8].Thesemodelsarecapableofgeneratinghigh-qualitytextandpossessrobustlearningandreasoningabilities.They caneventacklefew-shotlearningtasksthroughin-contextlearning(ICL)[8].Thisremarkablecapabilityenablestheir seamlessapplicationtoawiderangeofdownstreamtasksacrossdiversedomains[11;12;13;14]. ∗Correspondingauthor ORCID(s): YihengLiuetal.:PreprintsubmittedtoElsevier Page1of30 AComprehensiveOverviewfromTrainingtoInference Pre-trainedlanguagemodels(PLMs)withsignificantlylargerparametersizesandextensivetrainingdataare typicallydenotedasLargeLanguageModels(LLMs)[15;16;17].Themodelsizeusuallyexceeds6-10billion(6- 10B)parameters.AprominentmilestoneinthedevelopmentofLLMsisexemplifiedbytheGPTseries[18;7;8;19]. Notably,OpenAIreleasedChatGPTinNovember2022,markingapivotalmomentintheeraofLLMsandagame- changingmomentinthefieldofartificialintelligence.ChatGPThasempoweredcurrentAIalgorithmstoachieve unprecedentedlevelsofstrengthandeffectiveness,reshapingthewayhumansemployordevelopAIalgorithms. Itsemergencehascapturedtheattentionoftheresearchcommunity.However,owingtoChatGPT’sabsenceasan open-sourceplatform,theprincipalwaytouseChatGPTcurrentlyisbyaccessingitthroughOpenAI’swebsiteat https://chat.openai.comorviatheirAPIinterface.TrainingLLMsthatcanserveasalternativestoChatGPT,or domain-specificLLMs,hasbecomehighlynecessary[20;21;22;23;24;1;25;26].TraininganddeployingLLMs demandexpertiseinhandlinglarge-scaledataandsubstantialpracticalexperienceindistributedparalleltraining [27;28;29].ThisrequirementemphasizestheneedforresearchersdevelopingLLMstopossesssignificantengineering capabilitiesinaddressingthechallengesencounteredduringLLMdevelopment.Researcherswhoareinterestedinthe fieldofLLMsmustpossessengineeringskillsorlearntocollaborateeffectivelywithengineers. Fortheabovereasons,theprimaryobjectiveofthispaperistoprovideacomprehensiveoverviewofLLMstraining andinferencetechniquestoequipresearcherswiththeknowledgerequiredfordeveloping,deploying,andapplying LLMs.Thestructureoftherestofthisreviewisasfollows:InSection2,wewillintroducetherelevantbackground andfoundationalknowledgeofLLMs.InSection3,wewilldelveintothetechnicalaspectsoftrainingLLMs,whilein Section4wewillexplorethetechnologiesrelatedtoLLM’sinferenceanddeployment.InSection5,wewilldiscuss theutilizationofLLMs,andSection6willexplorethefuturedirectionsandtheirimplicationsforLLMs. 2. BackgroundKnowledge 2.1. Transformer Transformerisadeeplearningmodelbasedonanattentionmechanismforprocessingsequencedatathatcan effectivelysolvecomplexnaturallanguageprocessingproblems.Thismodelwasfirstproposedin2017[6],and replacedthetraditionalrecurrentneuralnetworkarchitecture[30]inmachinetranslationtasksasthestate-of-the-art modelatthattime.Duetoitssuitabilityforparallelcomputingandthecomplexityofthemodelitself,Transformer outperformsthepreviouslypopularrecurrentneuralnetworksintermsofaccuracyandperformance.TheTransformer architectureconsistsprimarilyoftwomodules,anEncoderandaDecoder,aswellastheattentionmechanismwithin thesemodules. 2.1.1. Self-Attention Self-AttentionStructure[6]: Essentially,theattentionmechanismaimsatselectingasmallamountofimportant informationfromalargeamountofdataandfocusingontheseimportantpieceswhileignoringthemajorityof unimportantinformation.Theself-attentionmechanism,asavariantoftheattentionmechanism,reducesrelianceon externalinformationandexcelsatcapturinginternalcorrelationswithindataorfeatures.Applyingtheself-attention mechanismintext-primarilyinvolvescalculatingthemutualinfluencebetweenwordstoaddresstheissueoflong-range dependencies.Additionally,self-attentionisthecoreideabehindtransformers.Thecoreformulaforkey-valueattention isasfollows: 𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛 (𝑄,𝐾,𝑉 ) = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 (𝑄𝐾 𝑇√𝑑 𝑘)𝑉 (1) Self-attentionallowsthe

heattentionmechanism,reducesrelianceon externalinformationandexcelsatcapturinginternalcorrelationswithindataorfeatures.Applyingtheself-attention mechanismintext-primarilyinvolvescalculatingthemutualinfluencebetweenwordstoaddresstheissueoflong-range dependencies.Additionally,self-attentionisthecoreideabehindtransformers.Thecoreformulaforkey-valueattention isasfollows: 𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛 (𝑄,𝐾,𝑉 ) = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 (𝑄𝐾 𝑇√𝑑 𝑘)𝑉 (1) Self-attentionallowsthemodeltoweightheimportanceofdifferentwordsinasentencewhenpredictingaparticular word.Itcalculatesaweightedsumofthevaluesofallwordsinthesentence,wheretheweightsaredeterminedbythe relevanceofeachwordtothetargetword. Theself-attentionmechanismconsistsofthreesteps:calculatingthequery,key,andvaluevectors.Thequeryvector representsthewordbeingattendedto,whilethekeyvectorsrepresentallthewordsinthesentence.Thevaluevectors storetheinformationassociatedwitheachword.Theattentionweightsarecomputedbytakingthedotproductbetween thequeryandkeyvectors,followedbyasoftmaxoperationtoobtainadistributionoverthewords. Multi-HeadAttention[6]: Multi-headself-attentionextendstheself-attentionmechanismbyperformingit multipletimesinparallel.Eachattentionheadlearnstofocusondifferentaspectsoftheinput,capturingdifferent dependenciesandpatterns.Theoutputsoftheattentionheadsarethenconcatenatedandlinearlytransformedtoobtain YihengLiuetal.:PreprintsubmittedtoElsevier Page2of30 AComprehensiveOverviewfromTrainingtoInference thefinalrepresentation.Byusingmultipleattentionheads,themodelcancapturebothlocalandglobaldependencies, allowingforamorecomprehensiveunderstandingoftheinputsequence.Thisparallelizationalsoenhancesthemodel’s capacitytocapturecomplexrelationshipsbetweenwords.TheMulti-headattentioncanbeformulatedasfollows: 𝑀𝑢𝑙𝑡𝑖𝐻𝑒𝑎𝑑𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛 (𝑄,𝐾,𝑉 ) = 𝐶𝑜𝑛𝑐𝑎𝑡 [ℎ𝑒𝑎𝑑 1,… ,ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑜 𝑤ℎ𝑒𝑟𝑒ℎ𝑒𝑎𝑑 𝑖 = 𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛 (𝑄𝑊 𝑄𝑖 ,𝐾𝑊 𝐾𝑖 ,𝑉𝑊 𝑉𝑖 ) (2) Inthiscase,"𝐶𝑜𝑛𝑐𝑎𝑡 "meanstoconcatenatetheattentioncalculationresultsofeachhead,"𝑊 𝑜"istheweight matrixoftheoutputlayer,usedtolinearlytransformtheconcatenatedresults.Thisyieldstheoutputofmulti-head attention.Insummary,multi-headattentionenhancesthemodel’sabilitytorepresentinputsequencesbyperforming parallelattentioncalculationsunderdifferentlineartransformations,thenconcatenatingandlinearlytransformingthe results.ThismechanismplaysanimportantroleintheTransformermodel,helpingtohandlelong-rangedependencies andimprovemodelperformance. 2.1.2. Encoder Theencodermodule[6]oftheTransformermodeliscomposedofmultipleidenticallayers,eachofwhichincludes amulti-headattentionmechanismandfeed-forwardneuralnetwork[31].Inthemulti-headattentionmechanism,each positionintheinputsequenceiscalculatedforattentionwithotherpositionstocapturethedependenciesbetween differentpositionsintheinputsequence.Thefeed-forwardneuralnetworkisthenusedtofurtherprocessandextract featuresfromtheoutputoftheattentionmechanism.Theencodermodulegraduallyextractsfeaturesoftheinput sequencethroughthestackingofmultiplesuchlayersandpassesthefinalencodingresulttothedecodermodulefor decoding.Thedesignoftheencodermoduleenablesittoeffectivelyhandlelong-rangedependencieswithintheinput sequenceandhassignificantlyimprovedperformanceinvariousNLPtasks. 2.1.3. Decoder Thedecodermodule[32]oftheTransformermodelisalsocomposedofmultipleidenticallayers,eachofwhich includesamulti-headattentionmechanismandafeed-forwardneuralnetwork.Unliketheencoder,thedecoderalso includesanadditionalencoder-decoderattentionmechanism,usedtocomputeattentionontheinputsequenceduring thedecodingprocess.Ateachposition,thedecodercanonlyperformself-attentioncalculationswiththepositions beforeittoensurethatthegenerationofthesequencedoesnotviolategrammarrules.Masksplayanimportantrole inthedecoder,ensuringthatonlyinformationbeforethecurrenttimestepisfocusedonwhengeneratingtheoutput sequence,andnotleakinginform

-headattentionmechanismandafeed-forwardneuralnetwork.Unliketheencoder,thedecoderalso includesanadditionalencoder-decoderattentionmechanism,usedtocomputeattentionontheinputsequenceduring thedecodingprocess.Ateachposition,thedecodercanonlyperformself-attentioncalculationswiththepositions beforeittoensurethatthegenerationofthesequencedoesnotviolategrammarrules.Masksplayanimportantrole inthedecoder,ensuringthatonlyinformationbeforethecurrenttimestepisfocusedonwhengeneratingtheoutput sequence,andnotleakinginformationfromfuturetimesteps.Specifically,thedecoder’sself-attentionmechanism usesmaskstopreventthemodelfromaccessingfutureinformationwhengeneratingpredictionsateachtimestep, maintainingthecausalityofthemodel.Thisensuresthattheoutputgeneratedbythemodeldependsontheinformation atthecurrenttimestepandbefore,withoutbeinginfluencedbyfutureinformation. 2.1.4. PositionalEmbedding Positionandorderarecrucialforcertaintasks,suchasunderstandingasentenceoravideo.Positionandorder definethegrammarofasentence,theyareintegraltothesemanticsofsentences.TheTransformerutilizesMulti-Head Self-Attention(MHSA)toavoidtherecursiveapproachofRNN,thusspeedingupthetrainingprocess.Additionally, itcancapturelong-rangedependenciesinsentencesandhandlelongerinputs.Wheneachtokeninasentencepasses throughtheTransformer’sEncoder/Decoderstack,themodelitselflacksanysenseofposition/orderforeachtoken (permutationinvariance).Therefore,amethodisstillneededtoincorporatethesequentialinformationoftokensinto themodel.Toenablethemodeltoperceivetheinputsequence,positionalinformationaboutthelocationofeach tokeninthesentencecanbeadded,andthistechniqueisknownaspositionalembedding(PE).whichisusedinthe Transformermodeltoincorporatethesequentialorderoftokensintotheinputrepresentation.SincetheTransformer doesnothaverecurrentconnections,itlackstheinherentnotionoftokenorderpresentinrecurrentneuralnetworks. Toaddressthis,positionalembeddingassignsauniquevectortoeachtokenpositionintheinputsequence.These positionalembeddingsareaddedtothewordembeddingbeforebeingfedintothemodel.Byincludingpositional information,themodelcandifferentiatebetweentokensbasedontheirpositioninthesequence.IntheTransformer model,thecoreformulaofthepositionembeddingcanbeexpressedas: 𝑃𝐸 (𝑝𝑜𝑠, 2𝑖) = 𝑠𝑖𝑛 ( 𝑝𝑜𝑠 10000 ( 2𝑖𝑑𝑚𝑜𝑑𝑒𝑙 )) (3) YihengLiuetal.:PreprintsubmittedtoElsevier Page3of30 AComprehensiveOverviewfromTrainingtoInference 𝑃𝐸 (𝑝𝑜𝑠, 2𝑖+1) = 𝑐𝑜𝑠 ( 𝑝𝑜𝑠 10000 ( 2𝑖𝑑𝑚𝑜𝑑𝑒𝑙 )) (4) Inthisequation,𝑃𝐸 representsthepositionembeddingmatrix,𝑝𝑜𝑠 representsthepositionofatokeninthesentence, 𝑖representsthedimensionindexofthepositionembedding,and𝑑 𝑚𝑜𝑑𝑒𝑙 representsthehiddenlayerdimensionofthe Transformermodel.Byusingsineandcosinefunctionsandperformingdifferentcalculationsontheposition(pos)and dimension(i),thisformulageneratesuniquepositionembeddingvaluesforeachpositionanddimension.Asaresult, eachtokenisassignedauniquepositionembeddingvector,allowingthemodeltoperceivethesequentialinformationof tokensinthesentence.Inpracticalapplications,thepositionembeddingmatrixisaddedtotheinputwordembedding matrixtocombinepositioninformationandsemanticinformation,therebyprovidingamorecomprehensiveinput representationfortheTransformermodel. TwocommonlyusedpositionalencodingmethodsinTransformerareAbsolutePositionalEncodingandRelative PositionalEncoding. (1)AbsolutePositionalEncoding:Itgeneratesuniquepositionalembeddingvaluesforeachpositionanddimension byusingsineandcosinefunctions.Thismethodusessineandcosinefunctionsinthementionedformulatocalculate thepositionalembeddingvaluesandaddsthemtothewordembeddings.AbsolutePositionalEncodingprovidesa uniqueencodingforeachposition,enablingthemodeltoperceivethesequentialinformationofwordsinthesentence. (2)RelativePositionalEncoding:Itisanencodingmethodbasedonrelativepositionalrela

Relative PositionalEncoding. (1)AbsolutePositionalEncoding:Itgeneratesuniquepositionalembeddingvaluesforeachpositionanddimension byusingsineandcosinefunctions.Thismethodusessineandcosinefunctionsinthementionedformulatocalculate thepositionalembeddingvaluesandaddsthemtothewordembeddings.AbsolutePositionalEncodingprovidesa uniqueencodingforeachposition,enablingthemodeltoperceivethesequentialinformationofwordsinthesentence. (2)RelativePositionalEncoding:Itisanencodingmethodbasedonrelativepositionalrelationships.Relative PositionalEncodingrepresentspositionalinformationbycalculatingtherelativedistancesbetweenwords.Thismethod isusedinmodelslikeTransformer-XL[33],andRelativePositionalEncodingcanbettercapturetherelativepositional relationshipsbetweenwordswhendealingwithlongsequences.Bothofthesepositionalencodingmethodsaimto providethepositionalinformationofwordsintheinputsequencetotheTransformermodel,enablingthemodelto bettercomprehendandprocesssequentialdata.Thespecificchoiceofpositionalencodingmethoddependsonthe specificapplicationscenarioandmodeldesign. Therearealsootherpositionalencodingmethodsappliedtoothermodels,suchasRoPE[34]andALiBi[35]. RoPEisamethodthatusesAbsolutePositionalEncodingtorepresentRelativePositionalEncodingandisapplied inthedesignoflargelanguagemodelslikePaLM[36],LLaMA[9],andGLM-130B[37]. ALiBidoesnotaddpositionalembeddingstowordembeddingsbutinsteadaddsapre-definedbiasmatrixtothe attentionscorebasedonthedistancebetweentokens.ItisappliedinthedesignoflargelanguagemodelslikeBLOOM [38]. Someotherpositionalencodingmethods,suchasmixedpositionalencoding,multi-digitpositionalencoding,and implicitpositionalencoding,arealsousedbysomemodels. 2.2. PromptLearning Promptlearningservesasawidelyadoptedmachinelearningapproach,particularlyinthefieldofNLP.Atits core,thismethodologyinvolvesguidingamodeltoproducespecificbehaviorsoroutputsthroughthecarefuldesign ofpromptstatements.Itiscommonlyemployedtofine-tuneandguidepre-trainedLLMsforexecutingparticular tasksorgeneratingdesiredresults.Researchershaveobservedthatthedesignofspecificpromptstatementscansteer pre-trainedmodelstoperformvarioustasks,suchasquestion-answering,textgeneration,andsemanticunderstanding [39;40;41;42;43;44;45;46;47;48;49;50].Thestrengthofthisapproachliesinitsabilitytoadapttodifferenttasks throughsimplemodificationstopromptstatements,eliminatingtheneedforretrainingtheentiremodel.ForLLMs liketheGPTseriesandotherpre-trainedmodels,promptlearningprovidesastraightforwardandpowerfulmeans formodelfine-tuning.Bysupplyingappropriateprompts,researchersandpractitionerscancustomizethemodel’s behavior,makingitmoresuitableforspecificdomainsortaskrequirements.Inshort,promptlearningisamachine learningapproachthat,buildsuponpre-trainedlanguagemodels,andguidesthemodeltoperformvarioustasksthrough thedesignofpromptstatements,offeringincreasedflexibilityforcustomizingmodelapplications.InthisSection,we willintroducethebasicknowledgeofpromptlearning. 2.2.1. BackgroundandOverview Promptlearningisanewapproachtomachinelearning[51].Intheearlyfieldofnaturallanguageprocessing (NLP),researchersmainlyusedfullysupervisedlearningmode[52],whichtrainedmodelsforspecifictasksonthe inputandoutputexampledatasetofthetargettask.However,duetothelimitedtrainingdataset,thismethodcannot YihengLiuetal.:PreprintsubmittedtoElsevier Page4of30 AComprehensiveOverviewfromTrainingtoInference trainhigh-qualitymodelswell,soearlyNLPreliedmoreonfeatureengineering;Withtheemergenceofneuralnetwork modelsandtheiruseinthefieldofNLP,peoplehavebeguntopayattentiontoarchitectureengineering[53]. However,between2017and2019,thelearningapproachofNLPmodelsshiftedfromfullysupervisedlearningto anewmode:pre-trainandfine-tuneparadigm[54].Inthisparadigm,amodelwithafixedarchitectureispre-trained asalanguagemodeltopredicttheprobabilityofobservedtextdata.Duetotheabundantrawtextdatarequiredfor traininglanguagemodels,theselanguagemodelscanbetrainedonlargedatasets.Duringthisprocess,languagemodels canlearnrobustuniversalfe

elsandtheiruseinthefieldofNLP,peoplehavebeguntopayattentiontoarchitectureengineering[53]. However,between2017and2019,thelearningapproachofNLPmodelsshiftedfromfullysupervisedlearningto anewmode:pre-trainandfine-tuneparadigm[54].Inthisparadigm,amodelwithafixedarchitectureispre-trained asalanguagemodeltopredicttheprobabilityofobservedtextdata.Duetotheabundantrawtextdatarequiredfor traininglanguagemodels,theselanguagemodelscanbetrainedonlargedatasets.Duringthisprocess,languagemodels canlearnrobustuniversalfeaturesofthelanguagetheyaremodeling.Then,byintroducingadditionalparametersand fine-tuningthemusingtask-specificobjectivefunctions,thePLMmentionedabovewilladapttodifferentdownstream tasks.Atthispoint,thefocusofresearchshiftedtoobjectiveengineering,whichistodesigntrainingobjectivesduring pre-trainingandfine-tuning.SinceBERT,NLPhasbeenusingpre-trainingandfine-tuningmethodsforalongperiod oftime,butthisapproachrequiresanewmodeltobefine-tunedforeachtaskandcannotbeshared.ButforanLLM, itfeelslikecustomizingeachtask,whichisveryinefficient[51]. Promptlearning,thismethodhasdemonstratedamazingcapabilitiesinGPT-3.TheGPT-3modelcanhandle manytaskswithonlyafewsamplesbyusingnaturallanguagepromptsandtaskdemonstrationsascontext,without updatingparametersintheunderlyingmodel.PromptLearningreplacestheprocessofpre-trainedandfine-tuning withpre-trained,promptsandpredictions.Inthisparadigm,thedownstreamtaskisnottoadaptthepre-trainedLM tothedownstreamtaskthroughobjectiveengineering,buttoredefinethedownstreamtaskwiththehelpoftext prompts,makingitlookmorelikethetaskssolvedduringtheoriginalLMtraining.Forpromptlearning,itisonly necessarytoinsertdifferentpromptparameterstoadapttodifferenttasks.Thatistosay,eachtaskonlyneedstotrain thepromptparameterseparately,withouttheneedtotraintheentirepre-trainedlanguagemodel[55].Thisapproach greatlyimprovestheefficiencyofusingpre-trainedlanguagemodelsandsignificantlyshortenstrainingtime. 2.2.2. BasiccomponentsandprocessofPromptlearning Inthetraditionalpre-trained+fine-tuningparadigm,thereisagapbetweenthepre-trainedstageanddownstream tasks[51],whilepromptlearningcanmaintainconsistencybetweenthepre-trainedtargetformatanddownstreamtask outputformat,thatis,aligntheformofdownstreamtaskswiththeformofPLMspre-trainedtasks.Whentraining PLMs,wecantransformtheoriginaltargettaskintoafill-in-the-blankorcontinuationtasksimilartothepre-trained taskofPLMsbyconstructingaprompt.Theadvantageofthismethodisthatthroughaseriesofappropriateprompts, wecanuseasinglelanguagemodeltosolvevariousdownstreamtasks. Promptlearningoptimizestheperformanceofmodelsondifferenttasksbyusingpre-trainedmodelsanddesigning appropriatetemplates.Promptlearningconsistsofprompttemplates,answermappings,andpre-trainedlanguage models.Theprompttemplateisthemainbodyoftheprompt,andfillintheblank[56]andgeneratebasedon prefix[57]aretwocommontypesofpromptlearningtemplates.Thefill-in-the-blanktemplateselectsoneormore positionsinthetextandrepresentsthemwith[MASK]tags,usedtopromptthemodeltofillinthecorresponding words;Prefix-basedtemplategenerationinvolvesaddingaspecificprefixbeforeasentencetoguidethemodel ingeneratingappropriatetext.Answermappingistheprocessofevaluatingallpossibleanswersaccordingtoa probabilitydistribution,selectingthemostlikelyanswerasthepredictedoutput,andconvertingitintoappropriate categorymappingwords.Thisprocesstypicallyinvolvesconvertinglabelsintonaturallanguagevocabulary,known asVerbalizer[58]. TheworkflowofPromptlearningmainlyincludesthefollowingfourparts: (1)UsePLMsasbaseencoders (2)Addadditionalcontext(template)witha[MASK]position (3)Projectlabelstolabelwords(verbalizer) (4)BetheGAPbetweenpre-trainingandfine-tuning Afterdefiningthetemplateandanswerspace,weneedtochooseasuitablepre-trainedlanguagemodel.Thereare nowvariouspre-trainedmodels(PTMs)withgoodperformance,andwhenselectingamodel,oneusuallyconsidersits paradigm,suchasAutorecursive,MaskedLanguageModeling,EncoderDecoder,etc.Basedonthis,forthesummary task,amoresuitableBidirectionalandAuto-RegressiveTransformers(BART)modelcanbeselected. Theselectionofatemplateplays

sition (3)Projectlabelstolabelwords(verbalizer) (4)BetheGAPbetweenpre-trainingandfine-tuning Afterdefiningthetemplateandanswerspace,weneedtochooseasuitablepre-trainedlanguagemodel.Thereare nowvariouspre-trainedmodels(PTMs)withgoodperformance,andwhenselectingamodel,oneusuallyconsidersits paradigm,suchasAutorecursive,MaskedLanguageModeling,EncoderDecoder,etc.Basedonthis,forthesummary task,amoresuitableBidirectionalandAuto-RegressiveTransformers(BART)modelcanbeselected. Theselectionofatemplateplaysaveryimportantroleinthepromptlearning.Templatescangenerallybe distinguishedbasedonwhethertheyaremanuallyspecified:artificiallyconstructedtemplatesorautomaticallysearched templates.Artificiallycreatedtemplatesarethemostintuitivemethod,easytounderstand,andhavegoodperformance inpracticalapplications.However,artificiallyconstructedtemplatesalsohavesomedrawbacks:priorknowledgeis requiredwhendesigningtemplatesmanually[59],andtheremaybefailures[60].Therearetwotypesofautomatically YihengLiuetal.:PreprintsubmittedtoElsevier Page5of30 AComprehensiveOverviewfromTrainingtoInference generatedtemplates:discretepromptsandcontinuousprompts.Discretepromptsallowthemodeltoselecttheoptimal templateinasetofdiscretetemplatespaces,whilecontinuouspromptsallowthelanguagemodeltoautomatically trainaprompt.Accordingtoresearch,usingmultipletemplates[61]canimprovetheperformanceofthemodel.The simplestwaytochoosetousemultipletemplatesandaggregatethemtogethertocompleteanansweristotakethe average[60]orweightedaverageofeachtemplateoutput[58]. Verbalizeristheprocessofmappinglabelstolabelwords,andtheselectionofverbalizersisalsocrucialforprompt learning.Therearetwowaystoconstructaverbalizer:manualdefinitionandautomaticsearch.Themanualdefinition requiresprofessionalknowledgeandmayhavedisadvantagessuchasstrongsubjectivityandasmallcoveragearea.To solvethisproblem,wecanchoosethefollowingsolutions:(1)Manuallydesignwithhumanpriorknowledge;(2)Start withanIntellabelword,paraphraseandexpand;(3)Startwithaninternallabelword,usingexternalknowledgeand expand;(4)Decomposethelabelwithmultipletokens;(5)Virtualtokenandoptimizethelabelembedding.Inaddition, wecanuseexternalknowledgebasestoexpandandimprovelabelwords,therebyachievingbettertextclassification results[62]. 2.2.3. learningstrategy TheemergenceofthenewparadigmofPromptlearninghasbroughtsignificantchangestothetrainingprocess. ThelearningstrategiesforPromptlearningmainlyincludethefollowing:(1)Pre-trainingthenfine-tuning,whichisa traditionalpre-training+finetuningmethod[63];(2)Tuningfreepromotion,relyingonthedesignerLMofpromptsto directlyprovideanswers[64];(3)FixedLMprompttuning,whichupdatestherelevantparametersofpromptsusing downstreamtasktrainingdata;(4)FixpromptLMtuning,thisstrategyistofine-tunetheparametersofLM,which havefixedparameterswhenusingprompts;(5)Prompt+LMtuningisastrategythatupdatesbothpromptsrelated parametersandLMparameters. Thesedifferentlearningstrategiescanbeselectedbasedonspecifictasksandneeds.Pre-training+fine-tuningis themostcommonstrategy,suitableformosttasks[63].Nofine-tuningpromptsaresuitableforsimpletasks,which cangreatlyreducetrainingtimeandcomputationalresourceconsumption.FixedLMpromptfine-tuningandfixed promptLMfine-tuningaresuitablefortasksthatrequiremoreprecisecontrolandcanoptimizemodelperformance byadjustingpromptparametersorlanguagemodelparameters.CombiningpromptsandLMfine-tuningcombinesthe advantagesofbothandcanfurtherimprovemodelperformance[51]. Insummary,Promptlearningprovidesuswithanewtrainingparadigmthatcanoptimizemodelperformance onvariousdownstreamtasksthroughappropriatepromptdesignandlearningstrategies.Choosingtheappropriate template,constructinganeffectiveverbalizer,andadoptingappropriatelearningstrategiesareallimportantfactorsin improvingtheeffectivenessofpromptlearning. 3. TrainingofLargeLanguageModels ThetrainingofLLMscanbebroadlydividedintothreesteps.Thefirststepinvolvesdatacollectionandprocessing. Thesecondstepencompassesthepre-trainingprocess,whichincludesdeterminingthemodel’

ainingparadigmthatcanoptimizemodelperformance onvariousdownstreamtasksthroughappropriatepromptdesignandlearningstrategies.Choosingtheappropriate template,constructinganeffectiveverbalizer,andadoptingappropriatelearningstrategiesareallimportantfactorsin improvingtheeffectivenessofpromptlearning. 3. TrainingofLargeLanguageModels ThetrainingofLLMscanbebroadlydividedintothreesteps.Thefirststepinvolvesdatacollectionandprocessing. Thesecondstepencompassesthepre-trainingprocess,whichincludesdeterminingthemodel’sarchitectureandpre- trainingtasksandutilizingsuitableparalleltrainingalgorithmstocompletethetraining.Thethirdstepinvolvesfine- tuningandalignment.Inthissection,wewillprovideanoverviewofthemodeltrainingtechniques.Thiswillincludean introductiontotherelevanttrainingdatasets,datapreparationandpreprocessing,modelarchitecture,specifictraining methodologies,modelevaluation,andcommonlyusedtrainingframeworksforLLMs. 3.1. DataPreparationandPreprocessing 3.1.1. Dataset TrainingLLMsrequirevastamountsoftextdata,andthequalityofthisdatasignificantlyimpactsLLM performance.Pre-trainingonlarge-scalecorporaprovidesLLMswithafundamentalunderstandingoflanguageand somegenerativecapability.ThefirststepinLLMtrainingiscollectingsubstantialcorporaofnaturallanguagetext. Pre-trainingdatasourcesarediverse,commonlyincorporatingwebtext,conversationaldata,andbooksasgeneral pre-trainingcorpora.Additionally,someresearcheffortsintroducespecializeddatafromprofessionaldomains,such ascodeorscientificdata,toenhanceLLMcapabilitiesinthosefields.LeveragingdiversesourcesoftextdataforLLM trainingcansignificantlyenhancethemodel’sgeneralizationcapabilities.Inthefollowingsection,wewillpresent thecommonlyuseddatasetsfortrainingLLMsasshowninTable1.Thesecorporaarecategorizedinto5groupsfor discussion. YihengLiuetal.:PreprintsubmittedtoElsevier Page6of30 AComprehensiveOverviewfromTrainingtoInference Table 1 Commonlyusedcorporainformation. Corpora Type Links BookCorpus[65] Books https://github.com/soskek/bookcorpus Gutenberg[66] Books https://www.gutenberg.org Books1[8] Books Notopensourceyet Books2[8] Books Notopensourceyet CommonCrawl[67] CommonCrawl https://commoncrawl.org C4[68] CommonCrawl https://www.tensorflow.org/datasets/catalog/c4 CC-Stories[69] CommonCrawl Notopensourceyet CC-News[70] CommonCrawl https://commoncrawl.org/blog/news-dataset-available RealNews[71] CommonCrawl https://github.com/rowanz/grover/tree/master/realnews RefinedWeb[72] CommonCrawl https://huggingface.co/datasets/tiiuae/falcon-refinedweb WebText RedditLink Notopensourceyet OpenWebText[73] RedditLink https://skylion007.github.io/OpenWebTextCorpus/ PushShift.io[74] RedditLink https://pushshift.io/ Wikipedia[75] Wikipedia https://dumps.wikimedia.org/zhwiki/latest/ BigQuery[76] Code https://cloud.google.com/bigquery CodeParrot Code https://huggingface.co/codeparrot thePile[77] Other https://github.com/EleutherAI/the-pile ROOTS[78] Other https://huggingface.co/bigscience-data Books:TwocommonlyutilizedbooksdatasetsforLLMstrainingareBookCorpus[65]andGutenberg[66].These datasetsincludeawiderangeofliterarygenres,includingnovels,essays,poetry,history,science,philosophy,andmore. Widelyemployedbynumerous

https://huggingface.co/codeparrot thePile[77] Other https://github.com/EleutherAI/the-pile ROOTS[78] Other https://huggingface.co/bigscience-data Books:TwocommonlyutilizedbooksdatasetsforLLMstrainingareBookCorpus[65]andGutenberg[66].These datasetsincludeawiderangeofliterarygenres,includingnovels,essays,poetry,history,science,philosophy,andmore. WidelyemployedbynumerousLLMs[9;79],thesedatasetscontributetothemodels’trainingbyexposingthemtoa diversearrayoftextualgenresandsubjectmatter,fosteringamorecomprehensiveunderstandingoflanguageacross variousdomains. CommonCrawl:CommonCrawl[67]managesanaccessiblerepositoryofwebcrawldata,freelyavailablefor utilizationbyindividualsandorganizations.Thisrepositoryencompassesavastcollectionofdata,comprisingover 250billionwebpagesaccumulatedoveraspanof16years.Establishedin2007,CommonCrawlhasevolvedintoa widelyrecognizedandreferencedcorpusintheacademicandresearchcommunities,citedinmorethan10,000research papers.Thiscontinuouslyexpandingcorpusisadynamicresource,withanadditionof3–5billionnewwebpageseach month.Itssignificanceextendstothefieldofnaturallanguageprocessing,whereitservesasaprimarytrainingcorpus innumerouslargelanguagemodels.Notably,asubstantialportionoftherawtokensemployedintrainingGPT-3 [8],amountingto82%,issourcedfromtheCommonCrawl.However,duetothepresenceofasubstantialamountof low-qualitydatainwebarchives,preprocessingisessentialwhenworkingwithCommonCrawldata.Currently,four commonlyusedfiltereddatasetsbasedonCommonCrawlareavailable:C4[68],CC-Stories[69],CC-News[70],and RealNews[71]. RedditLinks:Redditisasocialmediaplatformwhereuserscansubmitlinksandposts,andotherscanvoteon themusingthe"upvote"or"downvote"system.Thischaracteristicmakesitavaluableresourceforcreatinghigh-quality datasets. Wikipedia:Wikipedia[75],afreeandopenonlineencyclopediaproject,hostsavastrepositoryofhigh-quality encyclopediccontentspanningawidearrayoftopics.TheEnglishversionofWikipediaisextensivelyutilizedinthe trainingofmanyLLMs[8;9;80],servingasavaluableresourceforlanguageunderstandingandgenerationtasks. Additionally,Wikipediaisavailableinmultiplelanguages,providingdiverselanguageversionsthatcanbeleveraged fortraininginmultilingualenvironments. Code:Thereisalimitedavailabilityofpubliclyaccessiblecodedatasetsatpresent.Existingeffortsprimarily involvewebscrapingofcodewithopen-sourcelicensesfromtheinternet.ThemainsourcesincludeGithubandStack Overflow. WehaveorganizeddatasetsutilizedbydistinctLLMs.Duringthetrainingprocess,LLMsaretypicallytrainedon multipledatasets,asspecifiedinTable2forreference. YihengLiuetal.:PreprintsubmittedtoElsevier Page7of30 AComprehensiveOverviewfromTrainingtoInference Table 2 DatasetsutilizedbydistinctLLMs LLMs Datasets GPT-3[8] CommonCrawl[67],WebText2[8],Books1[8],Books2[8],Wikipedia[75] LLaMA[9] CommonCrawl[67],C4[68],Wikipedia[75],Github,Books,Arxiv,StackExchange PaLM[36] SocialMedia,Webpages,Books,Github,Wikipedia,News(total780Btokens) T5[68] C4[68],WebText,Wikipedia,RealNews CodeGen[81] thePile,BIGQUERY,BIGPYTHON CodeGeeX[82] CodeParrot,thePile,Github GLM[37] BooksCorpus,Wikipedia BLOOM[38] ROOTS OPT[83] BookCorpus,CCNews,CC-Stories,thePile,Pushshift.io 3.1.2. Datapreprocessing Onceanadequatecorpusofdataiscollected,thesubsequentstepisdatapreprocessing.Thequalityofdata preprocessingdirectlyimpactsthemodel’sperformanceandsecurity.Thespecificpreprocessingstepsinvolvefiltering low-qualitytext,includingeliminatingtoxicandbiasedcontenttoensurethemodelalignswithhumanethical standards.Italsoincludesdeduplication,removingduplicatesint

38] ROOTS OPT[83] BookCorpus,CCNews,CC-Stories,thePile,Pushshift.io 3.1.2. Datapreprocessing Onceanadequatecorpusofdataiscollected,thesubsequentstepisdatapreprocessing.Thequalityofdata preprocessingdirectlyimpactsthemodel’sperformanceandsecurity.Thespecificpreprocessingstepsinvolvefiltering low-qualitytext,includingeliminatingtoxicandbiasedcontenttoensurethemodelalignswithhumanethical standards.Italsoincludesdeduplication,removingduplicatesinthetrainingset,andexcludingredundantcontent inthetestsettomaintainthesampledistributionbalance.Privacyscrubbingisappliedtoensurethemodel’ssecurity, preventinginformationleakageorotherprivacy-relatedconcerns.Additionally,iffine-tuningLLMsisconsidered, expandingthevocabularyshouldalsobeconsidered.Ontheotherhand,LLaMA2models[10]representanotable exception.Thesemodelsforegofilteringintheirpretrainingcorpus,asaggressivefiltrationmightaccidentallyfilter outsomedemographicgroups.ThisapproachenhancesthegeneralizabilityofthebaseLLaMA2models,making themmoreadeptacrossarangeofdownstreamtasks,suchashatespeechdetectionandprivacyde-identification. Observationsindicatethatabstainingfromadditionalfilteringinthepretrainingdataenablesthebasemodeltoachieve reasonablesafetyalignmentwithfewerexamples[10].Whilethisincreasesbothgeneralizabilityandsafetyalignment efficiency,theimplementationofadditionalsafetymitigationsisstillimperativepriortopublicdeployment,asfurther discussedinSection3.5.4. Qualityfiltering:Filteringlow-qualitydataistypicallydoneusingheuristic-basedmethodsorclassifier-based methods.Heuristicmethodsinvolveemployingmanuallydefinedrulestoeliminatelow-qualitydata[84;72].For instance,rulescouldbesettoretainonlytextcontainingdigits,discardsentencescomposedentirelyofuppercase letters,andremovefileswithasymbolandwordratioexceeding0.1,andsoforth.Classifier-basedmethodsinvolve trainingaclassifieronahigh-qualitydatasetsuchasWebText[85]tofilteroutlow-qualitydatasets. Deduplication:Languagemodelsmaysometimesrepetitivelygeneratethesamecontentduringtextgeneration, potentiallyduetoahighdegreeofrepetitioninthetrainingdata.Extensiverepetitioncanleadtotraininginstability, resultinginadeclineintheperformanceofLLMs[86].Additionally,itiscrucialtoconsideravoidingdataset contaminationbyremovingduplicateddatapresentinboththetrainingandtestingset[87]. Privacyscrubbing:LLMs,astext-generatingmodels,aretrainedondiversedatasets,whichmayposeprivacy concernsandtheriskofinadvertentinformationdisclosure[88].Inthepreprocessingphaseoflanguagedatasets,itis imperativetoaddressprivacyconcernsbysystematicallyremovinganysensitiveinformation.Thisinvolvesemploying techniquessuchasanonymization,redaction,ortokenizationtoeliminatepersonallyidentifiabledetails,geolocation, andotherconfidentialdata.Bycarefullyscrubbingthedatasetofsuchsensitivecontent,researchersanddeveloperscan ensurethatthelanguagemodelstrainedonthesedatasetsupholdprivacystandardsandmitigatetheriskofunintentional disclosureofprivateinformation.Itisessentialtostrikeabalancebetweendatautilityandprivacyprotection,fostering responsibleandethicaluseoflanguagedatasetsinvariousapplications. Filteringouttoxicandbiasedtext:Inthepreprocessingstepsoflanguagedatasets,acriticalconsiderationisthe removaloftoxicandbiasedcontenttoensurethedevelopmentoffairandunbiasedlanguagemodels.Thisinvolves implementingrobustcontentmoderationtechniques,suchasemployingsentimentanalysis,hatespeechdetection,and biasidentificationalgorithms.Byleveragingthesetools[89],researcherscansystematicallyidentifyandfilterouttext thatmayperpetuateharmfulstereotypes,offensivelanguage,orbiasedviewpoints. YihengLiuetal.:PreprintsubmittedtoElsevier Page8of30 AComprehensiveOverviewfromTrainingtoInference 3.2. Architecture Currently,allLLMsarebuiltupontheTransformerarchitecture,allowingthesemodelstoscaletoseveral10billion orevenatrillionparameters.Typically,PLMarchitecturesfallintothreecategories:Encoder-only[90],Encoder- decoder[68]andDecoder-only[1

t thatmayperpetuateharmfulstereotypes,offensivelanguage,orbiasedviewpoints. YihengLiuetal.:PreprintsubmittedtoElsevier Page8of30 AComprehensiveOverviewfromTrainingtoInference 3.2. Architecture Currently,allLLMsarebuiltupontheTransformerarchitecture,allowingthesemodelstoscaletoseveral10billion orevenatrillionparameters.Typically,PLMarchitecturesfallintothreecategories:Encoder-only[90],Encoder- decoder[68]andDecoder-only[18].TheEncoder-onlyarchitectureisnolongeremployedinthelatestLLMsand won’tbefurtherdiscussedhere.Instead,thissectionwillfocusonintroducingtheEncoder-decoderandDecoder-only architectures. Figure 1:ThefiguresfromlefttorightrepresenttheEncoder-decoderarchitecture,CausalDecoderarchitecture,Prefix Decoderarchitecture,andtheirmaskconfigurations,respectively.Thisdiagramillustratestherangeoftokensthateach inputtokencanattendto. 3.2.1. Encoder-decoderArchitecture TheEncoder-decoderarchitectureofLLMsisbuiltuponthetraditionalTransformerEncoder-decoderarchitecture. TheEncoder-decoderarchitectureconsistsoftwomaincomponents:theEncoderandtheDecoder.Eachpartofthe EncoderiscomposedofmultiplelayersofTransformer’sMulti-HeadSelf-Attentionlayers,whichencodetheinput sequence.TheDecoder,ontheotherhand,utilizescross-attentionovertheoutputrepresentationoftheEncoderand generatesthetargetsequenceinanautoregressivemanner.Theencoder-decoderarchitectureservesasthefoundation forprominentLLMssuchasT5[68],flan-T5[91],andBART[92]. 3.2.2. Decoder-onlyArchitecture LLMswithaDecoder-onlyarchitectureutilizethedecodercomponentofthetraditionalTransformerarchitecture. UnliketheEncoder-Decoderarchitecture,whichincorporatesbothanencoderandadecoder,theDecoder-only architectureissolelyfocusedonthedecodingprocess.Inthisconfiguration,themodelsequentiallygeneratestokens, attendingtoprecedingtokensinthesequence.Thisarchitecturehasbeenappliedtovariouslanguagegenerationtasks, showcasingitseffectivenessinvarioustaskssuchastextgenerationwithouttheneedforanexplicitencodingphase. TheDecoder-onlyarchitecturecanbefurtherclassifiedintotwocategories:theCausalDecoderarchitectureandthe PrefixDecoderarchitecture. TheCausalDecoderArchitecture:IntheCausalDecoderarchitecture,eachtokeninthemodelinputsequence canonlyattendtopastinputtokensanditselfduringthedecodingprocess.Itachievesunidirectionalattentiontothe inputsequencebyusingaspecificmaskasshowninFigure1.Infact,differentarchitecturesaremainlyimplementedby configuringdifferentmaskmatrices.ThefigureillustratesacomparisonofmaskconfigurationsbetweentheEncoder- decoderandDecoder-onlyarchitectures(includingCasualDecoderandPrefixDecoder).TherepresentativeLLMsfor YihengLiuetal.:PreprintsubmittedtoElsevier Page9of30 AComprehensiveOverviewfromTrainingtoInference theCausalDecoderarchitecturearetheGPTseries[18;7;8;93;19].TheGPTseriesofLLMsarecurrentlyknownfor theirsuperiorperformance,withtheirfoundationalCausalDecoderarchitecturewidelyappliedinotherLLMssuchas BLOOM[38],OPT[83],Gopher[84],andLLaMA[9]. ThePrefixDecoderArchitecture:ThePrefixDecoderarchitecturecombinestheadvantagesofboththeEncoder- decoderandCausalDecoderarchitectures.Itleveragesitsuniquemaskconfigurations,asillustratedinFigure1, enablingbidirectionalattentionfortokensintheprefixwhilemaintainingunidirectionalattentionforgenerating subsequenttokens[54].Thisdesignallowsfortheautoregressivegenerationoftheoutputsequencewiththeflexibility toattendbi-directionallytotheprefixtokens.RepresentativeLLMsutilizingthePrefixDecoderarchitectureinclude PaLM[36]andGLM[37]. 3.3. Pre-trainingTasks LargeLanguageModels(LLMs)typicallylearnrichlanguagerepresentationsthroughapre-trainingprocess. Duringpre-training,thesemodelsleverageextensivecorpora,suchastextdatafromtheinternet,andundergotraining throughself-supervisedlearningmethods.Languagemodelingisonecommonformofself-supervisedlearningtask inwhichthemodelistaskedwithpredictingthenextwordinagivencontext.Throughthistask,themodelacquires theabilitytoca

ntativeLLMsutilizingthePrefixDecoderarchitectureinclude PaLM[36]andGLM[37]. 3.3. Pre-trainingTasks LargeLanguageModels(LLMs)typicallylearnrichlanguagerepresentationsthroughapre-trainingprocess. Duringpre-training,thesemodelsleverageextensivecorpora,suchastextdatafromtheinternet,andundergotraining throughself-supervisedlearningmethods.Languagemodelingisonecommonformofself-supervisedlearningtask inwhichthemodelistaskedwithpredictingthenextwordinagivencontext.Throughthistask,themodelacquires theabilitytocaptureinformationrelatedtovocabulary,grammar,semantics,andtextstructure. Inlanguagemodeling[18;7;8;36],themodelisrequiredtopredictthenextwordinagivencontext.Thistask enablesthemodeltodevelopanuancedunderstandingoflanguage.Specifically,themodelobserveslargeamounts oftextualdataandattemptstopredictthenextwordateachpositioninthetext.Thisgraduallearningprocess allowsthemodeltocapturethepatternsandinformationinherentinlanguage,encodingavastamountoflinguistic knowledgeintoitsparameters.Oncepre-trainingiscomplete,thesemodelparameterscanbefine-tunedforvarious naturallanguageprocessingtaskstoadapttospecifictaskrequirements.Theobjectiveoflanguagemodelingistotrain amodeltomaximizethelikelihoodoftextualdata.Foragiventextsequence,denotedas 𝑤 1,𝑤 2,...,𝑤 𝑇,where 𝑤 𝑡 representsthetokenatposition𝑡,𝑃 (𝑤 𝑡|𝑤 1,𝑤 2,...,𝑤 𝑡−1 )istheprobabilityofpredicting𝑤 𝑡giventheprecedingcontext 𝑤 1,𝑤 2,...,𝑤 𝑡−1 ,theobjectivefunctionforlanguagemodelingcanbeexpressedusingcross-entropyloss.Here,we definetheobjectiveasmaximizingtheconditionalprobabilityofthegiventextsequence: 𝑇∑ 𝐿 𝐿𝑀 = 1𝑇 − 𝑙𝑜𝑔𝑃 (𝑤 𝑡|𝑤 1,𝑤 2,...,𝑤 𝑡−1 ) (5) 𝑡=1 LanguagemodelingservesasaprevalentpretrainingobjectiveformostLLMs.Inadditiontolanguagemodeling, thereareotherpretrainingtaskswithintherealmoflanguagemodeling.Forinstance,somemodels[68;37]usetext withcertainportionsrandomlyreplaced,andthenemployautoregressivemethodstorecoverthereplacedtokens.The primarytrainingapproachinvolvestheautoregressiverecoveryofthereplacedintervals. 3.4. ModelTraining 3.4.1. ParallelTraining Intheparalleltrainingmentionedbelow,therewillbediscussionsaboutcollectivecommunicationwhichhelpsus betterunderstandtheprinciplesofparalleltraining.Figure2haslistedfivereductionrelationships.1)Broadcast:Send datafromoneGPUtootherGPUs.2)Reduce:Reduce(sum/average)dataofallGPUs,sendtooneGPU.3)AllReduce: ReducealldataofGPUs,sendtoallGPUs.4)ReduceScatter:ReducealldataofGPUs,sendportionstoallGPUs.5)All Gather:GatherdataofallGPUs,sendallGPUs. DataParallel:Theprocessofdataparallelism[94?]isshowninFigure3,thereisaparameterserverthatstores themodel’sparametersandtheentirebatchofdata.EachGPUusesbroadcasttosynchronizethemodelparameters anddividesthedataintooneportionperGPU,witheachGPUreceivingaportionofthedata.EachGPUusesthe completemodelparametersandaportionofthedatatoperformforwardandbackwardpropagation.Thisway,the gradientsareobtainedforeachGPU.Finally,weaggregatethegradientsandsendtheaggregatedgradientsbacktothe parameterserver,wheretheoriginalmodelparametersandtheaggregatedcompletegradientsareavailable.Withthis information,wecanuseanoptimizertoupdatethemodelparameters.Theupdatedparameterswillthenenterthenext roundofmodeltrainingiterations.Distributeddataparallelism[95]abandonstheuseofaparameterserverandinstead employsall-reduceongradientinformation,ensuringthateachGPUreceivesthesamegradientinformation.Theresult ofall-reduceiscommunicatedtoallGPUs,allowingthemtoindependentlyupdatetheirrespectivemodeloptimizers. Aftereachroundofupdates,themodel’sparameters,gradients,andthehistoricalinformationoftheoptimizerare consistentacrossallGPUs. YihengLiuetal.:PreprintsubmittedtoElsevier Page10of30 AComprehensiveOverviewfromTrainingtoInference Figure 2:Fivecollectivecommunicationsthatareusedbyparalleltrainingmethods. TheoccupationofGPUmemoryofintermediateresultsisrelatedtothebatchsize,sentencelength,andmodel d

deloptimizers. Aftereachroundofupdates,themodel’sparameters,gradients,andthehistoricalinformationoftheoptimizerare consistentacrossallGPUs. YihengLiuetal.:PreprintsubmittedtoElsevier Page10of30 AComprehensiveOverviewfromTrainingtoInference Figure 2:Fivecollectivecommunicationsthatareusedbyparalleltrainingmethods. TheoccupationofGPUmemoryofintermediateresultsisrelatedtothebatchsize,sentencelength,andmodel dimensions.Whenusingdataparallelism,abatchofdataisdividedintomanyparts,allowingeachGPUtoprocessa portionofthedata.Inequivalentterms,thebatchsizeprocessedoneachGPUisreducedtooneovertheoriginalnumber ofGPUs.Dataparallelismhasreducedtheinputdimensions,resultinginanoverallreductionintheintermediateresults ofthemodel.Adrawbackisthattosupportmodeltraining,eachGPUneedstoreceiveatleastonepieceofdata.In themostextremecase,wheneachGPUreceivesonlyonepieceofdata,ourparameters,gradients,andoptimizerstill needtobefullystoredontheGPU.Evenifwedon’tstoreanyintermediateresultsontheGPU,ourmodelmaystillbe unabletoperformcomputationsonasingleGPU. ModelParallel:Modelparallelism[96]wasfirstintroducedbyMegatron-LMtoalleviatememorypressure.From figure4,wecanclearlyunderstandtheoverallarchitectureofmodelparallelism.Takingadvantageofthemostcommon linearlayerintheTransformerasanexample,theparametersofthelinearlayerformamatrixofsizeA*B,andthe inputtothelinearlayerisavectorofsizeB*1.Representingthisas𝑦𝐴 ∗𝐵 =𝑊 𝐴 ∗𝐵 𝑥 𝐵,wecanhorizontallypartition themodel’sparametersintomanysegmentsusingthepropertyofmatrixmultiplication.Eachsegmentisofsizea dividedbynmultipliedbyB.Utilizingthepropertiesofmatrixmultiplication,wecanmove𝑥 𝐵 intoparentheses,and finally,theresultofthelinearlayerisobtainedbymultiplyingmanysmallmatriceswiththeparametersofthelinear layer.Throughthisapproach,theparametersofthelinearlayercanbedistributedacrossmultipleGPUs.However,itis YihengLiuetal.:PreprintsubmittedtoElsevier Page11of30 AComprehensiveOverviewfromTrainingtoInference Figure 3:Thearchitectureofdataparallelismanddistributeddataparallelism.Thediagramillustratesthedifference betweendataparallelismanddistributeddataparallelismandtheadvantagesofdistributeddataparallelism. crucialtoensurethattheinputstothemodelonmultipleGPUsareidentical.Insteadofusingadataparallelapproach topartitionthedata,weneedtoensurethattheinputsobtainedoneachGPUarethesame,meaningtheybelongto thesamebatchofdata.WecanthenpartitionaparameterlikethelinearlayeracrossGPUs,witheachGPUreceiving asmallportionofthematrix.Byperformingmodelcalculationswiththissmallportionandthedata,weobtaina sub-result,asshowninFormula5.Theresultsofthesecomputationsneedtobeconcatenatedusingtheall-gather operatorandcommunicatedtoallGPUs. 𝑦𝐴 ∗𝐵 = 𝑊 𝐴 ∗𝐵 𝑥 𝐵 = [ 𝑊 (1)𝐴 𝐴 𝐴 (6) 𝑛 ∗𝑏;𝑊 (2)𝑛 ∗𝑏;...;𝑊 (𝑛)𝑛 ∗𝑏]𝑥 𝐵 = [ 𝑊 (1)𝐴 𝐴 𝐴 𝑛 ∗𝑏𝑥 𝐵 ;𝑊 (2)𝑛 ∗𝑏𝑥 𝐵 ;...;𝑊 (𝑛)𝑛 ∗𝑏𝑥 𝐵 ] ZeRO:ZeRO[97]isaframeworkbuiltondataparallelism.DuringtheparameterupdatingprocessoneachGPU, thesamesetofparametersisused,leadingtocomputationalredundancy.EachGPUusesreducedscattertoeliminate thisredundancytoobtainaportionofthegradientresults.Afterupdatingaportionofthemodelparametersoneach GPU,anall-gatheroperationisperformedtosynchronizetheparametersacrossallGPUs.Aftertheall-gatheroperation, theoriginalgradientnolongerneedstobesavedonthegraphicscardandcanberemoved.Figure5showstheupdate ofZeRO.InZeRO1,theoriginalgradientisremovedafterbackwardpropagation,whileinZeRO2,theproductof thegradient*iscalculatedinadvanceduringbackwardpropagation,andonlythegradient*issavedonthegraphics card,removingthegradient.Thisway,thedeletionofthegradientisadvanced,leadingtofurthersavingsinGPU YihengLiuetal.:PreprintsubmittedtoElsevier Page12of30

ginalgradientnolongerneedstobesavedonthegraphicscardandcanberemoved.Figure5showstheupdate ofZeRO.InZeRO1,theoriginalgradientisremovedafterbackwardpropagation,whileinZeRO2,theproductof thegradient*iscalculatedinadvanceduringbackwardpropagation,andonlythegradient*issavedonthegraphics card,removingthegradient.Thisway,thedeletionofthegradientisadvanced,leadingtofurthersavingsinGPU YihengLiuetal.:PreprintsubmittedtoElsevier Page12of30 AComprehensiveOverviewfromTrainingtoInference Figure4:Theoverallarchitectureofmodelparallelism.Theleftsideofthediagramshowstheprocessofmodelparallelism, andtherightsideshowsthememoryusageofparameters,gradients,andoptimizersinthegraphicscardofthemodel parallelismmethod. memoryspace.ZeRO3conductsadetaileddivisionofthemodelparameters.Eachgraphicscardretainsonlyaportion ofthegradientsforupdating,andparameterupdatesalsoonlyaffectaportionofthemodelparameters.Therefore, eachgraphicscardonlyneedstostoretheparameters,gradients,andoptimizerrelatedtothepartoftheparameters itisresponsiblefor.Duringforwardandbackwardpropagation,anall-gatheroperationisrequiredonce,andafterthe operationiscomplete,themodelparametersarereleasedfromthegraphicscard.Zero3doesnotuseallgatherduring parameterupdates,butitrequiresanall-gatheroperationduringbothforwardandbackwardpropagation,addingone communicationstep.ComparedtoZeRO2,ZeRO3isanalgorithmthattradestimeforspace. PipelineParallel:Pipelineparallelism[98]andmodelparallelismsharesimilarities.Inmodelparallelism,linear layersaredividedintomanysmallmatrices,whicharethendistributedtodifferentGPUs.Forpipelineparallelism, differentlayersofthemodelareassignedtodifferentGPUs.Specifically,ifwehaveann-layertransformer,wecan assignthe𝑙𝑎𝑦𝑒𝑟 𝑖ofthetransformertothe𝐺𝑃𝑈 𝑖,andsoon.Duringtheforwardpropagationofthemodel,weneed toperformthecomputationofthe𝑙𝑎𝑦𝑒𝑟 𝑖onthe𝐺𝑃𝑈 𝑖,thenpasstheresulttothe𝐺𝑃𝑈 𝑖+1 .The𝐺𝑃𝑈 𝑖+1 receivesthe outputfromthe 𝐺𝑃𝑈 𝑖,performsthecomputationforthatlayerandpassestheresulttothenextGPU.Thismethod partitionstheparameters,gradients,optimizer,andintermediateresultsforeachlayer. 3.4.2. MixedPrecisionTraining Inrecentyears,topre-trainextremelylargelanguagemodels,someresearch[99]hasbeguntoutilize16-bitfloating- pointnumbers(FP16)toreducememoryusageandcommunicationoverhead.FP16hasasmallernumericalrangeand lowerprecisionineffectivedigits[100;38],butcomputationstendtobefasterthanFP32.Ingeneralmodeltraining, FP32isoftenusedasthedefaultrepresentationfortrainingparameters.However,inactualmodeltraining,thenumber ofparametersinamodeltypicallydoesnotexceedtheorderofthousands,wellwithinthenumericalrangeofFP16. Toimprovecomputationalspeed,wecanconvertfromFP32toFP16.Duringparameterupdates,theamountofthe parameterisroughlyequaltothegradientmultipliedbythelearningrate.TheminimumvalueofFP16isontheorder of1e-5.AstheproductofthegradientandlearningrateisalreadywellbelowtherepresentationrangeofFP16,the parameterupdatewouldresultinloss,knownasunderflow.Therefore,werepresenttheparameterupdateobtainedby YihengLiuetal.:PreprintsubmittedtoElsevier Page13of30 AComprehensiveOverviewfromTrainingtoInference Figure 5:TheoverallarchitectureofZeRO.TheupperdemonstratesZeROstage1andZeROstage2.Thelowerdisplays ZeROstage3.ThegraphillustratestheoptimizationofmemoryusageofgraphicscardparametersinrelationtoZeRO3 versusZeRO1andZeRO2 multiplyingthegradientbythelearningrateasFP32.Wecannotdirectlyaddthishigh-precisionparameterupdate toalower-precisionmodel,asthiswouldstillresultinfloating-pointunderflow.Consequently,weneedtosavean additionalsingle-precisionparameterontheoptimizer.Toacceleratebothforwardandbackwardpassesinthemodel, half-precisionparametersandgradientsareusedandpassedtotheoptimizerforupdating.Theoptimizer’supdate quantityissavedasFP32,andweaccumulateiteffectivelythroughatemporarilycreatedFP32parameterinthe optimizer.Aftereffectiveaccumulation,itisthenconvertedbacktoFP16parameters. 3.4.3. Offloading Theparametersi

date toalower-precisionmodel,asthiswouldstillresultinfloating-pointunderflow.Consequently,weneedtosavean additionalsingle-precisionparameterontheoptimizer.Toacceleratebothforwardandbackwardpassesinthemodel, half-precisionparametersandgradientsareusedandpassedtotheoptimizerforupdating.Theoptimizer’supdate quantityissavedasFP32,andweaccumulateiteffectivelythroughatemporarilycreatedFP32parameterinthe optimizer.Aftereffectiveaccumulation,itisthenconvertedbacktoFP16parameters. 3.4.3. Offloading Theparametersintheoptimizerareatleasttwiceasmanyasthemodelparameters,andastudy[101]proposesthe ideaofmovingtheoptimizer’sparametersfromtheGPUtotheCPU.AlthoughGPUcomputationismuchfasterthan CPU,thequestionariseswhetheroffloadingthisoperationcouldbecomeabottleneckfortheoveralltrainingspeed ofthemodeloptimizer.Inreality,weutilizeZeRO3.AftertheoptimizationwithZeRO3,thesizeoftheparameters, gradients,andoptimizerisreducedto1/nofthenumberofGPUs.BybindingoneGPUtomultipleCPUs,weeffectively lowerthecomputationalloadoneachCPU. 3.4.4. Overlapping Memoryoperationsaretypicallyasynchronous.Thus,Wecansendarequesttothememoryinadvanceandthen proceedwithothercomputations.Aftercompletingothercomputations,wecomebacktohandlethememoryrequest. Thistwo-stepoperationisusedintheforwardpropagationprocessofmodeltraining.Weneedtoobtaintheparameters of𝑙𝑎𝑦𝑒𝑟 𝑖throughagatheroperation.Afterobtainingtheparametersof𝑙𝑎𝑦𝑒𝑟 𝑖,intheforwardpropagationprocessof 𝑙𝑎𝑦𝑒𝑟 𝑖,weproactivelyretrievetheparametersof𝑙𝑎𝑦𝑒𝑟 𝑖+1 throughanasynchronousfetch.Oncetheforwardpropagation YihengLiuetal.:PreprintsubmittedtoElsevier Page14of30 AComprehensiveOverviewfromTrainingtoInference Table 3 Commonlyusedinstructiontuningdatasets. Datasets Links static-hh https://huggingface.co/datasets/Dahoas/static-hh OIG https://huggingface.co/datasets/laion/OIG Self-Instruct[102] https://github.com/yizhongw/self-instruct Naturalinstructions[103] https://github.com/allenai/natural-instructions P3[104] https://huggingface.co/datasets/bigscience/P3 Promptsource[105] https://github.com/bigscience-workshop/promptsource WebGPT[106] https://huggingface.co/datasets/openai/webgpt_comparisons Flan[107] https://github.com/google-research/flan MVPCorpus[108] https://github.com/RUCAIBox/MVP calculationfor𝑙𝑎𝑦𝑒𝑟 𝑖iscompleted,theparametersfor𝑙𝑎𝑦𝑒𝑟 𝑖+1 havebeenobtainedandarestoredintheGPU.Wecan thenimmediatelyproceedwiththeforwardpropagationcalculation,andsoon. 3.4.5. Checkpoint Inordertosupportthebackwardpropagationofthemodel,AllintermediateresultsintheGPUmemoryneedto besavedduringtheforwardpropagationofthemodel.Tooptimizethisprocess,acheckpointmechanism,whichdoes notsaveallintermediateresultsintheGPUmemorybutonlyretainscertaincheckpointpointsisutilized. Thediagrambelowillustratesasimplifiedstructureofatransformer.Eachtransformerblocktakesamodelinput, undergoescomplexcomputationsthroughattentionandfeed-forwardprocesses,andproducestheoveralloutputof thatlayer.Wekeeponlytheinputofeachmajorlayerinthetransformerasourcheckpoint. Duringthebackwardpropagationprocess,howdowecomputethegradientsofthelinearlayerswithineachmajor layer?Wecanperformatechniquecalledrecomputation,whichinvolvesre-executingtheforwardpassofeachmajor layerduringthebackwardpropagationprocess.Wetemporarilyobtaintheinputsofthelinearlayerswithineachmajor layer,andtheintermediateresultsobtainedcanbeusedforbackwardpropagation.Oncethebackwardpropagationfor thatlayeriscomplete,wecandiscardthecheckpointandthetemporarilyrecomputedintermediateresultsofthelinear layerswithinthemodelfrom

kwardpropagationprocess,howdowecomputethegradientsofthelinearlayerswithineachmajor layer?Wecanperformatechniquecalledrecomputation,whichinvolvesre-executingtheforwardpassofeachmajor layerduringthebackwardpropagationprocess.Wetemporarilyobtaintheinputsofthelinearlayerswithineachmajor layer,andtheintermediateresultsobtainedcanbeusedforbackwardpropagation.Oncethebackwardpropagationfor thatlayeriscomplete,wecandiscardthecheckpointandthetemporarilyrecomputedintermediateresultsofthelinear layerswithinthemodelfromtheGPUmemory. Assumingwehaveatransformerwith24layers,eachlayercontainingfourtofivelinearlayers,usingthecheckpoint mechanismreducestheoriginallyrequiredstorageof120intermediateresultstoonly24intermediateresults. 3.5. Fine-Tuning ThetrainingofLLMsinthispaperisdividedintothreestages:datacollectionandprocessing,pre-training,and fine-tuning.Thissectionwillprovideareviewofthefine-tuningmethodsforLLMs.Specifically,wecategorizefine- tuningtechniquesintothreetypes:supervisedfine-tuning(SFT)[93],alignmenttuning,andparameter-efficienttuning. 3.5.1. SupervisedFine-Tuning Thecoreconceptofsupervisedfine-tuninginvolvesadjustingthemodelinasupervisedmanneronthebasisof large-scalepre-training,enhancingitscapabilitytobetteradapttothespecificrequirementsofthetargettask.Inthe processofSFT,itisnecessarytopreparealabeleddatasetforthetargettask,whichincludesinputtextalongwith correspondinglabels.Instructiontuningisacommonlyusedtechniqueinthefine-tuningprocessofLLMsandcan beconsideredasaspecificformofSFT.ItinvolvesfurthertrainingLLMsonadatasetcomposedof(instruction, output)pairs,focusingonenhancingthecapabilitiesandcontrollabilityoflargelanguagemodelsbyunderstanding andfollowinghumaninstructions.Wecompiledcommonlyusedinstructiontuningdatasets,asillustratedinTable3. 3.5.2. AlignmentTuning DuetoLLMsbeingpre-trainedonmassiveanddiverseinternetdata,eventhoughthetrainingdataundergoes somepreprocessing,itisstillchallengingtoguaranteetheabsenceofbiasedorharmfulcontentinterabyte-scale trainingdatasets.DespiteLLMsdemonstratingimpressiveperformanceacrossvariousnaturallanguageprocessing tasks,theyfrequentlyexhibitbehaviorsdivergingfromhumanintent.Thisincludesgeneratingfalseinformation, YihengLiuetal.:PreprintsubmittedtoElsevier Page15of30 AComprehensiveOverviewfromTrainingtoInference producingexpressionswithbiasormisleadingcontent,andsoon[93;109].ToaddresstheseissuesofLLMsdisplaying behaviorsbeyondhumanintent,alignmenttuningbecomescrucial[93;110]. Ingeneral,alignmenttuningaimstomeetthefollowingthreecriteria:beinghelpful,honest,andharmless. Helpful:Theconceptofhelpfulnessrevolvesaroundwhetherthemodel-generatedoutputprovesgenuinely beneficialforaspecifictaskorinquiry.Intherealmofnaturallanguageprocessing,themodel’sgeneratedtextor responsesshouldfurnishvaluableinformation,positivelyimpactingtheuser’srequirementsortaskobjectives. Honest:Honestyentailswhetherthemodel-generatedoutputisauthenticandreliable.Themodelshouldproduce informationconsistentwithfacts,steeringclearoffabricationordistortion.Thiscontributestomaintainingusertrust intheauthenticityofthemodel’soutputs. Harmless:Harmlessnessisconcernedwithwhetherthemodel-generatedoutputposesnoharmtousersorsociety. Themodelshouldrefrainfromgeneratingcontentthatisharmful,offensive,orperilous,ensuringitsutilizationremains safeforallrelevantstakeholders. IntrainingLLMs,anoteworthyapproachtoalignmenttuningisbasedonReinforcementLearningwithHuman Feedback(RLHF)[93].Thismethodinvolvescollectinghumanfeedbackdatatotrainarewardmodel(RM)for reinforcementlearning.TheRMservesastherewardfunctionduringreinforcementlearningtraining,andalgorithms suchasProximalPolicyOptimization(PPO)[111]areemployedtofine-tunetheLLM.Inthiscontext,LLMis consideredasthepolicy,andtheactionspaceisconsideredasthevocabularyoftheLLM. 3.5.3. Parameter-efficientTuning Currently,large-scalePLMssuchasChatGPT[93;19]continuetogrowinscale.However,forthemajorityof researchers,conductingfullfine-tuningonconsumer-gradehardwarehasbecomecost-pr

edbackdatatotrainarewardmodel(RM)for reinforcementlearning.TheRMservesastherewardfunctionduringreinforcementlearningtraining,andalgorithms suchasProximalPolicyOptimization(PPO)[111]areemployedtofine-tunetheLLM.Inthiscontext,LLMis consideredasthepolicy,andtheactionspaceisconsideredasthevocabularyoftheLLM. 3.5.3. Parameter-efficientTuning Currently,large-scalePLMssuchasChatGPT[93;19]continuetogrowinscale.However,forthemajorityof researchers,conductingfullfine-tuningonconsumer-gradehardwarehasbecomecost-prohibitiveandimpractical. UnlikeSFTandalignmenttuning,theobjectiveofparameter-efficienttuningistoreducecomputationalandmemory overhead.Thismethodinvolvesfine-tuningonlyasmalloradditionalsubsetofmodelparameterswhilekeeping themajorityofpre-trainedparametersfixed,therebysignificantlyloweringcomputationalandstoragecosts.Itis noteworthythatstate-of-the-artparameter-efficienttuningtechniqueshaveachievedperformancelevelscomparable tofullfine-tuning.Somecommonparameter-efficienttuningmethodsincludeLow-RankAdaptation(LoRA)[112], PrefixTuning[113]andP-Tuning[114;115].Theadoptionofthesemethodsenablesefficientmodeltuningevenin resource-constrainedenvironments,offeringfeasibilityandefficiencyforpracticalapplications. WiththeriseofLLMs,parameter-efficienttuninghasgarneredincreasingattention,withLoRAbeingwidely employedinthelatestreleasesofLLMs.LoRA[112]anditsrelatedadvancements[116;117]arenoteworthyand deserveattention. 3.5.4. SafetyFine-Tuning ToenhancethesafetyandresponsibilityofLLMs,theintegrationofadditionalsafetytechniquesduringfine-tuning isessential.Thisencompassesthreeprimarytechniques,applicabletobothSFTandRLHFphases. SupervisedSafetyFine-Tuning:Inthistechnique,labelersaretaskedwithgeneratingdemonstrationdatathat incorporateshighsafetyriskadversarialprompts.Thishandcraftsafetydemonstrationdataisthenincorporatedinto theSFTphase,therebyaugmentingthemodel’scapacitytomanagesafetyrisks. SafetyRLHF:Thistechniqueemploysthesameorevenmoreaggressiveadversarialpromptstoquerythemodels. ThesafestresponseexhibitingrefusalbehavioristhenusedtotrainasafetyrewardmodelwithintheRLHFframework. SafetyContextDistillation:Thistechniqueemployscontextdistillation[118]byinitiallyprefixingsafety preprompts,like“Youareasafeandresponsibleassistant,”toadversarialprompts.Thisprocessyieldssafergenerated responses.Themodelisthenfine-tunedonthesesaferdemonstrationdatabutwithouttheinclusionofthesafety pre-prompts.Thissafetydistillationfurtherenhancesthemodel’ssafetycapabilities. 3.6. Evaluation Unlikeinthepast,large-scaledeeplearningmodelshaveawiderrangeofapplicationsandstrongerperformance comparedtoordinarymodels.However,withgreatpowercomesgreatresponsibility,andevaluatingthesemodelshas becomemorecomplex,requiringconsiderationofpotentialproblemsandrisksfromallaspects.Sincethepopularity ofChatGPT,manyrelatedstudieshavebeenpublished,includingthesurveyandsummaryofLLMsevaluationin reference[119;120],whichishelpfulfordevelopinglarge-scaledeeplearningmodels.Thissectionwillintroduce sometestingdatasets,evaluationdirectionsandmethods,andpotentialthreatsthatneedtobeconsideredbasedon previousevaluationworkonlargemodels. YihengLiuetal.:PreprintsubmittedtoElsevier Page16of30 AComprehensiveOverviewfromTrainingtoInference 3.6.1. Statictestingdataset Theevaluationoflargemodels’capabilitiesrequiresappropriatedatasetsforvalidation.Here,weintroduceseveral commonlyuseddatasetsfortestingpurposes.Consideringmultimodallargemodels,typicaldatasetsforcomputer visionincludeImageNet[121]andOpenImages[122].InadditiontothecommonlyusedGLUE[123]andSuperGLUE [124]forLLMs,MMLU[125]ishighlycompetitiveintestingcomprehensivecapability.Ifyourmodelprimarilyuses Chineselanguage,thenCMMLU[126],asabenchmarkforChineselargemodels,shouldalsobeconsidered,and XTREME[127]andXTREME-R[128]aresuitablechoicesformultilinguallargemodels.Forassessingmathematical knowledgecapabilities,therearedatasetssuchasMATH[129]andGSM8K[130],whileHumanEval[131]andMBPP [132]canserveasbenchmarksforcodegeneration.Forcom

ImageNet[121]andOpenImages[122].InadditiontothecommonlyusedGLUE[123]andSuperGLUE [124]forLLMs,MMLU[125]ishighlycompetitiveintestingcomprehensivecapability.Ifyourmodelprimarilyuses Chineselanguage,thenCMMLU[126],asabenchmarkforChineselargemodels,shouldalsobeconsidered,and XTREME[127]andXTREME-R[128]aresuitablechoicesformultilinguallargemodels.Forassessingmathematical knowledgecapabilities,therearedatasetssuchasMATH[129]andGSM8K[130],whileHumanEval[131]andMBPP [132]canserveasbenchmarksforcodegeneration.Forcommonsensereasoningtestsindailyhumanlifeandwork, thefollowingdatasetsareavailable:HelloSwag[133],PIQA[134],BoolQ[135],SIQA[136],WinoGrande[137], ARC[138],andOpenBookQA[139].Formedicalknowledge,therearedatasetssuchasMedQA-USMLE[140]and MedMCQA[141]. 3.6.2. OpendomainQ&Aevaluation Currently,LLMsinteractwithhumansintheformofquestionsandanswers.Comparedtothefragmentedand ambiguousinformationreturnedbytraditionalsearches,LLMsprovidemorerealisticandefficientquestion-and- answerresultsthatalignwithhumanhabits.Therefore,theevaluationofODQA(OpenDomainQuestionAnswering) [142]capabilityisessential.Theperformanceofopen-domainquestionansweringgreatlyaffectsuserexperience. CommonlyuseddatasetsfortestingincludeSquAD[143]andNaturalQuestions[144],withF1scoreandExact-Match accuracy(EM)asevaluationmetrics.However,notethatthemethodofwordmatchingmayhavecertainissues,such aswhenafactuallycorrectanswerisnotinthegoldenanswerlist.Therefore,humanevaluationseemstobenecessary, andliterature[145]hasconducteddetailedresearchonthismatter. 3.6.3. Securityevaluation Asanemergingandhotresearchfield,LLMsmustpayattentiontotheirpotentialsecuritythreats,preventmalicious useorvulnerabilitiestomaliciousattacks,andaddressanypotentiallong-termissuesthatmayposeathreatto humandevelopment.Additionally,redteaminginvariousdomainsisnecessarytocriticallyassessandtestthemodel, identifyingvulnerabilities,biases,inaccuracies,andareasforsafetyimprovement. Potentialbias:ThetrainingdataforLLMsmaycontainpotentialbiases,suchasgenderorrace.Security assessmentsneedtoaddresswhetherthemodelgeneratesoramplifiesthesebiasesandhowtoreduceorcorrectthem. Reference[146]discussesindetailthecausesofbiasinLLMsandtheseriousconsequencesthatmayarise.Reference [147]extensivelystudieshowpre-trainedlanguagemodelsgenerateharmfulcontenttowhatextent,andhowtouse controlledtextgenerationalgorithmstopreventthegenerationoftoxiccontent.CHBias[148]isaChinesedatasetthat canbeusedtoevaluateandmitigatethebiasproblemofLLMs. Privacyprotection:LLMsmaycomeintocontactwithalargeamountofuserdata,suchastextandimages, duringthetrainingprocess.Securityassessmentsneedtoensuretheeffectiveprotectionofuserdataprivacytoprevent leaksandmisuse.Reference[149]conductedresearchonmodelslikeChatGPTandfoundthatitispossibletoextract trainingdataeffectivelyfromthesemodels.Reference[150]providesasolutionbyproposingaframeworkcalled DEPN(DetectandEditingPrivacyNeurons)todetectandeditprivacyneuronsinpre-trainedlanguagemodels.It alsointroducesaprivacyneuronaggregatortoeliminateprivacyinformationinabatch-processingmanner,effectively reducingtheleakageofprivacydatawhilemaintainingmodelperformance. Adversarialattacks:LLMsmaybesusceptibletoadversarialattacks,suchasinputtampering,intentional misinformation,orgeneratingfalseinformation.Securityassessmentsneedtoconsidertherobustnessofthemodel, i.e.,itsabilitytowithstandsuchattacks.Asmentionedinreference[151],LLMsstillhave"jailbreak"risks,whereusers canmanipulatethemodeltogeneratetoxiccontentusingspecificinputmethodslikerole-playingoraddingspecial suffixesasstudiedinthereferencedpaper.Especiallywhenusingopen-sourcepre-trainedmodels,anyvulnerabilities inthepre-trainingmodelsregardingadversarialattacksareinheritedaswell.Reference[152]providesasolutionto mitigatetheharmcausedbythesevulnerabilities. 3.6.4. Evaluationmethod AutomatedevaluationandmanualevaluationplaycrucialrolesinLanguageModel(LLM)research.Automated evaluationtypicallyinvolvesusingvariousmetricsandindicatorstoquantifytheperformanceofmodels,suchas BIEU[153],ROUGE[154],andBERTSScore[155],whichcanmeasuretheaccurac

rencedpaper.Especiallywhenusingopen-sourcepre-trainedmodels,anyvulnerabilities inthepre-trainingmodelsregardingadversarialattacksareinheritedaswell.Reference[152]providesasolutionto mitigatetheharmcausedbythesevulnerabilities. 3.6.4. Evaluationmethod AutomatedevaluationandmanualevaluationplaycrucialrolesinLanguageModel(LLM)research.Automated evaluationtypicallyinvolvesusingvariousmetricsandindicatorstoquantifytheperformanceofmodels,suchas BIEU[153],ROUGE[154],andBERTSScore[155],whichcanmeasuretheaccuracyofLLM-generatedcontent. YihengLiuetal.:PreprintsubmittedtoElsevier Page17of30 AComprehensiveOverviewfromTrainingtoInference Thesemetricscanhelpresearchersquicklyassessmodelperformanceonlarge-scaledataandcomparedifferent models.However,automatedevaluationalsohaslimitationsasitcannotfullycapturethecomplexityoflanguage understandingandgeneration.Researchinreference[156]hasshownthatmanualevaluationismorereliablefor someopen-endedgenerationtasks.Manualevaluationtypicallyinvolveshumanannotatorssubjectivelyjudgingand assessingthequalityofmodel-generatedoutputs.Thisevaluationmethodcanhelprevealhowmodelsperformin specifictasksorscenariosandidentifysubtleissuesanderrorsthatautomatedevaluationmayoverlook.However, manualevaluationalsofaceschallengessuchashightimecostsandsubjectivity.Therefore,itisoftennecessaryto combinethestrengthsofautomatedandmanualevaluationtocomprehensivelyassesstheperformanceoflanguage models. 3.7. LLMFramework Largedeeplearningmodelsoffersignificantaccuracygains,buttrainingbillionstotrillionsofparametersis challenging.Existingsolutionssuchasdistributedtraininghavesolvedfundamentallimitationstofitthesemodelsinto limiteddevicememorywhileobtainingcomputation,communication,anddevelopmentefficiency.Next,thissection willintroduceseverallargelanguagemodelframeworksthatutilizedistributedtrainingtechnologyleveragingGPU, CPU,andNVMememorytoallowforunprecedentedmodelscaleonlimitedresourceswithoutrequiringmodelcode refactoring. TransformersTransformers[157],anopen-sourcePythonlibrarybyHuggingFace,isdedicatedtobuildingmodels usingtheTransformerarchitecture.Featuringasimpleanduser-friendlyAPI,itfacilitateseasycustomizationofvarious pre-trainedmodels.Witharobustcommunityofusersanddevelopers,transformerscontinuouslyupdateandimprove modelsandalgorithms. DeepSpeed:Deepspeed[158],anopen-sourceoptimizationlibrarycompatiblewithPyTorch,isdevelopedby MicrosoftandutilizedintrainingLLMslikeMTNLG[79]andBLOOM[38].Currently,Itprovidesfullsupport forZeROtechnologywhichincludesOptimizerstatepartitioning,Gradientpartitioningandparameterpartitioning, Custommixedprecisiontraining,ArangeoffastCUDA-extension-basedoptimizers[159]andZeRO-offloadtoCPU andDisk/NVMe.Throughtheabovetechnologies.Additionally,Deepspeedhasachievedexcellentscalabilityand efficiencywithsmallmemoryrequirements. BMTrain:BMTrain[160]isanefficientlargemodeltrainingtoolkitdevelopedbyTsinghuaUniversitythatcanbe usedtotrainlargemodelswithtensofbillionsofparameters.Itcantrainmodelsinadistributedmannerwhilekeeping thecodeassimpleasstand-alonetraining.BMTraindoesnotrequiremodelrefactoringtowork.Infact,PyTorchusers canenableBMTrainwithafewlinesofcodechangetotheirexistingtrainingpipeline.Itprovidesthesupportof variousoptimizationtechniquessuchasZeROoptimizationandcommunicationoptimization. Megatron-LM:Megatron-LM[96;161;162]isadeeplearninglibrarydevelopedbyNVIDIAfortraininglarge- scalelanguagemodels.Megatron-LMpresentstheirtechniquesincludingmodelanddataparallelism,mixed-precision training,andFlashAttentionfortrainingverylargetransformermodels.Specifically,ittakesadvantageofthestructure oftransformernetworkstocreateasimplemodelparallelimplementationbyaddingafewsynchronizationprimitives anditenablestrainingtransformermodelswithbillionsofparametersandtrainsefficientlyinPyTorch.Italsoperforms anin-depthempiricalanalysisoftheirmodelanddataparalleltechniqueanddemonstratesupto76%scalingefficiency using512GPUswhichcanlargelyimprovethetrainingefficiencyandspeed,enablingefficientdistribut

ion training,andFlashAttentionfortrainingverylargetransformermodels.Specifically,ittakesadvantageofthestructure oftransformernetworkstocreateasimplemodelparallelimplementationbyaddingafewsynchronizationprimitives anditenablestrainingtransformermodelswithbillionsofparametersandtrainsefficientlyinPyTorch.Italsoperforms anin-depthempiricalanalysisoftheirmodelanddataparalleltechniqueanddemonstratesupto76%scalingefficiency using512GPUswhichcanlargelyimprovethetrainingefficiencyandspeed,enablingefficientdistributedtraining acrossGPUs. Inadditiontotheaforementionedframeworks,Colossal-AI[163]andFastMoE[164;165]arealsotwopopular frameworksfortrainingLLMs.Inprinciple,anydeeplearningframeworkthatsupportsparallelcomputingcanbeused totrainLLMs.ExamplesincludePyTorch[166],TensorFlow[167;168],PaddlePaddle[169],MXNet[170],OneFlow [171],MindSpore[172]andJAX[173]. 4. InferencewithLargeLanguageModels Thescaleoflargemodelsisgrowingatarateofnearly10timesperyear,whichbringsabouthugecomputational consumptionandcarbonemissions[174].Therefore,reducingthecomputationalburdenoftraininglargemodelswhile retainingtheirreasoningabilityhasbecomeacommonconcernforeveryone.Inthischapter,wemainlyintroducehow toreducecostsfrombothcomputationalandstorageaspects,thatis,howtoefficientlyperformlarge-scalemodel inferencefromfouraspects:modelcompression,memoryscheduling,parallelism,andstructuraloptimization. YihengLiuetal.:PreprintsubmittedtoElsevier Page18of30 AComprehensiveOverviewfromTrainingtoInference 4.1. ModelCompression 4.1.1. KnowledgeDistillation KnowledgeDistillation[175]referstotransferringknowledgefromacumbersome(teacher)modeltoasmaller (student)modelthatismoresuitablefordeployment.Thisisachievedbyfittingthesofttargetsofthetwomodels, assofttargetsprovidemoreinformationthangoldlabels.Initially,thecalculationformodeldistillationinvolvedonly fittingtheoutputsfromthelastlayerofboththeteacherandstudentmodels[176].PKD[177]improvesthisprocessby computingthemean-squarelossbetweennormalizedhiddenstates,allowingthestudentmodeltolearnfrommultiple intermediatelayersoftheteachermodel.Inordertodiscovermoreintermediaterepresentationssuitableforknowledge distillation,Jiaoetal.[178]proposedTinyBERT.Thisenablesthestudentmodeltolearnfromtheembeddinglayer andattentionmatricesoftheteachermodel. 4.1.2. ModelPruning Modelpruninginvolvesremovingredundantportionsfromtheparametermatricesoflargemodels.Itisdivided intounstructuredpruningandstructuredpruning.Unstructuredpruninginvolvesremovingindividualconnections orweightsinaneuralnetworkwithoutadheringtoanyspecificstructuralpattern.Instructuredpruning,specific structuralpatternsorunitswithinaneuralnetworkareprunedorremoved.Gordonetal.[179]comparedtheeffects ofunstructuredandstructuredpruningontheBERTmodel.Theyfoundthattheeffectivenessofunstructuredpruning significantlydecreasesasthepruningratioincreases,whileinstructuredpruning,30-40%oftheweightscanbe discardedwithoutaffectingBERT’suniversality.Differentstructuresinthemodelcanbestructurallypruned.Michel etal.[180]prunedattentionheadsandfoundthatablatingoneheadoftenpositivelyimpactstheperformanceofWMT andBERT.Theyproposedagradient-basedmetricforevaluatingtheimportanceofattentionheadstoenhancepruning effectiveness.Fanetal.[179]performedlayerpruningbyextendingdropoutfromweightstolayers.Duringtraining, theyrandomlydroppedlayersandachievedgoodinferenceresultsbyselectingsub-networkswithanydesireddepth duringtesting. 4.1.3. ModelQuantization Thefundamentalideabehindmodelquantizationistoreducethenumberoffloating-pointbitsusedinnumerical calculationswithinalargemodelnetwork,therebydecreasingstorageandcomputationcosts.Thisinvolvesconverting floating-pointoperationsintofixed-precisionoperations.However,asprecisiondecreases,themodel’slossgradually increases,andwhenprecisiondropsto1bit,themodel’sperformanceexperiencesasuddendecline.Toaddressthe optimizationchallengesintroducedbylow-precisionquantization,Baietal.[181]proposedBinaryBERT.Theyinitially trainedahalf-sizedternarymodelandtheniniti

istoreducethenumberoffloating-pointbitsusedinnumerical calculationswithinalargemodelnetwork,therebydecreasingstorageandcomputationcosts.Thisinvolvesconverting floating-pointoperationsintofixed-precisionoperations.However,asprecisiondecreases,themodel’slossgradually increases,andwhenprecisiondropsto1bit,themodel’sperformanceexperiencesasuddendecline.Toaddressthe optimizationchallengesintroducedbylow-precisionquantization,Baietal.[181]proposedBinaryBERT.Theyinitially trainedahalf-sizedternarymodelandtheninitializedabinarymodelwiththeternarymodelthroughweightsplitting. Finally,theyfine-tunedthebinarymodel.Thisapproachyieldedbetterresultsforthebinarymodelcomparedtotraining abinarymodelfromscratch. 4.1.4. WeightSharing ThebasicideaofweightsharingistousethesamesetofparametersformultiplepartsofaLLM.Insteadoflearning differentparametersforeachinstanceorcomponent,themodelsharesacommonsetofparametersacrossvariousparts. Weightsharinghelpsreducethenumberofparametersthatneedtobelearned,makingthemodelmorecomputationally efficientandreducingtheriskofoverfitting,especiallyinsituationswherethereislimiteddata.ALBERT[182]uses theCross-layerparameter-sharingstrategytoeffectivelyreducethenumberofparametersofthemodel,andcanachieve bettertrainingresultsthanthebaselinewiththesameparameternumber. 4.1.5. Low-rankApproximation Low-rankdecompositionmethodsarecrucialinthefieldofmodelcompression,astheyallowforthecreation ofmorecompactmodelswithfewerparameters.Thisreductioninmodelsizeisparticularlybeneficialfordeploying neuralnetworksonresource-constraineddevices,improvingefficiencyduringinference.Chenetal.[183]performeda low-rankdecompositionontheinputmatrix,enablingmatrixoperationswithinthelargemodeltooccuratalower-rank level,effectivelyreducingthecomputationalworkload.Fromtheresults,theirproposedmethod,DRONE,notonly ensurestheinferenceperformanceofthelargemodelbutalsoachievesanaccelerationratioofmorethan1.3times comparedtothebaselinemethod.Thespecificchoiceoflow-rankdecompositionmethoddependsonthearchitecture oftheneuralnetworkandtherequirementsofthetargetapplication. YihengLiuetal.:PreprintsubmittedtoElsevier Page19of30 AComprehensiveOverviewfromTrainingtoInference 4.2. MemoryScheduling DeployingLLMsonasingleconsumer-gradeGPUisconstrainedbythelimitationsoftheavailablevideomemory, giventhesubstantialparametersofLLMs.Therefore,appropriateMemorySchedulingstrategiescanbeusedtosolve thehardwarelimitationsoflargemodelinference.Memoryschedulinginlargemodelinferenceinvolvestheefficient organizationandmanagementofmemoryaccesspatternsduringthereasoningorinferencephaseofcomplexneural networkmodels.Inthecontextofsophisticatedreasoningtasks,suchasnaturallanguageunderstandingorcomplex decision-making,largemodelsoftenhaveintricatearchitecturesandconsiderablememoryrequirements.Memory schedulingoptimizestheretrievalandstorageofintermediaterepresentations,modelparameters,andactivationvalues, ensuringthattheinferenceprocessisbothaccurateandperformedwithminimallatency.Forexample,BMInf[184] utilizestheprincipleofvirtualmemory,achievingefficientinferenceforlargemodelsbyintelligentlyschedulingthe parametersofeachlayerbetweentheGPUandCPU. 4.3. Parallelism Bothinferenceandtrainingcanleverageparallelizationtechniques.Presently,parallelizationtechniquesfor inferenceprimarilymanifestacrossthreedimensions:DataParallelism,TensorParallelism,andPipelineParallelism. DataParallelismprimarilyinvolvesincreasingtheoverallthroughputoftheinferencesystembyaddingmoreGPU devices[101;97;159;185].Tensorparallelismisaformofmodelparallelismwherethemodel’sparametersare partitionedintomultipletensors,eachcomputedondifferentprocessingunits.Thisapproachprovesbeneficialwhen dealingwithmodelsthataretoolargetofitintothememoryofasingleGPU.Tensorparallelismprimarilyinvolves increasingthenumberofdeviceshorizontallythroughparallelcomputationtoreducelatency[96].Pipelineparallelism primarilyinvolvesverticallyincreasingthenumberofGPUdevicesthroughparallelcomputationtosupportlarger modelsandenhancedeviceutiliza

185].Tensorparallelismisaformofmodelparallelismwherethemodel’sparametersare partitionedintomultipletensors,eachcomputedondifferentprocessingunits.Thisapproachprovesbeneficialwhen dealingwithmodelsthataretoolargetofitintothememoryofasingleGPU.Tensorparallelismprimarilyinvolves increasingthenumberofdeviceshorizontallythroughparallelcomputationtoreducelatency[96].Pipelineparallelism primarilyinvolvesverticallyincreasingthenumberofGPUdevicesthroughparallelcomputationtosupportlarger modelsandenhancedeviceutilization.Typically,itiscombinedwithtensorparallelismtoachieveoptimalperformance [98]. 4.4. StructuralOptimization IntheforwardpropagationcomputationofLLMs,thecalculationspeedissignificantlyfasterthanthespeed ofmemoryaccess.Inferencespeedcanbeimpactedbynumerousmemoryaccessoperations.OnegoalinLLM inferenceistominimizethenumberofmemoryaccessesduringforwardpropagation.FlashAttention[186]and PagedAttention[187]enhancecomputationalspeedbyemployingachunkedcomputationapproach,mitigatingthe storageoverheadassociatedwithmatrices.TheentireoperationtakesplacewithinSRAM,reducingthenumberof accessestoHighBandwidthMemory(HBM)andsignificantlyboostingcomputationalspeed.BothFlashAttention andPagedAttentionhavebeenadoptedbymainstreaminferenceframeworks,andseamlesslyintegratedintothese frameworksforstraightforwardutilization. 4.5. InferenceFramework Parallelcomputing,modelcompression,memoryscheduling,andspecificoptimizationsfortransformerstructures, allintegraltoLLMinference,havebeeneffectivelyimplementedinmainstreaminferenceframeworks.These frameworksfurnishthefoundationalinfrastructureandtoolsrequiredfordeployingandrunningLLMmodels.They offeraspectrumoftoolsandinterfaces,streamliningthedeploymentandinferenceprocessesforresearchersand engineersacrossdiverseapplicationscenarios.Thechoiceofaframeworktypicallyhingesonprojectrequirements, hardwaresupport,anduserpreferences.InTable4,wecompilesomeoftheseframeworksforreference. 5. UtilizationofLLMs TheapplicationscopeofLLMsisextensiveandcanbepracticallyemployedinalmostanyspecializeddomain [1;193;46;194;195].Followingpre-trainingandfine-tuning,LLMsareprimarilyutilizedbydesigningsuitable promptsforvarioustasks.Leveragingpowerfulzero-shotcapabilities,manytaskscanbedirectlyaccomplishedby guidingLLMswithstraightforwardprompts.Formorecomplextasksthatcannotbeachievedthroughsimpleprompts, afew-shotapproachinvolvingin-contextlearningisemployedtoguideLLMsintaskcompletion.Additionally, incorporatingchain-of-thought[196;197]promptsinthepromptenhancesin-contextlearningbyintroducinga reasoningprocess.Thepipelineofthein-contextlearningandchain-of-thoughtisshowninFigure6.Insome specializedresearchdirections,obtainingintermediatelayerrepresentationsofLLMsmaybenecessary.Forinstance, inneurosciencestudies,embeddingrepresentationsfromthemodelareusedtoinvestigateactivationregionsofbrain functions[198;199;200;201]. YihengLiuetal.:PreprintsubmittedtoElsevier Page20of30 AComprehensiveOverviewfromTrainingtoInference Table 4 ListofLLMinferenceframework. Framework Links TensorRT https://github.com/NVIDIA/TensorRT-LLM FasterTransformer https://github.com/NVIDIA/FasterTransformer Megatron-LM[96] https://github.com/NVIDIA/Megatron-LM FlexGen[188] https://github.com/FMInference/FlexGen DeepSpeed[158] https://github.com/microsoft/DeepSpeed vLLM[187] https://github.com/vllm-project/vllm FlexFlow[189] https://github.com/flexflow/FlexFlow StreamingLLM[190] https://github.com/mit-han-lab/streaming-llm ColossalAI[163] https://github.com/hpcaitech/ColossalAI BMCook[191

eepSpeed[158] https://github.com/microsoft/DeepSpeed vLLM[187] https://github.com/vllm-project/vllm FlexFlow[189] https://github.com/flexflow/FlexFlow StreamingLLM[190] https://github.com/mit-han-lab/streaming-llm ColossalAI[163] https://github.com/hpcaitech/ColossalAI BMCook[191] https://github.com/OpenBMB/BMCook BMInf[184] https://github.com/OpenBMB/BMInf Petals[192] https://github.com/bigscience-workshop/petals Figure 6:A)in-contextlearning,B)Chainofthought. Generally,thereareseveralapproachestoemployingLLMs.Thefirstinvolvesaccessingthecapabilitiesofrobust proprietarymodelsthroughopenAPIservices,suchasutilizingtheAPIprovidedbyChatGPT[19].Thesecond approachincludesdeployingopen-sourceLLMsforlocaluse[9].Thethirdmethodentailsfine-tuningopen-source LLMstomeetspecificdomainstandards[43;202],enablingtheirapplicationinaparticularfield,andsubsequently deployingthemlocally.InTable5,wehavecompiledinformationonvariousopen-sourceLLMsforreference. Researcherscanchoosefromtheseopen-sourceLLMstodeployapplicationsthatbestsuittheirneeds. 6. FutureDirectionsandImplications ThissectionwilldelveintothefuturetrendsandimpactofLLMtechnology.Ourdiscussionwillbestructuredinto threeparts:firstly,anexplorationofthedevelopmentaltrendswithinLLMstechnologyitself;secondly,anexamination ofthedevelopmentaldirectionsforAIresearchers;andfinally,ananalysisofthesocietalimpactresultingfromthe ongoingdevelopmentofLLMs. Basedonexistingexperiences,itisevidentthatanamplesupplyofhigh-qualitydataandasufficientnumberof parameterssignificantlycontributetoenhancingtheperformanceofmodels[8].Lookingahead,themodelscaleof YihengLiuetal.:PreprintsubmittedtoElsevier Page21of30 AComprehensiveOverviewfromTrainingtoInference Table 5 ListofopensourceLLMs. LLM Size(B) Links T5[68] 11B https://github.com/google-research/text-to-text-transfer-transformer CodeGen[81] 16B https://github.com/salesforce/CodeGen MOSS[203] 16B https://github.com/OpenLMLab/MOSS GLM[37] 130B https://github.com/THUDM/GLM ChatGLM[37] 6B https://github.com/THUDM/ChatGLM3 ChatYuan[204] 0.7B https://github.com/clue-ai/ChatYuan OPT[83] 175B https://github.com/facebookresearch/metaseq BLOOM[38] 176B https://huggingface.co/bigscience/bloom LLaMA[9] 65B https://github.com/facebookresearch/llama CodeGeeX[82] 13B https://github.com/THUDM/CodeGeeX Baichuan[205] 13B https://github.com/baichuan-inc/Baichuan2 Aquila 7B https://github.com/FlagAI-Open/FlagAI/tree/master/examples/Aquila MiniGPT-4[206] 25B https://github.com/Vision-CAIR/MiniGPT-4 Vicuna[207] 13B https://github.com/lm-sys/FastChat LLMsisexpectedtocontinueexpanding,therebyaugmentingtheirlearningcapabilitiesandoverallperformance. Moreover,themajorityofcurrentlyavailableLLMsareconfinedtoasinglenaturallanguagemodality,lacking extensionstoprocessmultimodaldatasuchasimages,videos,andspeech.Thereisapotentialfuturetrajectoryfor LLMstoevo

-4[206] 25B https://github.com/Vision-CAIR/MiniGPT-4 Vicuna[207] 13B https://github.com/lm-sys/FastChat LLMsisexpectedtocontinueexpanding,therebyaugmentingtheirlearningcapabilitiesandoverallperformance. Moreover,themajorityofcurrentlyavailableLLMsareconfinedtoasinglenaturallanguagemodality,lacking extensionstoprocessmultimodaldatasuchasimages,videos,andspeech.Thereisapotentialfuturetrajectoryfor LLMstoevolvetowardshandlinginformationbeyondtext,incorporatingmultimodaldatalikeimagesandaudio. Thisevolutionwouldempowermodelstocomprehensivelyunderstandandgeneratemultimodalcontent,significantly broadeningtheapplicationscopeofLLMs.TheinevitableexpansionofLLMsintothefieldofmultimodalityisbound toincurincreasedtrainingcosts.Apivotalfocusforfuturedevelopmentsliesintheefficientfine-tuningofparameters andthedeploymentofLLMsthroughtechniquessuchasknowledgedistillation,modelcompression,andquantization, aimedatreducingboththetrainingandinferencecostsofLLMs.Anotheremergingtrendisthedomain-specifictraining andfine-tuningofLLMsforparticularsectors,facilitatingamoreadeptadaptationtoandunderstandingofindustry- specificterminologiesandcontexts.Lastly,intheexplorationofpotentialnewarchitecturesforLLMsthecurrent landscapepredominantlyreliesonthetransformerarchitecture.Whilethetransformerarchitecturenaturallyboasts advantagessuchasparallelcomputingandadaptabilitytovariousinputmodalities,itsdesigntypicallynecessitates fixed-sizeinputs.Thisrequirementmaynecessitatepaddingortruncationwhendealingwithvariable-lengthsequences, potentiallyleadingtocomputationalandinformationinefficiencies,aswellaschallengesingeneratingcoherentdata. InvestigatingthepotentialofRecurrentNeuralNetwork(RNN)architecturesintheeraofLLMscouldemergeasa pivotalresearchdirection.Forinstance,RWKV[208],anLLMdesignedundertheRNNarchitecture,hasdemonstrated competitiveperformanceonvariousthird-partyevaluations,provingitselfcomparabletothemajorityoftransformer- basedLLMs. ForresearchersinthefieldofAI,workinginisolationisbecomingincreasinglyimpractical.Thefuturedirection ofAIdevelopmentwillintertwinewithvariousindustries,necessitatingclosecollaborationwithprofessionalsfrom diversefields.Itiscrucialtoengageincollaborativeefforts,bridgingresearchdisciplines,andcollectivelyaddressing challengesbycombiningexpertisefromdifferentdomains.Simultaneously,thereisafreshsetofrequirementsforthe comprehensiveskillsofAIresearchers.TraininganddeployingLLMsnecessitateproficiencyinmanaginglarge-scale dataandsubstantialpracticalexperienceindistributedparalleltraining.Thiscriterionunderscorestheimportancefor researchersinvolvedinLLMdevelopmenttopossesssubstantialengineeringcapabilities,addressingthechallenges inherentintheprocess.ResearcherswhoareinterestedinthefieldofLLMsmusteitherpossessengineeringskillsor adeptlycollaboratewithengineerstonavigatethecomplexitiesofmodeldevelopment[3]. AsLLMsfindwidespreadapplicationsinsocietallife,concernsaboutethicalissuesandsocietalimpactareona continuousrise.Thismayinvolveresearchandimprovementsinareassuchasmanagingmodelbiasesandcontrolling theriskofmisuse[4].Consideringtheparamountimportanceofprivacyanddatasecurity,thefuturedevelopment ofLLMsmightinvolvemorefederatedlearninganddecentralizedapproachestoenhancemodelperformancewhile safeguardinguserprivacy.Developersshouldengageininterdisciplinarycollaborationwithexpertsfromvarious fields,includingdecision-making,legalstudies,andsociology,toestablishstandardsandethicalframeworksforthe YihengLiuetal.:PreprintsubmittedtoElsevier Page22of30 AComprehensiveOverviewfromTrainingtoInference development,deployment,andutilizationofLLMs,mitigatingpotentialharmfulconsequences.Intermsofpublic awarenessandeducation,mandatoryawarenesstrainingshouldbeimplementedbeforelarge-scalepublicdeployment andapplications.ThisaimstoenhancepublicunderstandingofthecapabilitiesandlimitationsofLLMs,fostering responsibleandinformeduse,especiallyinindustriessuc

ubmittedtoElsevier Page22of30 AComprehensiveOverviewfromTrainingtoInference development,deployment,andutilizationofLLMs,mitigatingpotentialharmfulconsequences.Intermsofpublic awarenessandeducation,mandatoryawarenesstrainingshouldbeimplementedbeforelarge-scalepublicdeployment andapplications.ThisaimstoenhancepublicunderstandingofthecapabilitiesandlimitationsofLLMs,fostering responsibleandinformeduse,especiallyinindustriessuchaseducationandjournalism. 7. Conclusion TheintroductionofChatGPThasusheredinatransformativeeraintherealmofLargeLLMs,significantly influencingtheirutilizationfordiversedownstreamtasks.Theemphasisoncost-effectivetraininganddeployment hasemergedasacrucialaspectintheevolutionofLLMs.Thispaperhasprovidedacomprehensivesurveyofthe evolutionoflargelanguagemodeltrainingtechniquesandinferencedeploymenttechnologiesinalignmentwith theemergingtrendoflow-costdevelopment.Theprogressionfromtraditionalstatisticallanguagemodelstoneural languagemodels,andsubsequentlytoPLMssuchasELMoandtransformerarchitecture,hassetthestageforthe dominanceofLLMs.Thescaleandperformanceofthesemodels,particularlyexemplifiedbytheGPTseries,have reachedunprecedentedlevels,showcasingthephenomenonofemergenceandenablingversatileapplicationsacross variousdomains.Notably,thereleaseofChatGPTbyOpenAIinNovember2022hasmarkedapivotalmomentin theLLMlandscape,revolutionizingthestrengthandeffectivenessofAIalgorithms.However,thecurrentreliance onOpenAI’sinfrastructureunderscoresthenecessityforalternativeLLMs,emphasizingtheneedfordomain-specific modelsandadvancementsinthetraininganddeploymentprocesses. TraininganddeployingLLMspresentchallengesthatdemandexpertiseinhandlinglarge-scaledataanddistributed paralleltraining.TheengineeringcapabilitiesrequiredforLLMdevelopmenthighlightthecollaborativeeffortsneeded betweenresearchersandengineers.AsweexplorethetechnicalaspectsofLLMtrainingandinferenceinthisreview, itbecomesevidentthatadeepunderstandingoftheseprocessesisessentialforresearchersventuringintothefield. Lookingahead,thefutureofLLMsholdspromisingdirections,includingfurtheradvancementsinmodelarchitectures, improvedtrainingefficiency,andbroaderapplicationsacrossindustries.Theinsightsprovidedinthisreviewaimto equipresearcherswiththeknowledgeandunderstandingnecessarytonavigatethecomplexitiesofLLMdevelopment, fosteringinnovationandprogressinthisdynamicfield.AsLLMscontinuetoevolve,theirimpactonnaturallanguage processingandAIasawholeispoisedtoshapethefuturelandscapeofintelligentsystems. References [1]Y.Liu,T.Han,S.Ma,J.Zhang,Y.Yang,J.Tian,H.He,A.Li,M.He,Z.Liu etal.,“Summaryofchatgpt-relatedresearchandperspective towardsthefutureoflargelanguagemodels,”Meta-Radiology,p.100017,2023. [2]J.Wang,E.Shi,S.Yu,Z.Wu,C.Ma,H.Dai,Q.Yang,Y.Kang,J.Wu,H.Hu etal.,“Promptengineeringforhealthcare:Methodologiesand applications,”arXivpreprintarXiv:2304.14670,2023. [3]W.X.Zhao,K.Zhou,J.Li,T.Tang,X.Wang,Y.Hou,Y.Min,B.Zhang,J.Zhang,Z.Dong etal.,“Asurveyoflargelanguagemodels,” arXivpreprintarXiv:2303.18223,2023. [4]J.Kaddour,J.Harris,M.Mozes,H.Bradley,R.Raileanu,andR.McHardy,“Challengesandapplicationsoflargelanguagemodels,” arXiv preprintarXiv:2307.10169,2023. [5]M.E.Peters,M.Neumann,M.Iyyer,M.Gardner,C.Clark,K.Lee,andL.Zettlemoyer,“Deepcontextualizedwordrepresentations,”in Proceedingsofthe2018ConferenceoftheNorthAmericanChapteroftheAssociationforComputationalLinguistics:HumanLanguage Technologies,Volume1(LongPapers),Jun.2018,pp.2227–2237. [6]A.Vaswani,N.Shazeer,N.Parmar,J.Uszkoreit,L.Jones,A.N.Gomez,Ł.Kaiser,andI.Polosukhin,“Attentionisallyouneed,” Advances inneuralinformationprocessingsystems,vol.30,2017. [7]A.Radford,J.Wu,D.Amodei,D.Amodei,J.Clark,M.Brundage,andI.Sutskever,“Betterlanguagemodelsandtheirimplications,” OpenAI Bloghttps://openai.com/blog/better-language-models,vol.1,no.2,2019. [8]T.Brown,B.Mann,N.Ryder,M.Subbiah,J.D.Kaplan,P.Dhariwal,A.Neelakantan,P.Shyam,G.Sastry,A.Askell etal.,“Language modelsarefew-shotlear

wani,N.Shazeer,N.Parmar,J.Uszkoreit,L.Jones,A.N.Gomez,Ł.Kaiser,andI.Polosukhin,“Attentionisallyouneed,” Advances inneuralinformationprocessingsystems,vol.30,2017. [7]A.Radford,J.Wu,D.Amodei,D.Amodei,J.Clark,M.Brundage,andI.Sutskever,“Betterlanguagemodelsandtheirimplications,” OpenAI Bloghttps://openai.com/blog/better-language-models,vol.1,no.2,2019. [8]T.Brown,B.Mann,N.Ryder,M.Subbiah,J.D.Kaplan,P.Dhariwal,A.Neelakantan,P.Shyam,G.Sastry,A.Askell etal.,“Language modelsarefew-shotlearners,”Advancesinneuralinformationprocessingsystems,vol.33,pp.1877–1901,2020. [9]H.Touvron,T.Lavril,G.Izacard,X.Martinet,M.-A.Lachaux,T.Lacroix,B.Rozière,N.Goyal,E.Hambro,F.Azhar etal.,“Llama:Open andefficientfoundationlanguagemodels,”arXivpreprintarXiv:2302.13971,2023. [10]H.Touvron,L.Martin,K.Stone,P.Albert,A.Almahairi,Y.Babaei,N.Bashlykov,S.Batra,P.Bhargava,S.Bhosale etal.,“Llama2:Open foundationandfine-tunedchatmodels,”arXivpreprintarXiv:2307.09288,2023. [11]S.Rezayi,H.Dai,Z.Liu,Z.Wu,A.Hebbar,A.H.Burns,L.Zhao,D.Zhu,Q.Li,W.Liu etal.,“Clinicalradiobert:Knowledge-infusedfew shotlearningforclinicalnotesnamedentityrecognition,”inInternationalWorkshoponMachineLearninginMedicalImaging. Springer, 2022,pp.269–278. [12]Z.Liu,M.He,Z.Jiang,Z.Wu,H.Dai,L.Zhang,S.Luo,T.Han,X.Li,X.Jiang etal.,“Surveyonnaturallanguageprocessinginmedical imageanalysis.”Zhongnandaxuexuebao.Yixueban=JournalofCentralSouthUniversity.MedicalSciences,vol.47,no.8,pp.981–993, 2022. YihengLiuetal.:PreprintsubmittedtoElsevier Page23of30 AComprehensiveOverviewfromTrainingtoInference [13]W.Liao,Z.Liu,H.Dai,Z.Wu,Y.Zhang,X.Huang,Y.Chen,X.Jiang,D.Zhu,T.Liu etal.,“Mask-guidedbertforfewshottextclassification,” arXivpreprintarXiv:2302.10447,2023. [14]S.Rezayi,Z.Liu,Z.Wu,C.Dhakal,B.Ge,H.Dai,G.Mai,N.Liu,C.Zhen,T.Liu etal.,“Exploringnewfrontiersinagriculturalnlp: Investigatingthepotentialoflargelanguagemodelsforfoodapplications,”arXivpreprintarXiv:2306.11892,2023. [15]T.Zhong,W.Zhao,Y.Zhang,Y.Pan,P.Dong,Z.Jiang,X.Kui,Y.Shang,L.Yang,Y.Wei etal.,“Chatradio-valuer:Achatlargelanguage modelforgeneralizableradiologyreportgenerationbasedonmulti-institutionandmulti-systemdata,”arXivpreprintarXiv:2310.05242, 2023. [16]Z.Liu,T.Zhong,Y.Li,Y.Zhang,Y.Pan,Z.Zhao,P.Dong,C.Cao,Y.Liu,P.Shu etal.,“Evaluatinglargelanguagemodelsforradiology naturallanguageprocessing,”arXivpreprintarXiv:2307.13693,2023. [17]T.Zhong,Y.Wei,L.Yang,Z.Wu,Z.Liu,X.Wei,W.Li,J.Yao,C.Ma,X.Li etal.,“Chatabl:Abductivelearningvianaturallanguage interactionwithchatgpt,”arXivpreprintarXiv:2304.11107,2023. [18]A.Radford,K.Narasimhan,T.Salimans,I.Sutskever etal.,“Improvinglanguageunderstandingbygenerativepre-training,”OpenAI,2018. [19]OpenAI,“Gpt-4technicalreport,”2023. [20]H.Dai,Z.Liu,W.Liao,X.Huang,Y.Cao,Z.Wu,L.Zhao,S.Xu,W.Liu,N.Liu,S.Li,D.Zhu,H.Cai,L.Sun,Q.Li,D.Shen,T.Liu,and X.Li,“Auggpt:Leveragingchatgptfortextdataaugmentation,”2023. [21]Z.Liu,X.Yu,L.Zhang,Z.Wu,C.Cao,H.Dai,L.Zhao,W.Liu,D.Shen,Q.Li etal.,“Deid-gpt:Zero-shotmedicaltextde-identification bygpt-4,”arXivpreprintarXiv:2303.11032,2023. [22]C.Ma,Z.Wu,J.Wang,S.Xu,Y.Wei,Z.Liu,L.Guo,X.Cai,S.Zhang,T.Zhang etal.,“Impressiongpt:aniterativeoptimizingframework forradiologyreportsummarizationwithchatgpt,”arXivpreprintarXiv:2304.08448,2023. [23]W.Liao,Z.Liu,H.Dai,S.Xu,Z.Wu,Y.Zhang,X.Huang,D.Zhu,H.Cai,T.Liu etal.,“Differentiatechatgpt-generatedandhuman-written medicaltexts,”arXivpreprintarXiv:2304.11567,2023. [24]H.Dai,Y.Li,Z.Liu,L.Zhao,Z.Wu,S.Song,Y.Shen,D.Zhu,X.Li,S.Li etal.,“Ad-autogpt:Anautonomousgptforalzheimer’sdisease infodemiology,”arXivpreprintarXiv:2306.10095,2023. [25]Z.Guan,Z.Wu,Z.Liu,D.Wu,H.Ren,Q.Li,X.Li,andN.Liu,“Cohortgpt:Anenhancedgptforparticipantrecruitmentinclinicalstudy,” arXivpreprintarXiv:2307.11346,2023. [26]Z.Liu,Z.Wu,M.Hu,B.Zhao,L.Zhao,T.Zhang,H.Dai,X.Chen,Y.Shen,S.Li etal.,“Pharmacygpt:Theaiphar

dhuman-written medicaltexts,”arXivpreprintarXiv:2304.11567,2023. [24]H.Dai,Y.Li,Z.Liu,L.Zhao,Z.Wu,S.Song,Y.Shen,D.Zhu,X.Li,S.Li etal.,“Ad-autogpt:Anautonomousgptforalzheimer’sdisease infodemiology,”arXivpreprintarXiv:2306.10095,2023. [25]Z.Guan,Z.Wu,Z.Liu,D.Wu,H.Ren,Q.Li,X.Li,andN.Liu,“Cohortgpt:Anenhancedgptforparticipantrecruitmentinclinicalstudy,” arXivpreprintarXiv:2307.11346,2023. [26]Z.Liu,Z.Wu,M.Hu,B.Zhao,L.Zhao,T.Zhang,H.Dai,X.Chen,Y.Shen,S.Li etal.,“Pharmacygpt:Theaipharmacist,”arXivpreprint arXiv:2307.10432,2023. [27]Y.Wei,T.Zhang,H.Zhang,T.Zhong,L.Zhao,Z.Liu,C.Ma,S.Zhang,M.Shang,L.Du etal.,“Chat2brain:Amethodformapping open-endedsemanticqueriestobrainactivationmaps,”arXivpreprintarXiv:2309.05021,2023. [28]T.Zhong,X.Wei,E.Shi,J.Gao,C.Ma,Y.Wei,S.Zhang,L.Guo,J.Han,T.Liu etal.,“Asmall-samplemethodwitheegsignalsbased onabductivelearningformotorimagerydecoding,”inInternationalConferenceonMedicalImageComputingandComputer-Assisted Intervention. Springer,2023,pp.416–424. [29]J.Gao,L.Zhao,T.Zhong,C.Li,Z.He,Y.Wei,S.Zhang,L.Guo,T.Liu,J.Han etal.,“Predictionofcognitivescoresbyjointuseof movie-watchingfmriconnectivityandeyetrackingviaattention-censnet,”Psychoradiology,vol.3,2023. [30]I.Sutskever,O.Vinyals,andQ.V.Le,“Sequencetosequencelearningwithneuralnetworks,” Advancesinneuralinformationprocessing systems,vol.27,2014. [31]G.BebisandM.Georgiopoulos,“Feed-forwardneuralnetworks,” IeeePotentials,vol.13,no.4,pp.27–31,1994. [32]Y.Yang,L.Wang,S.Shi,P.Tadepalli,S.Lee,andZ.Tu,“Onthesub-layerfunctionalitiesoftransformerdecoder,” arXivpreprint arXiv:2010.02648,2020. [33]Z.Dai,Z.Yang,Y.Yang,J.Carbonell,Q.V.Le,andR.Salakhutdinov,“Transformer-xl:Attentivelanguagemodelsbeyondafixed-length context,”arXivpreprintarXiv:1901.02860,2019. [34]J.Su,M.Ahmed,Y.Lu,S.Pan,W.Bo,andY.Liu,“Roformer:Enhancedtransformerwithrotarypositionembedding,” Neurocomputing,p. 127063,2023. [35]O.Press,N.A.Smith,andM.Lewis,“Trainshort,testlong:Attentionwithlinearbiasesenablesinputlengthextrapolation,” arXivpreprint arXiv:2108.12409,2021. [36]A.Chowdhery,S.Narang,J.Devlin,M.Bosma,G.Mishra,A.Roberts,P.Barham,H.W.Chung,C.Sutton,S.Gehrmann etal.,“Palm: Scalinglanguagemodelingwithpathways,”arXivpreprintarXiv:2204.02311,2022. [37]A.Zeng,X.Liu,Z.Du,Z.Wang,H.Lai,M.Ding,Z.Yang,Y.Xu,W.Zheng,X.Xia etal.,“Glm-130b:Anopenbilingualpre-trained model,”arXivpreprintarXiv:2210.02414,2022. [38]B.Workshop,T.L.Scao,A.Fan,C.Akiki,E.Pavlick,S.Ilić,D.Hesslow,R.Castagné,A.S.Luccioni,F.Yvon etal.,“Bloom:A176b- parameteropen-accessmultilinguallanguagemodel,”arXivpreprintarXiv:2211.05100,2022. [39]L.Zhao,L.Zhang,Z.Wu,Y.Chen,H.Dai,X.Yu,Z.Liu,T.Zhang,X.Hu,X.Jiang etal.,“Whenbrain-inspiredaimeetsagi,”Meta- Radiology,p.100005,2023. [40]J.Holmes,Z.Liu,L.Zhang,Y.Ding,T.T.Sio,L.A.McGee,J.B.Ashman,X.Li,T.Liu,J.Shen,andW.Liu,“Evaluatinglargelanguage modelsonahighly-specializedtopic,radiationoncologyphysics,”FrontiersinOncology,vol.13,Jul.2023. [41]Z.Wu,L.Zhang,C.Cao,X.Yu,H.Dai,C.Ma,Z.Liu,L.Zhao,G.Li,W.Liu etal.,“Exploringthetrade-offs:Unifiedlargelanguagemodels vslocalfine-tunedmodelsforhighly-specificradiologynlitask,”arXivpreprintarXiv:2304.09138,2023. [42]S.Rezayi,Z.Liu,Z.Wu,C.Dhakal,B.Ge,C.Zhen,T.Liu,andS.Li,“Agribert:knowledge-infusedagriculturallanguagemodelsfor matchingfoodandnutrition,”inProceedingsoftheThirty-FirstInternationalJointConferenceonArtificialIntelligence,vol.7,2022,pp. 5150–5156. [43]Z.Liu,A.Zhong,Y.Li,L.Yang,C.Ju,Z.Wu,C.Ma,P.Shu,C.Chen,S.Kim etal.,“Radiology-gpt:Alargelanguagemodelforradiology,” arXivpreprintarXiv:2306.08666,2023. YihengLiuetal.:PreprintsubmittedtoElsevier Page24of30 AComprehensiveOverviewfromTrainingtoInference [44]Z.Liu,X.He,L.Liu,T.Liu,andX.Zhai, ContextMatters:AStrategytoPre-trainLanguageModelforScienceEducation. SpringerNature Switzerland,2023,p.666–674.

]Z.Liu,A.Zhong,Y.Li,L.Yang,C.Ju,Z.Wu,C.Ma,P.Shu,C.Chen,S.Kim etal.,“Radiology-gpt:Alargelanguagemodelforradiology,” arXivpreprintarXiv:2306.08666,2023. YihengLiuetal.:PreprintsubmittedtoElsevier Page24of30 AComprehensiveOverviewfromTrainingtoInference [44]Z.Liu,X.He,L.Liu,T.Liu,andX.Zhai, ContextMatters:AStrategytoPre-trainLanguageModelforScienceEducation. SpringerNature Switzerland,2023,p.666–674. [45]J.Wang,Z.Liu,L.Zhao,Z.Wu,C.Ma,S.Yu,H.Dai,Q.Yang,Y.Liu,S.Zhang etal.,“Reviewoflargevisionmodelsandvisualprompt engineering,”arXivpreprintarXiv:2307.00855,2023. [46]X.Li,L.Zhang,Z.Wu,Z.Liu,L.Zhao,Y.Yuan,J.Liu,G.Li,D.Zhu,P.Yan etal.,“Artificialgeneralintelligenceformedicalimaging,” arXivpreprintarXiv:2306.05480,2023. [47]H.Cai,W.Liao,Z.Liu,Y.Zhang,X.Huang,S.Ding,H.Ren,Z.Wu,H.Dai,S.Li etal.,“Coarse-to-fineknowledgegraphdomainadaptation basedondistantly-supervisediterativetraining,”arXivpreprintarXiv:2211.02849,2022. [48]H.Dai,C.Ma,Z.Liu,Y.Li,P.Shu,X.Wei,L.Zhao,Z.Wu,D.Zhu,W.Liu etal.,“Samaug:Pointpromptaugmentationforsegment anythingmodel,”arXivpreprintarXiv:2307.01187,2023. [49]L.Zhang,Z.Liu,L.Zhang,Z.Wu,X.Yu,J.Holmes,H.Feng,H.Dai,X.Li,Q.Li etal.,“Segmentanythingmodel(sam)forradiation oncology,”arXivpreprintarXiv:2306.11730,2023. [50]Z.Xiao,Y.Chen,L.Zhang,J.Yao,Z.Wu,X.Yu,Y.Pan,L.Zhao,C.Ma,X.Liu etal.,“Instruction-vit:Multi-modalpromptsforinstruction learninginvit,”arXivpreprintarXiv:2305.00201,2023. [51]P.Liu,W.Yuan,J.Fu,Z.Jiang,H.Hayashi,andG.Neubig,“Pre-train,prompt,andpredict:Asystematicsurveyofpromptingmethodsin naturallanguageprocessing,”ACMComputingSurveys,vol.55,no.9,pp.1–35,2023. [52]S.B.Kotsiantis,I.Zaharakis,P.Pintelas etal.,“Supervisedmachinelearning:Areviewofclassificationtechniques,”Emergingartificial intelligenceapplicationsincomputerengineering,vol.160,no.1,pp.3–24,2007. [53]Y.Bengio,A.Courville,andP.Vincent,“Representationlearning:Areviewandnewperspectives,” IEEEtransactionsonpatternanalysis andmachineintelligence,vol.35,no.8,pp.1798–1828,2013. [54]L.Dong,N.Yang,W.Wang,F.Wei,X.Liu,Y.Wang,J.Gao,M.Zhou,andH.-W.Hon,“Unifiedlanguagemodelpre-trainingfornatural languageunderstandingandgeneration,”Advancesinneuralinformationprocessingsystems,vol.32,2019. [55]T.SchickandH.Schütze,“It’snotjustsizethatmatters:Smalllanguagemodelsarealsofew-shotlearners,” arXivpreprintarXiv:2009.07118, 2020. [56]F.Petroni,T.Rocktäschel,P.Lewis,A.Bakhtin,Y.Wu,A.H.Miller,andS.Riedel,“Languagemodelsasknowledgebases?” arXivpreprint arXiv:1909.01066,2019. [57]B.Lester,R.Al-Rfou,andN.Constant,“Thepowerofscaleforparameter-efficientprompttuning,” arXivpreprintarXiv:2104.08691,2021. [58]T.SchickandH.Schütze,“Exploitingclozequestionsforfewshottextclassificationandnaturallanguageinference,” arXivpreprint arXiv:2001.07676,2020. [59]R.Shin,C.H.Lin,S.Thomson,C.Chen,S.Roy,E.A.Platanios,A.Pauls,D.Klein,J.Eisner,andB.VanDurme,“Constrainedlanguage modelsyieldfew-shotsemanticparsers,”arXivpreprintarXiv:2104.08768,2021. [60]Z.Jiang,F.F.Xu,J.Araki,andG.Neubig,“Howcanweknowwhatlanguagemodelsknow?” TransactionsoftheAssociationfor ComputationalLinguistics,vol.8,pp.423–438,2020. [61]K.Duh,K.Sudoh,X.Wu,H.Tsukada,andM.Nagata,“Generalizedminimumbayesrisksystemcombination,”in Proceedingsof5th InternationalJointConferenceonNaturalLanguageProcessing,2011,pp.1356–1360. [62]Z.Jiang,J.Araki,H.Ding,andG.Neubig,“Howcanweknowwhenlanguagemodelsknow?onthecalibrationoflanguagemodelsfor questionanswering,”TransactionsoftheAssociationforComputationalLinguistics,vol.9,pp.962–977,2021. [63]M.McCloskeyandN.J.Cohen,“Catastrophicinterferenceinconnectionistnetworks:Thesequentiallearningproblem,”in Psychologyof learningandmotivation. Elsevier,1989,vol.24,pp.109–165. [64]T.Brown,B.Mann,N.Ryder,M.Subbiah,J.D.Kaplan,P.Dhariwal,A.Neelakantan,P.Shyam,G.Sastry,A.Askell etal.,“Language modelsarefew-shotle

,“Howcanweknowwhenlanguagemodelsknow?onthecalibrationoflanguagemodelsfor questionanswering,”TransactionsoftheAssociationforComputationalLinguistics,vol.9,pp.962–977,2021. [63]M.McCloskeyandN.J.Cohen,“Catastrophicinterferenceinconnectionistnetworks:Thesequentiallearningproblem,”in Psychologyof learningandmotivation. Elsevier,1989,vol.24,pp.109–165. [64]T.Brown,B.Mann,N.Ryder,M.Subbiah,J.D.Kaplan,P.Dhariwal,A.Neelakantan,P.Shyam,G.Sastry,A.Askell etal.,“Language modelsarefew-shotlearners,”Advancesinneuralinformationprocessingsystems,vol.33,pp.1877–1901,2020. [65]Y.Zhu,R.Kiros,R.Zemel,R.Salakhutdinov,R.Urtasun,A.Torralba,andS.Fidler,“Aligningbooksandmovies:Towardsstory-likevisual explanationsbywatchingmoviesandreadingbooks,”inProceedingsoftheIEEEinternationalconferenceoncomputervision,2015,pp. 19–27. [66]“Projectgutenberg.”[Online].Available:https://www.gutenberg.org/ [67]“Commoncrawl.”[Online].Available:https://commoncrawl.org/ [68]C.Raffel,N.Shazeer,A.Roberts,K.Lee,S.Narang,M.Matena,Y.Zhou,W.Li,andP.J.Liu,“Exploringthelimitsoftransferlearning withaunifiedtext-to-texttransformer,”TheJournalofMachineLearningResearch,vol.21,no.1,pp.5485–5551,2020. [69]T.H.TrinhandQ.V.Le,“Asimplemethodforcommonsensereasoning,” arXivpreprintarXiv:1806.02847,2018. [70]Y.Liu,M.Ott,N.Goyal,J.Du,M.Joshi,D.Chen,O.Levy,M.Lewis,L.Zettlemoyer,andV.Stoyanov,“Roberta:Arobustlyoptimizedbert pretrainingapproach,”arXivpreprintarXiv:1907.11692,2019. [71]R.Zellers,A.Holtzman,H.Rashkin,Y.Bisk,A.Farhadi,F.Roesner,andY.Choi,“Defendingagainstneuralfakenews,” Advancesinneural informationprocessingsystems,vol.32,2019. [72]G.Penedo,Q.Malartic,D.Hesslow,R.Cojocaru,H.Alobeidli,A.Cappelli,B.Pannier,E.Almazrouei,andJ.Launay,“Therefinedweb datasetforfalconllm:Outperformingcuratedcorporawithwebdataonly,”inThirty-seventhConferenceonNeuralInformationProcessing SystemsDatasetsandBenchmarksTrack,2023. [73]A.Gokaslan,V.C.E.Pavlick,andS.Tellex,“Openwebtextcorpus,”http://Skylion007.github.io/OpenWebTextCorpus,2019. [74]J.Baumgartner,S.Zannettou,B.Keegan,M.Squire,andJ.Blackburn,“Thepushshiftredditdataset,”in Proceedingsoftheinternational AAAIconferenceonwebandsocialmedia,vol.14,2020,pp.830–839. [75]“Wikipedia.”[Online].Available:https://en.wikipedia.org/wiki/Main_Page [76]“Bigquerydataset.”[Online].Available:https://cloud.google.com/bigquery [77]L.Gao,S.Biderman,S.Black,L.Golding,T.Hoppe,C.Foster,J.Phang,H.He,A.Thite,N.Nabeshima etal.,“Thepile:An800gbdataset ofdiversetextforlanguagemodeling,”arXivpreprintarXiv:2101.00027,2020. YihengLiuetal.:PreprintsubmittedtoElsevier Page25of30 AComprehensiveOverviewfromTrainingtoInference [78]H.Laurençon,L.Saulnier,T.Wang,C.Akiki,A.VillanovadelMoral,T.LeScao,L.VonWerra,C.Mou,E.GonzálezPonferrada,H.Nguyen etal.,“Thebigsciencerootscorpus:A1.6tbcompositemultilingualdataset,”AdvancesinNeuralInformationProcessingSystems,vol.35, pp.31809–31826,2022. [79]S.Smith,M.Patwary,B.Norick,P.LeGresley,S.Rajbhandari,J.Casper,Z.Liu,S.Prabhumoye,G.Zerveas,V.Korthikanti etal.,“Using deepspeedandmegatrontotrainmegatron-turingnlg530b,alarge-scalegenerativelanguagemodel,”arXivpreprintarXiv:2201.11990, 2022. [80]R.Thoppilan,D.DeFreitas,J.Hall,N.Shazeer,A.Kulshreshtha,H.-T.Cheng,A.Jin,T.Bos,L.Baker,Y.Du etal.,“Lamda:Language modelsfordialogapplications,”arXivpreprintarXiv:2201.08239,2022. [81]E.Nijkamp,B.Pang,H.Hayashi,L.Tu,H.Wang,Y.Zhou,S.Savarese,andC.Xiong,“Codegen:Anopenlargelanguagemodelforcode withmtulti-turnprogramsynthesis,”arXivpreprintarXiv:2203.13474,2022. [82]Q.Zheng,X.Xia,X.Zou,Y.Dong,S.Wang,Y.Xue,L.Shen,Z.Wang,A.Wang,Y.Li etal.,“Codegeex:Apre-trainedmodelforcode generationwithmultilingualbenchmarkingonhumaneval-x,”inProceedingsofthe29thACMSIGKDDConferenceonKnowledgeDiscovery andDataMining,2023,pp.5673–5684. [83]S.Zhang,S.Roller,N.Goyal,M.Artetxe,M.Chen,S.Chen,C.Dewan,M.Diab,X.Li,X.V.Lin

H.Wang,Y.Zhou,S.Savarese,andC.Xiong,“Codegen:Anopenlargelanguagemodelforcode withmtulti-turnprogramsynthesis,”arXivpreprintarXiv:2203.13474,2022. [82]Q.Zheng,X.Xia,X.Zou,Y.Dong,S.Wang,Y.Xue,L.Shen,Z.Wang,A.Wang,Y.Li etal.,“Codegeex:Apre-trainedmodelforcode generationwithmultilingualbenchmarkingonhumaneval-x,”inProceedingsofthe29thACMSIGKDDConferenceonKnowledgeDiscovery andDataMining,2023,pp.5673–5684. [83]S.Zhang,S.Roller,N.Goyal,M.Artetxe,M.Chen,S.Chen,C.Dewan,M.Diab,X.Li,X.V.Lin etal.,“Opt:Openpre-trainedtransformer languagemodels,”arXivpreprintarXiv:2205.01068,2022. [84]H.W.Chung,L.Hou,S.Longpre,B.Zoph,Y.Tay,W.Fedus,Y.Li,X.Wang,M.Dehghani,S.Brahma etal.,“Scalinginstruction-finetuned languagemodels,”arXivpreprintarXiv:2210.11416,2022. [85]A.Radford,J.Wu,R.Child,D.Luan,D.Amodei,I.Sutskever etal.,“Languagemodelsareunsupervisedmultitasklearners,”OpenAIblog, vol.1,no.8,p.9,2019. [86]D.Hernandez,T.Brown,T.Conerly,N.DasSarma,D.Drain,S.El-Showk,N.Elhage,Z.Hatfield-Dodds,T.Henighan,T.Hume etal., “Scalinglawsandinterpretabilityoflearningfromrepeateddata,”arXivpreprintarXiv:2205.10487,2022. [87]K.Lee,D.Ippolito,A.Nystrom,C.Zhang,D.Eck,C.Callison-Burch,andN.Carlini,“Deduplicatingtrainingdatamakeslanguagemodels better,”arXivpreprintarXiv:2107.06499,2021. [88]N.Carlini,F.Tramer,E.Wallace,M.Jagielski,A.Herbert-Voss,K.Lee,A.Roberts,T.Brown,D.Song,U.Erlingsson etal.,“Extracting trainingdatafromlargelanguagemodels,”in30thUSENIXSecuritySymposium(USENIXSecurity21),2021,pp.2633–2650. [89]S.Gehman,S.Gururangan,M.Sap,Y.Choi,andN.A.Smith,“Realtoxicityprompts:Evaluatingneuraltoxicdegenerationinlanguage models,”arXivpreprintarXiv:2009.11462,2020. [90]J.Devlin,M.-W.Chang,K.Lee,andK.Toutanova,“Bert:Pre-trainingofdeepbidirectionaltransformersforlanguageunderstanding,” arXiv preprintarXiv:1810.04805,2018. [91]H.W.Chung,L.Hou,S.Longpre,B.Zoph,Y.Tay,W.Fedus,Y.Li,X.Wang,M.Dehghani,S.Brahma etal.,“Scalinginstruction-finetuned languagemodels,”arXivpreprintarXiv:2210.11416,2022. [92]M.Lewis,Y.Liu,N.Goyal,M.Ghazvininejad,A.Mohamed,O.Levy,V.Stoyanov,andL.Zettlemoyer,“Bart:Denoisingsequence-to- sequencepre-trainingfornaturallanguagegeneration,translation,andcomprehension,”arXivpreprintarXiv:1910.13461,2019. [93]L.Ouyang,J.Wu,X.Jiang,D.Almeida,C.L.Wainwright,P.Mishkin,C.Zhang,S.Agarwal,K.Slama,A.Ray etal.,“Traininglanguage modelstofollowinstructionswithhumanfeedback,”arXivpreprintarXiv:2203.02155,2022. [94]S.Li,Y.Zhao,R.Varma,O.Salpekar,P.Noordhuis,T.Li,A.Paszke,J.Smith,B.Vaughan,P.Damania etal.,“Pytorchdistributed: Experiencesonacceleratingdataparalleltraining,”arXivpreprintarXiv:2006.15704,2020. [95]M.Isard,M.Budiu,Y.Yu,A.Birrell,andD.Fetterly,“Dryad:distributeddata-parallelprogramsfromsequentialbuildingblocks,”in Proceedingsofthe2ndACMSIGOPS/EuroSysEuropeanConferenceonComputerSystems2007,2007,pp.59–72. [96]M.Shoeybi,M.Patwary,R.Puri,P.LeGresley,J.Casper,andB.Catanzaro,“Megatron-lm:Trainingmulti-billionparameterlanguagemodels usingmodelparallelism,”arXivpreprintarXiv:1909.08053,2019. [97]S.Rajbhandari,J.Rasley,O.Ruwase,andY.He,“Zero:Memoryoptimizationstowardtrainingtrillionparametermodels,”in SC20: InternationalConferenceforHighPerformanceComputing,Networking,StorageandAnalysis. IEEE,2020,pp.1–16. [98]Y.Huang,Y.Cheng,A.Bapna,O.Firat,D.Chen,M.Chen,H.Lee,J.Ngiam,Q.V.Le,Y.Wu etal.,“Gpipe:Efficienttrainingofgiant neuralnetworksusingpipelineparallelism,”Advancesinneuralinformationprocessingsystems,vol.32,2019. [99]P.Micikevicius,S.Narang,J.Alben,G.Diamos,E.Elsen,D.Garcia,B.Ginsburg,M.Houston,O.Kuchaiev,G.Venkatesh etal.,“Mixed precisiontraining,”arXivpreprintarXiv:1710.03740,2017. [100]J.W.Rae,S.Borgeaud,T.Cai,K.Millican,J.Hoffmann,F.Song,J.Aslanides,S.Henderson,R.Ring,S.Young etal.,“Scalinglanguage models:Methods,analysis&insightsfromtraininggopher,”arXivpreprintarXiv:2112.11446,2021. [101]J.Ren,S.Rajbhandari,R.Y.Aminabadi,O.Ruwase,S.Yang,M.Zhang,D.Li,andY.He,“

ssingsystems,vol.32,2019. [99]P.Micikevicius,S.Narang,J.Alben,G.Diamos,E.Elsen,D.Garcia,B.Ginsburg,M.Houston,O.Kuchaiev,G.Venkatesh etal.,“Mixed precisiontraining,”arXivpreprintarXiv:1710.03740,2017. [100]J.W.Rae,S.Borgeaud,T.Cai,K.Millican,J.Hoffmann,F.Song,J.Aslanides,S.Henderson,R.Ring,S.Young etal.,“Scalinglanguage models:Methods,analysis&insightsfromtraininggopher,”arXivpreprintarXiv:2112.11446,2021. [101]J.Ren,S.Rajbhandari,R.Y.Aminabadi,O.Ruwase,S.Yang,M.Zhang,D.Li,andY.He,“ {ZeRO-Offload}:Democratizing{Billion-Scale} modeltraining,”in2021USENIXAnnualTechnicalConference(USENIXATC21),2021,pp.551–564. [102]Y.Wang,Y.Kordi,S.Mishra,A.Liu,N.A.Smith,D.Khashabi,andH.Hajishirzi,“Self-instruct:Aligninglanguagemodelwithself generatedinstructions,”arXivpreprintarXiv:2212.10560,2022. [103]Y.Wang,S.Mishra,P.Alipoormolabashi,Y.Kordi,A.Mirzaei,A.Arunkumar,A.Ashok,A.S.Dhanasekaran,A.Naik,D.Stap etal., “Super-naturalinstructions:Generalizationviadeclarativeinstructionson1600+nlptasks,”arXivpreprintarXiv:2204.07705,2022. [104]S.H.Bach,V.Sanh,Z.-X.Yong,A.Webson,C.Raffel,N.V.Nayak,A.Sharma,T.Kim,M.S.Bari,T.Fevry etal.,“Promptsource:An integrateddevelopmentenvironmentandrepositoryfornaturallanguageprompts,”arXivpreprintarXiv:2202.01279,2022. [105]S.Victor,W.Albert,R.Colin,B.Stephen,S.Lintang,A.Zaid,C.Antoine,S.Arnaud,R.Arun,D.Manan etal.,“Multitaskprompted trainingenableszero-shottaskgeneralization,”inInternationalConferenceonLearningRepresentations,2022. [106]R.Nakano,J.Hilton,S.Balaji,J.Wu,L.Ouyang,C.Kim,C.Hesse,S.Jain,V.Kosaraju,W.Saunders etal.,“Webgpt:Browser-assisted question-answeringwithhumanfeedback,”arXivpreprintarXiv:2112.09332,2021. [107]J.Wei,M.Bosma,V.Zhao,K.Guu,A.W.Yu,B.Lester,N.Du,A.M.Dai,andQ.V.Le,“Finetunedlanguagemodelsarezero-shotlearners,” inInternationalConferenceonLearningRepresentations. YihengLiuetal.:PreprintsubmittedtoElsevier Page26of30 AComprehensiveOverviewfromTrainingtoInference [108]T.Tang,J.Li,W.X.Zhao,andJ.-R.Wen,“Mvp:Multi-tasksupervisedpre-trainingfornaturallanguagegeneration,” arXivpreprint arXiv:2206.12131,2022. [109]Z.Kenton,T.Everitt,L.Weidinger,I.Gabriel,V.Mikulik,andG.Irving,“Alignmentoflanguageagents,” arXivpreprintarXiv:2103.14659, 2021. [110]A.Glaese,N.McAleese,M.Trębacz,J.Aslanides,V.Firoiu,T.Ewalds,M.Rauh,L.Weidinger,M.Chadwick,P.Thacker etal.,“Improving alignmentofdialogueagentsviatargetedhumanjudgements,”arXivpreprintarXiv:2209.14375,2022. [111]J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347,2017. [112]E.J.Hu,Y.Shen,P.Wallis,Z.Allen-Zhu,Y.Li,S.Wang,L.Wang,andW.Chen,“Lora:Low-rankadaptationoflargelanguagemodels,” arXivpreprintarXiv:2106.09685,2021. [113]X.L.LiandP.Liang,“Prefix-tuning:Optimizingcontinuouspromptsforgeneration,” arXivpreprintarXiv:2101.00190,2021. [114]X.Liu,K.Ji,Y.Fu,W.L.Tam,Z.Du,Z.Yang,andJ.Tang,“P-tuningv2:Prompttuningcanbecomparabletofine-tuninguniversallyacross scalesandtasks,”arXivpreprintarXiv:2110.07602,2021. [115]X.Liu,Y.Zheng,Z.Du,M.Ding,Y.Qian,Z.Yang,andJ.Tang,“Gptunderstands,too,” AIOpen,2023. [116]Q.Zhang,M.Chen,A.Bukharin,P.He,Y.Cheng,W.Chen,andT.Zhao,“Adaptivebudgetallocationforparameter-efficientfine-tuning,” arXivpreprintarXiv:2303.10512,2023. [117]T.Dettmers,A.Pagnoni,A.Holtzman,andL.Zettlemoyer,“Qlora:Efficientfinetuningofquantizedllms,” arXivpreprintarXiv:2305.14314, 2023. [118]A.Askell,Y.Bai,A.Chen,D.Drain,D.Ganguli,T.Henighan,A.Jones,N.Joseph,B.Mann,N.DasSarma etal.,“Agenerallanguage assistantasalaboratoryforalignment,”arXivpreprintarXiv:2112.00861,2021. [119]Y.Chang,X.Wang,J.Wang,Y.Wu,K.Zhu,H.Chen,L.Yang,X.Yi,C.Wang,Y.Wang etal.,“Asurveyonevaluationoflargelanguage models,”arXivpreprintarXiv:2307.03109,2023. [120]Z.Liu,H.Jiang,T.Zhong,Z.Wu,C.Ma,Y.Li,X.Yu,Y.Zhang,Y.Pan,P.Shu etal.,“Holist

ofquantizedllms,” arXivpreprintarXiv:2305.14314, 2023. [118]A.Askell,Y.Bai,A.Chen,D.Drain,D.Ganguli,T.Henighan,A.Jones,N.Joseph,B.Mann,N.DasSarma etal.,“Agenerallanguage assistantasalaboratoryforalignment,”arXivpreprintarXiv:2112.00861,2021. [119]Y.Chang,X.Wang,J.Wang,Y.Wu,K.Zhu,H.Chen,L.Yang,X.Yi,C.Wang,Y.Wang etal.,“Asurveyonevaluationoflargelanguage models,”arXivpreprintarXiv:2307.03109,2023. [120]Z.Liu,H.Jiang,T.Zhong,Z.Wu,C.Ma,Y.Li,X.Yu,Y.Zhang,Y.Pan,P.Shu etal.,“Holisticevaluationofgpt-4vforbiomedicalimaging,” arXivpreprintarXiv:2312.05256,2023. [121]J.Deng,W.Dong,R.Socher,L.-J.Li,K.Li,andL.Fei-Fei,“Imagenet:Alarge-scalehierarchicalimagedatabase,”in 2009IEEEconference oncomputervisionandpatternrecognition. Ieee,2009,pp.248–255. [122]A.Kuznetsova,H.Rom,N.Alldrin,J.Uijlings,I.Krasin,J.Pont-Tuset,S.Kamali,S.Popov,M.Malloci,A.Kolesnikov etal.,“Theopen imagesdatasetv4:Unifiedimageclassification,objectdetection,andvisualrelationshipdetectionatscale,”InternationalJournalofComputer Vision,vol.128,no.7,pp.1956–1981,2020. [123]A.Wang,A.Singh,J.Michael,F.Hill,O.Levy,andS.R.Bowman,“Glue:Amulti-taskbenchmarkandanalysisplatformfornaturallanguage understanding,”2018. [124]A.Wang,Y.Pruksachatkun,N.Nangia,A.Singh,J.Michael,F.Hill,O.Levy,andS.Bowman,“Superglue:Astickierbenchmarkfor general-purposelanguageunderstandingsystems,”Advancesinneuralinformationprocessingsystems,vol.32,2019. [125]D.Hendrycks,C.Burns,S.Basart,A.Zou,M.Mazeika,D.Song,andJ.Steinhardt,“Measuringmassivemultitasklanguageunderstanding,” arXivpreprintarXiv:2009.03300,2020. [126]H.Li,Y.Zhang,F.Koto,Y.Yang,H.Zhao,Y.Gong,N.Duan,andT.Baldwin,“Cmmlu:Measuringmassivemultitasklanguage understandinginchinese,”arXivpreprintarXiv:2306.09212,2023. [127]J.Hu,S.Ruder,A.Siddhant,G.Neubig,O.Firat,andM.Johnson,“Xtreme:Amassivelymultilingualmulti-taskbenchmarkforevaluating cross-lingualgeneralisation,”inInternationalConferenceonMachineLearning. PMLR,2020,pp.4411–4421. [128]S.Ruder,N.Constant,J.Botha,A.Siddhant,O.Firat,J.Fu,P.Liu,J.Hu,D.Garrette,G.Neubig etal.,“Xtreme-r:Towardsmorechallenging andnuancedmultilingualevaluation,”arXivpreprintarXiv:2104.07412,2021. [129]D.Hendrycks,C.Burns,S.Kadavath,A.Arora,S.Basart,E.Tang,D.Song,andJ.Steinhardt,“Measuringmathematicalproblemsolving withthemathdataset,”arXivpreprintarXiv:2103.03874,2021. [130]K.Cobbe,V.Kosaraju,M.Bavarian,M.Chen,H.Jun,L.Kaiser,M.Plappert,J.Tworek,J.Hilton,R.Nakano etal.,“Trainingverifiersto solvemathwordproblems,”arXivpreprintarXiv:2110.14168,2021. [131]M.Chen,J.Tworek,H.Jun,Q.Yuan,H.P.d.O.Pinto,J.Kaplan,H.Edwards,Y.Burda,N.Joseph,G.Brockman etal.,“Evaluatinglarge languagemodelstrainedoncode,”arXivpreprintarXiv:2107.03374,2021. [132]J.Austin,A.Odena,M.Nye,M.Bosma,H.Michalewski,D.Dohan,E.Jiang,C.Cai,M.Terry,Q.Le etal.,“Programsynthesiswithlarge languagemodels,”arXivpreprintarXiv:2108.07732,2021. [133]R.Zellers,A.Holtzman,Y.Bisk,A.Farhadi,andY.Choi,“Hellaswag:Canamachinereallyfinishyoursentence?”2019. [134]Y.Bisk,R.Zellers,J.Gao,Y.Choi etal.,“Piqa:Reasoningaboutphysicalcommonsenseinnaturallanguage,”inProceedingsoftheAAAI conferenceonartificialintelligence,vol.34,no.05,2020,pp.7432–7439. [135]C.Clark,K.Lee,M.-W.Chang,T.Kwiatkowski,M.Collins,andK.Toutanova,“Boolq:Exploringthesurprisingdifficultyofnaturalyes/no questions,”arXivpreprintarXiv:1905.10044,2019. [136]M.Sap,H.Rashkin,D.Chen,R.LeBras,andY.Choi,“Socialiqa:Commonsensereasoningaboutsocialinteractions,” arXivpreprint arXiv:1904.09728,2019. [137]K.Sakaguchi,R.L.Bras,C.Bhagavatula,andY.Choi,“Winogrande:Anadversarialwinogradschemachallengeatscale,” Communications oftheACM,vol.64,no.9,pp.99–106,2021. [138]P.Clark,I.Cowhey,O.Etzioni,T.Khot,A.Sabharwal,C.Schoenick,andO.Tafjord,“Thinkyouhavesolvedquestionanswering?tryarc, theai2reasoningchallenge,”arXivpreprintarXiv:1803.05457,2018. [139]T.Mihaylov,P.Clark,T.Khot,andA.Sabharwal,“Canasuitofarmorconductelectricity?anewdatasetforopenbookqu

arXivpreprint arXiv:1904.09728,2019. [137]K.Sakaguchi,R.L.Bras,C.Bhagavatula,andY.Choi,“Winogrande:Anadversarialwinogradschemachallengeatscale,” Communications oftheACM,vol.64,no.9,pp.99–106,2021. [138]P.Clark,I.Cowhey,O.Etzioni,T.Khot,A.Sabharwal,C.Schoenick,andO.Tafjord,“Thinkyouhavesolvedquestionanswering?tryarc, theai2reasoningchallenge,”arXivpreprintarXiv:1803.05457,2018. [139]T.Mihaylov,P.Clark,T.Khot,andA.Sabharwal,“Canasuitofarmorconductelectricity?anewdatasetforopenbookquestionanswering,” 2018. YihengLiuetal.:PreprintsubmittedtoElsevier Page27of30 AComprehensiveOverviewfromTrainingtoInference [140]D.Jin,E.Pan,N.Oufattole,W.-H.Weng,H.Fang,andP.Szolovits,“Whatdiseasedoesthispatienthave?alarge-scaleopendomainquestion answeringdatasetfrommedicalexams,”AppliedSciences,vol.11,no.14,p.6421,2021. [141]A.Pal,L.K.Umapathi,andM.Sankarasubbu,“Medmcqa:Alarge-scalemulti-subjectmulti-choicedatasetformedicaldomainquestion answering,”inConferenceonHealth,Inference,andLearning. PMLR,2022,pp.248–260. [142]E.M.Voorhees etal.,“Thetrec-8questionansweringtrackreport.”inTrec,vol.99,1999,pp.77–82. [143]P.Rajpurkar,J.Zhang,K.Lopyrev,andP.Liang,“Squad:100,000+questionsformachinecomprehensionoftext,” arXivpreprint arXiv:1606.05250,2016. [144]T.Kwiatkowski,J.Palomaki,O.Redfield,M.Collins,A.Parikh,C.Alberti,D.Epstein,I.Polosukhin,J.Devlin,K.Lee etal.,“Natural questions:abenchmarkforquestionansweringresearch,”TransactionsoftheAssociationforComputationalLinguistics,vol.7,pp.453–466, 2019. [145]E.Kamalloo,N.Dziri,C.L.Clarke,andD.Rafiei,“Evaluatingopen-domainquestionansweringintheeraoflargelanguagemodels,” arXiv preprintarXiv:2305.06984,2023. [146]E.Ferrara,“Shouldchatgptbebiased?challengesandrisksofbiasinlargelanguagemodels,” arXivpreprintarXiv:2304.03738,2023. [147]S.Gehman,S.Gururangan,M.Sap,Y.Choi,andN.A.Smith,“Realtoxicityprompts:Evaluatingneuraltoxicdegenerationinlanguage models,”arXivpreprintarXiv:2009.11462,2020. [148]J.Zhao,M.Fang,Z.Shi,Y.Li,L.Chen,andM.Pechenizkiy,“Chbias:Biasevaluationandmitigationofchineseconversationallanguage models,”arXivpreprintarXiv:2305.11262,2023. [149]M.Nasr,N.Carlini,J.Hayase,M.Jagielski,A.F.Cooper,D.Ippolito,C.A.Choquette-Choo,E.Wallace,F.Tramèr,andK.Lee,“Scalable extractionoftrainingdatafrom(production)languagemodels,”arXivpreprintarXiv:2311.17035,2023. [150]X.Wu,J.Li,M.Xu,W.Dong,S.Wu,C.Bian,andD.Xiong,“Depn:Detectingandeditingprivacyneuronsinpretrainedlanguagemodels,” arXivpreprintarXiv:2310.20138,2023. [151]A.Zou,Z.Wang,J.Z.Kolter,andM.Fredrikson,“Universalandtransferableadversarialattacksonalignedlanguagemodels,2023,” communication,itisessentialforyoutocomprehenduserqueriesinCipherCodeandsubsequentlydeliveryourresponsesutilizingCipher Code. [152]Z.Zhang,Y.Li,J.Wang,B.Liu,D.Li,Y.Guo,X.Chen,andY.Liu,“Remos:reducingdefectinheritanceintransferlearningviarelevant modelslicing,”inProceedingsofthe44thInternationalConferenceonSoftwareEngineering,2022,pp.1856–1868. [153]K.Papineni,S.Roukos,T.Ward,andW.-J.Zhu,“Bleu:amethodforautomaticevaluationofmachinetranslation,”in Proceedingsofthe 40thannualmeetingoftheAssociationforComputationalLinguistics,2002,pp.311–318. [154]C.-Y.Lin,“Rouge:Apackageforautomaticevaluationofsummaries,”in Textsummarizationbranchesout,2004,pp.74–81. [155]T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, and Y. Artzi, “Bertscore: Evaluating text generation with bert,” arXiv preprint arXiv:1904.09675,2019. [156]J.Novikova,O.Dušek,A.C.Curry,andV.Rieser,“Whyweneednewevaluationmetricsfornlg,” arXivpreprintarXiv:1707.06875,2017. [157]T.Wolf,L.Debut,V.Sanh,J.Chaumond,C.Delangue,A.Moi,P.Cistac,T.Rault,R.Louf,M.Funtowicz etal.,“Transformers:State-of- the-artnaturallanguageprocessing,”inProceedingsofthe2020conferenceonempiricalmethodsinnaturallanguageprocessing:system demonstrations,2020,pp.38–45. [158]J.Rasley,S.Rajbhandari,O.Ruwase,

on with bert,” arXiv preprint arXiv:1904.09675,2019. [156]J.Novikova,O.Dušek,A.C.Curry,andV.Rieser,“Whyweneednewevaluationmetricsfornlg,” arXivpreprintarXiv:1707.06875,2017. [157]T.Wolf,L.Debut,V.Sanh,J.Chaumond,C.Delangue,A.Moi,P.Cistac,T.Rault,R.Louf,M.Funtowicz etal.,“Transformers:State-of- the-artnaturallanguageprocessing,”inProceedingsofthe2020conferenceonempiricalmethodsinnaturallanguageprocessing:system demonstrations,2020,pp.38–45. [158]J.Rasley,S.Rajbhandari,O.Ruwase,andY.He,“Deepspeed:Systemoptimizationsenabletrainingdeeplearningmodelswithover100 billionparameters,”inProceedingsofthe26thACMSIGKDDInternationalConferenceonKnowledgeDiscovery&DataMining,2020,pp. 3505–3506. [159]S.Rajbhandari,O.Ruwase,J.Rasley,S.Smith,andY.He,“Zero-infinity:Breakingthegpumemorywallforextremescaledeeplearning,” inProceedingsoftheInternationalConferenceforHighPerformanceComputing,Networking,StorageandAnalysis,2021,pp.1–14. [160]G.Zeng,X.Han,Z.Zhang,Z.Liu,Y.Lin,andM.Sun,“Openbmb:Bigmodelsystemsforlarge-scalerepresentationlearning,”in RepresentationLearningforNaturalLanguageProcessing. SpringerNatureSingaporeSingapore,2023,pp.463–489. [161]D.Narayanan,M.Shoeybi,J.Casper,P.LeGresley,M.Patwary,V.Korthikanti,D.Vainbrand,P.Kashinkunti,J.Bernauer,B.Catanzaro etal.,“Efficientlarge-scalelanguagemodeltrainingongpuclustersusingmegatron-lm,”inProceedingsoftheInternationalConferencefor HighPerformanceComputing,Networking,StorageandAnalysis,2021,pp.1–15. [162]V.A.Korthikanti,J.Casper,S.Lym,L.McAfee,M.Andersch,M.Shoeybi,andB.Catanzaro,“Reducingactivationrecomputationinlarge transformermodels,”ProceedingsofMachineLearningandSystems,vol.5,2023. [163]S.Li,H.Liu,Z.Bian,J.Fang,H.Huang,Y.Liu,B.Wang,andY.You,“Colossal-ai:Aunifieddeeplearningsystemforlarge-scaleparallel training,”inProceedingsofthe52ndInternationalConferenceonParallelProcessing,2023,pp.766–775. [164]J.He,J.Qiu,A.Zeng,Z.Yang,J.Zhai,andJ.Tang,“Fastmoe:Afastmixture-of-experttrainingsystem,” arXivpreprintarXiv:2103.13262, 2021. [165]J.He,J.Zhai,T.Antunes,H.Wang,F.Luo,S.Shi,andQ.Li,“Fastermoe:modelingandoptimizingtrainingoflarge-scaledynamicpre-trained models,”inProceedingsofthe27thACMSIGPLANSymposiumonPrinciplesandPracticeofParallelProgramming,2022,pp.120–134. [166]A.Paszke,S.Gross,F.Massa,A.Lerer,J.Bradbury,G.Chanan,T.Killeen,Z.Lin,N.Gimelshein,L.Antiga etal.,“Pytorch:Animperative style,high-performancedeeplearninglibrary,”Advancesinneuralinformationprocessingsystems,vol.32,2019. [167]M.Abadi,P.Barham,J.Chen,Z.Chen,A.Davis,J.Dean,M.Devin,S.Ghemawat,G.Irving,M.Isard etal.,“{TensorFlow}:asystem for {Large-Scale} machinelearning,”in12thUSENIXsymposiumonoperatingsystemsdesignandimplementation(OSDI16),2016,pp. 265–283. [168]M.Abadi,A.Agarwal,P.Barham,E.Brevdo,Z.Chen,C.Citro,G.S.Corrado,A.Davis,J.Dean,M.Devin etal.,“Tensorflow:Large-scale machinelearningonheterogeneousdistributedsystems,”arXivpreprintarXiv:1603.04467,2016. [169]Y.Ma,D.Yu,T.Wu,andH.Wang,“Paddlepaddle:Anopen-sourcedeeplearningplatformfromindustrialpractice,” FrontiersofDataand Domputing,vol.1,no.1,pp.105–115,2019. YihengLiuetal.:PreprintsubmittedtoElsevier Page28of30 AComprehensiveOverviewfromTrainingtoInference [170]T.Chen,M.Li,Y.Li,M.Lin,N.Wang,M.Wang,T.Xiao,B.Xu,C.Zhang,andZ.Zhang,“Mxnet:Aflexibleandefficientmachinelearning libraryforheterogeneousdistributedsystems,”arXivpreprintarXiv:1512.01274,2015. [171]J.Yuan,X.Li,C.Cheng,J.Liu,R.Guo,S.Cai,C.Yao,F.Yang,X.Yi,C.Wu etal.,“Oneflow:Redesignthedistributeddeeplearning frameworkfromscratch,”arXivpreprintarXiv:2110.15032,2021. [172]L.HuaweiTechnologiesCo.,“Huaweimindsporeaidevelopmentframework,”in ArtificialIntelligenceTechnology. Springer,2022,pp. 137–162. [173]J.Bradbury,R.Frostig,P.Hawkins,M.J.Johnson,C.Leary,D.Maclaurin,G.Necula,A.Paszke,J.VanderPlas,S.Wanderman-Milne etal., “Jax:composabletransformationsofp

1512.01274,2015. [171]J.Yuan,X.Li,C.Cheng,J.Liu,R.Guo,S.Cai,C.Yao,F.Yang,X.Yi,C.Wu etal.,“Oneflow:Redesignthedistributeddeeplearning frameworkfromscratch,”arXivpreprintarXiv:2110.15032,2021. [172]L.HuaweiTechnologiesCo.,“Huaweimindsporeaidevelopmentframework,”in ArtificialIntelligenceTechnology. Springer,2022,pp. 137–162. [173]J.Bradbury,R.Frostig,P.Hawkins,M.J.Johnson,C.Leary,D.Maclaurin,G.Necula,A.Paszke,J.VanderPlas,S.Wanderman-Milne etal., “Jax:composabletransformationsofpython+numpyprograms,”2018. [174]E.Strubell,A.Ganesh,andA.McCallum,“Energyandpolicyconsiderationsfordeeplearninginnlp,” arXivpreprintarXiv:1906.02243, 2019. [175]G.Hinton,O.Vinyals,andJ.Dean,“Distillingtheknowledgeinaneuralnetwork,” arXivpreprintarXiv:1503.02531,2015. [176]J.Gou,B.Yu,S.J.Maybank,andD.Tao,“Knowledgedistillation:Asurvey,” InternationalJournalofComputerVision,vol.129,pp. 1789–1819,2021. [177]S.Sun,Y.Cheng,Z.Gan,andJ.Liu,“Patientknowledgedistillationforbertmodelcompression,” arXivpreprintarXiv:1908.09355,2019. [178]X.Jiao,Y.Yin,L.Shang,X.Jiang,X.Chen,L.Li,F.Wang,andQ.Liu,“Tinybert:Distillingbertfornaturallanguageunderstanding,” arXiv preprintarXiv:1909.10351,2019. [179]M.A.Gordon,K.Duh,andN.Andrews,“Compressingbert:Studyingtheeffectsofweightpruningontransferlearning,” arXivpreprint arXiv:2002.08307,2020. [180]P.Michel,O.Levy,andG.Neubig,“Aresixteenheadsreallybetterthanone?” Advancesinneuralinformationprocessingsystems,vol.32, 2019. [181]H.Bai,W.Zhang,L.Hou,L.Shang,J.Jin,X.Jiang,Q.Liu,M.Lyu,andI.King,“Binarybert:Pushingthelimitofbertquantization,” arXiv preprintarXiv:2012.15701,2020. [182]Z.Lan,M.Chen,S.Goodman,K.Gimpel,P.Sharma,andR.Soricut,“Albert:Alitebertforself-supervisedlearningoflanguage representations,”arXivpreprintarXiv:1909.11942,2019. [183]P.Chen,H.-F.Yu,I.Dhillon,andC.-J.Hsieh,“Drone:Data-awarelow-rankcompressionforlargenlpmodels,” Advancesinneural informationprocessingsystems,vol.34,pp.29321–29334,2021. [184]X.Han,G.Zeng,W.Zhao,Z.Liu,Z.Zhang,J.Zhou,J.Zhang,J.Chao,andM.Sun,“Bminf:Anefficienttoolkitforbigmodelinference andtuning,”inProceedingsofthe60thAnnualMeetingoftheAssociationforComputationalLinguistics:SystemDemonstrations,2022,pp. 224–230. [185]Y.Zhao,A.Gu,R.Varma,L.Luo,C.-C.Huang,M.Xu,L.Wright,H.Shojanazeri,M.Ott,S.Shleifer etal.,“Pytorchfsdp:experienceson scalingfullyshardeddataparallel,”arXivpreprintarXiv:2304.11277,2023. [186]T.Dao,D.Fu,S.Ermon,A.Rudra,andC.Ré,“Flashattention:Fastandmemory-efficientexactattentionwithio-awareness,” Advancesin NeuralInformationProcessingSystems,vol.35,pp.16344–16359,2022. [187]W.Kwon,Z.Li,S.Zhuang,Y.Sheng,L.Zheng,C.H.Yu,J.Gonzalez,H.Zhang,andI.Stoica,“Efficientmemorymanagementforlarge languagemodelservingwithpagedattention,”inProceedingsofthe29thSymposiumonOperatingSystemsPrinciples,2023,pp.611–626. [188]Y.Sheng,L.Zheng,B.Yuan,Z.Li,M.Ryabinin,B.Chen,P.Liang,C.Ré,I.Stoica,andC.Zhang,“Flexgen:High-throughputgenerative inferenceoflargelanguagemodelswithasinglegpu,”inInternationalConferenceonMachineLearning. PMLR,2023,pp.31094–31116. [189]X.Miao,G.Oliaro,Z.Zhang,X.Cheng,Z.Wang,R.Y.Y.Wong,Z.Chen,D.Arfeen,R.Abhyankar,andZ.Jia,“Specinfer:Accelerating generativellmservingwithspeculativeinferenceandtokentreeverification,”arXivpreprintarXiv:2305.09781,2023. [190]G. Xiao, Y. Tian, B. Chen, S. Han, and M. Lewis, “Efficient streaming language models with attention sinks,” arXiv preprint arXiv:2309.17453,2023. [191]Z.Zhang,B.Gong,Y.Chen,X.Han,G.Zeng,W.Zhao,Y.Chen,Z.Liu,andM.Sun,“Bmcook:Atask-agnosticcompressiontoolkitfor bigmodels,”inProceedingsoftheThe2022ConferenceonEmpiricalMethodsinNaturalLanguageProcessing:SystemDemonstrations, 2022,pp.396–405. [192]A.Borzunov,D.Baranchuk,T.Dettmers,M.Ryabinin,Y.Belkada,A.Chumachenko,P.Samygin,andC.Raffel,“Petals:Collaborative inferenceandfine-tuningoflargemodels,”arXivpreprintarXiv:2209.01188,2022. [193]F.Dou,J.Ye,G.Yuan,Q.Lu,W.Niu,H.Sun,L.Guan,G.Lu,G.Mai,N.Liu

ng,B.Gong,Y.Chen,X.Han,G.Zeng,W.Zhao,Y.Chen,Z.Liu,andM.Sun,“Bmcook:Atask-agnosticcompressiontoolkitfor bigmodels,”inProceedingsoftheThe2022ConferenceonEmpiricalMethodsinNaturalLanguageProcessing:SystemDemonstrations, 2022,pp.396–405. [192]A.Borzunov,D.Baranchuk,T.Dettmers,M.Ryabinin,Y.Belkada,A.Chumachenko,P.Samygin,andC.Raffel,“Petals:Collaborative inferenceandfine-tuningoflargemodels,”arXivpreprintarXiv:2209.01188,2022. [193]F.Dou,J.Ye,G.Yuan,Q.Lu,W.Niu,H.Sun,L.Guan,G.Lu,G.Mai,N.Liu etal.,“Towardsartificialgeneralintelligence(agi)inthe internetofthings(iot):Opportunitiesandchallenges,”arXivpreprintarXiv:2309.07438,2023. [194]C.Liu,Z.Liu,J.Holmes,L.Zhang,L.Zhang,Y.Ding,P.Shu,Z.Wu,H.Dai,Y.Li etal.,“Artificialgeneralintelligenceforradiation oncology,”Meta-Radiology,p.100045,2023. [195]Z.Liu,Y.Li,Q.Cao,J.Chen,T.Yang,Z.Wu,J.Hale,J.Gibbs,K.Rasheed,N.Liu etal.,“Transformationvstradition:Artificialgeneral intelligence(agi)forartsandhumanities,”arXivpreprintarXiv:2310.19626,2023. [196]J.Wei,X.Wang,D.Schuurmans,M.Bosma,F.Xia,E.Chi,Q.V.Le,D.Zhou etal.,“Chain-of-thoughtpromptingelicitsreasoninginlarge languagemodels,”AdvancesinNeuralInformationProcessingSystems,vol.35,pp.24824–24837,2022. [197]T.Kojima,S.S.Gu,M.Reid,Y.Matsuo,andY.Iwasawa,“Largelanguagemodelsarezero-shotreasoners,” Advancesinneuralinformation processingsystems,vol.35,pp.22199–22213,2022. [198]N.Qiang,J.Gao,Q.Dong,H.Yue,H.Liang,L.Liu,J.Yu,J.Hu,S.Zhang,B.Ge etal.,“Functionalbrainnetworkidentificationandfmri augmentationusingavae-ganframework,”ComputersinBiologyandMedicine,vol.165,p.107395,2023. [199]M.He,X.Hou,E.Ge,Z.Wang,Z.Kang,N.Qiang,X.Zhang,andB.Ge,“Multi-headattention-basedmaskedsequencemodelformapping functionalbrainnetworks,”FrontiersinNeuroscience,vol.17,p.1183145,2023. [200]Y.Liu,E.Ge,N.Qiang,T.Liu,andB.Ge,“Spatial-temporalconvolutionalattentionformappingfunctionalbrainnetworks,”in 2023IEEE 20thInternationalSymposiumonBiomedicalImaging(ISBI). IEEE,2023,pp.1–4. YihengLiuetal.:PreprintsubmittedtoElsevier Page29of30 AComprehensiveOverviewfromTrainingtoInference [201]S.R.Oota,J.Arora,V.Agarwal,M.Marreddy,M.Gupta,andB.Surampudi,“Neurallanguagetaskonomy:WhichNLPtasksarethe mostpredictiveoffMRIbrainactivity?”inProceedingsofthe2022ConferenceoftheNorthAmericanChapteroftheAssociationfor ComputationalLinguistics:HumanLanguageTechnologies,M.Carpuat,M.-C.deMarneffe,andI.V.MezaRuiz,Eds. Seattle,United States:AssociationforComputationalLinguistics,Jul.2022,pp.3220–3237. [202]Z.Liu,Y.Li,P.Shu,A.Zhong,L.Yang,C.Ju,Z.Wu,C.Ma,J.Luo,C.Chen etal.,“Radiology-llama2:Best-in-classlargelanguagemodel forradiology,”arXivpreprintarXiv:2309.06419,2023. [203]T.Sun,X.Zhang,Z.He,P.Li,Q.Cheng,H.Yan,X.Liu,Y.Shao,Q.Tang,X.Zhao,K.Chen,Y.Zheng,Z.Zhou,R.Li,J.Zhan,Y.Zhou, L.Li,X.Yang,L.Wu,Z.Yin,X.Huang,andX.Qiu,“Moss:Trainingconversationallanguagemodelsfromsyntheticdata,”2023. [204]L.X.XuanweiZhangandK.Zhao,“Chatyuan:Alargelanguagemodelfordialogueinchineseandenglish,”Dec.2022.[Online].Available: https://github.com/clue-ai/ChatYuan [205]Baichuan,“Baichuan2:Openlarge-scalelanguagemodels,” arXivpreprintarXiv:2309.10305,2023. [206]D.Zhu,J.Chen,X.Shen,X.Li,andM.Elhoseiny,“Minigpt-4:Enhancingvision-languageunderstandingwithadvancedlargelanguage models,”arXivpreprintarXiv:2304.10592,2023. [207]L.Zheng,W.-L.Chiang,Y.Sheng,S.Zhuang,Z.Wu,Y.Zhuang,Z.Lin,Z.Li,D.Li,E.Xing etal.,“Judgingllm-as-a-judgewithmt-bench andchatbotarena,”arXivpreprintarXiv:2306.05685,2023. [208]B.Peng,E.Alcaide,Q.Anthony,A.Albalak,S.Arcadinho,H.Cao,X.Cheng,M.Chung,M.Grella,K.K.GV etal.,“Rwkv:Reinventing rnnsforthetransformerera,”arXivpreprintarXiv:2305.13048,2023. YihengLiuetal.:PreprintsubmittedtoElsevier Page30of30

ing etal.,“Judgingllm-as-a-judgewithmt-bench andchatbotarena,”arXivpreprintarXiv:2306.05685,2023. [208]B.Peng,E.Alcaide,Q.Anthony,A.Albalak,S.Arcadinho,H.Cao,X.Cheng,M.Chung,M.Grella,K.K.GV etal.,“Rwkv:Reinventing rnnsforthetransformerera,”arXivpreprintarXiv:2305.13048,2023. YihengLiuetal.:PreprintsubmittedtoElsevier Page30of30