{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/large-language-models-can-self-improve","title":"Large Language Models Can Self-Improve","arxiv_id":"2210.11610","date":"2022-10-20","proceeding":null,"authors":["Jiaxin Huang","Shixiang Shane Gu","Le Hou","Yuexin Wu","Xuezhi Wang","Hongkun Yu","Jiawei Han"],"abstract":"Large Language Models (LLMs) have achieved excellent performances in various tasks. However, fine-tuning an LLM requires extensive supervision. Human, on the other hand, may improve their reasoning abilities by self-thinking without external inputs. In this work, we demonstrate that an LLM is also capable of self-improving with only unlabeled datasets. We use a pre-trained LLM to generate \"high-confidence\" rationale-augmented answers for unlabeled questions using Chain-of-Thought prompting and self-consistency, and fine-tune the LLM using those self-generated solutions as target outputs. We show that our approach improves the general reasoning ability of a 540B-parameter LLM (74.4%->82.1% on GSM8K, 78.2%->83.0% on DROP, 90.0%->94.4% on OpenBookQA, and 63.4%->67.9% on ANLI-A3) and achieves state-of-the-art-level performance, without any ground truth label. We conduct ablation studies and show that fine-tuning on reasoning is critical for self-improvement.","url_abs":"https://arxiv.org/abs/2210.11610v2","url_pdf":"https://arxiv.org/pdf/2210.11610v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[],"tasks":[{"task_slug":"arithmetic-reasoning","task_name":"Arithmetic Reasoning"},{"task_slug":"common-sense-reasoning","task_name":"Common Sense Reasoning"},{"task_slug":"gsm8k","task_name":"GSM8K"},{"task_slug":"natural-language-inference","task_name":"Natural Language Inference"},{"task_slug":"question-answering","task_name":"Question Answering"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/arithmetic-reasoning-on-gsm8k","task":"Arithmetic Reasoning","dataset":"GSM8K","model":"PaLM 540B (Self Improvement, Self Consistency)","rank_in_archive_order":62,"of":164,"metrics":{"Accuracy":"82.1","Parameters (Billion)":"540"},"uses_additional_data":false},{"leaderboard":"/sota/arithmetic-reasoning-on-gsm8k","task":"Arithmetic Reasoning","dataset":"GSM8K","model":"PaLM 540B (Self Consistency)","rank_in_archive_order":88,"of":164,"metrics":{"Accuracy":"74.4","Parameters (Billion)":"540"},"uses_additional_data":false},{"leaderboard":"/sota/arithmetic-reasoning-on-gsm8k","task":"Arithmetic Reasoning","dataset":"GSM8K","model":"PaLM 540B (Self Improvement, CoT Prompting)","rank_in_archive_order":96,"of":164,"metrics":{"Accuracy":"73.5","Parameters (Billion)":"540"},"uses_additional_data":false},{"leaderboard":"/sota/arithmetic-reasoning-on-gsm8k","task":"Arithmetic Reasoning","dataset":"GSM8K","model":"PaLM 540B (CoT Prompting)","rank_in_archive_order":125,"of":164,"metrics":{"Accuracy":"56.5","Parameters (Billion)":"540"},"uses_additional_data":false},{"leaderboard":"/sota/arithmetic-reasoning-on-gsm8k","task":"Arithmetic Reasoning","dataset":"GSM8K","model":"PaLM 540B (Self Improvement, Standard-Prompting)","rank_in_archive_order":146,"of":164,"metrics":{"Accuracy":"32.2","Parameters (Billion)":"540"},"uses_additional_data":false},{"leaderboard":"/sota/arithmetic-reasoning-on-gsm8k","task":"Arithmetic Reasoning","dataset":"GSM8K","model":"PaLM 540B (Standard-Prompting)","rank_in_archive_order":154,"of":164,"metrics":{"Accuracy":"17.9","Parameters (Billion)":"540"},"uses_additional_data":false},{"leaderboard":"/sota/common-sense-reasoning-on-arc-challenge","task":"Common Sense Reasoning","dataset":"ARC (Challenge)","model":"PaLM 540B (Self Improvement, Self Consistency)","rank_in_archive_order":7,"of":54,"metrics":{"Accuracy":"89.8"},"uses_additional_data":false},{"leaderboard":"/sota/common-sense-reasoning-on-arc-challenge","task":"Common Sense Reasoning","dataset":"ARC (Challenge)","model":"PaLM 540B (Self Consistency)","rank_in_archive_order":8,"of":54,"metrics":{"Accuracy":"88.7"},"uses_additional_data":false},{"leaderboard":"/sota/common-sense-reasoning-on-arc-challenge","task":"Common Sense Reasoning","dataset":"ARC (Challenge)","model":"PaLM 540B (Self Improvement, CoT Prompting)","rank_in_archive_order":9,"of":54,"metrics":{"Accuracy":"88.3"},"uses_additional_data":false},{"leaderboard":"/sota/common-sense-reasoning-on-arc-challenge","task":"Common Sense Reasoning","dataset":"ARC (Challenge)","model":"PaLM 540B (Self Improvement, Standard-Prompting)","rank_in_archive_order":10,"of":54,"metrics":{"Accuracy":"87.2"},"uses_additional_data":false},{"leaderboard":"/sota/common-sense-reasoning-on-arc-challenge","task":"Common Sense Reasoning","dataset":"ARC (Challenge)","model":"PaLM 540B (Standard-Prompting)","rank_in_archive_order":11,"of":54,"metrics":{"Accuracy":"87.1"},"uses_additional_data":false},{"leaderboard":"/sota/common-sense-reasoning-on-arc-challenge","task":"Common Sense Reasoning","dataset":"ARC (Challenge)","model":"PaLM 540B (CoT Prompting)","rank_in_archive_order":15,"of":54,"metrics":{"Accuracy":"85.2"},"uses_additional_data":false},{"leaderboard":"/sota/natural-language-inference-on-anli-test","task":"Natural Language Inference","dataset":"ANLI test","model":"PaLM 540B (Self Improvement, Self Consistency)","rank_in_archive_order":20,"of":25,"metrics":{"A2":"66.5","A3":"67.9"},"uses_additional_data":false},{"leaderboard":"/sota/natural-language-inference-on-anli-test","task":"Natural Language Inference","dataset":"ANLI test","model":"PaLM 540B (Self Improvement, CoT Prompting)","rank_in_archive_order":21,"of":25,"metrics":{"A2":"65.3","A3":"67.3"},"uses_additional_data":false},{"leaderboard":"/sota/natural-language-inference-on-anli-test","task":"Natural Language Inference","dataset":"ANLI test","model":"PaLM 540B (Self Improvement, Standard-Prompting)","rank_in_archive_order":22,"of":25,"metrics":{"A2":"64.8","A3":"66.9"},"uses_additional_data":false},{"leaderboard":"/sota/natural-language-inference-on-anli-test","task":"Natural Language Inference","dataset":"ANLI test","model":"PaLM 540B (Self Consistency)","rank_in_archive_order":23,"of":25,"metrics":{"A2":"64.5","A3":"63.4"},"uses_additional_data":false},{"leaderboard":"/sota/natural-language-inference-on-anli-test","task":"Natural Language Inference","dataset":"ANLI test","model":"PaLM 540B (CoT Prompting)","rank_in_archive_order":24,"of":25,"metrics":{"A2":"58.9","A3":"60.6"},"uses_additional_data":false},{"leaderboard":"/sota/natural-language-inference-on-anli-test","task":"Natural Language Inference","dataset":"ANLI test","model":"PaLM 540B (Standard-Prompting)","rank_in_archive_order":25,"of":25,"metrics":{"A2":"55.8","A3":"55.8"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-drop","task":"Question Answering","dataset":"DROP","model":"PaLM 540B (Self Improvement, Self Consistency)","rank_in_archive_order":1,"of":6,"metrics":{"Accuracy":"83"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-drop","task":"Question Answering","dataset":"DROP","model":"PaLM 540B (Self Consistency)","rank_in_archive_order":2,"of":6,"metrics":{"Accuracy":"78.2"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-drop","task":"Question Answering","dataset":"DROP","model":"PaLM 540B (Self Improvement, CoT Prompting)","rank_in_archive_order":3,"of":6,"metrics":{"Accuracy":"76.2"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-drop","task":"Question Answering","dataset":"DROP","model":"PaLM 540B (Self Improvement, Standard-Prompting)","rank_in_archive_order":4,"of":6,"metrics":{"Accuracy":"71.7"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-drop","task":"Question Answering","dataset":"DROP","model":"PaLM 540B (CoT Prompting)","rank_in_archive_order":5,"of":6,"metrics":{"Accuracy":"70.6"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-drop","task":"Question Answering","dataset":"DROP","model":"PaLM 540B (Standard-Prompting)","rank_in_archive_order":6,"of":6,"metrics":{"Accuracy":"60"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-openbookqa","task":"Question Answering","dataset":"OpenBookQA","model":"PaLM 540B (Self Improvement, Self Consistency)","rank_in_archive_order":3,"of":45,"metrics":{"Accuracy":"94.4"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-openbookqa","task":"Question Answering","dataset":"OpenBookQA","model":"PaLM 540B (Self Improvement, CoT Prompting)","rank_in_archive_order":5,"of":45,"metrics":{"Accuracy":"93"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-openbookqa","task":"Question Answering","dataset":"OpenBookQA","model":"PaLM 540B (Self Improvement, Standard-Prompting)","rank_in_archive_order":6,"of":45,"metrics":{"Accuracy":"92"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-openbookqa","task":"Question Answering","dataset":"OpenBookQA","model":"PaLM 540B (Self Consistency)","rank_in_archive_order":8,"of":45,"metrics":{"Accuracy":"90"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-openbookqa","task":"Question Answering","dataset":"OpenBookQA","model":"PaLM 540B (CoT Prompting)","rank_in_archive_order":15,"of":45,"metrics":{"Accuracy":"86.4"},"uses_additional_data":false},{"leaderboard":"/sota/question-answering-on-openbookqa","task":"Question Answering","dataset":"OpenBookQA","model":"PaLM 540B (Standard-Prompting)","rank_in_archive_order":17,"of":45,"metrics":{"Accuracy":"84.4"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/2210.11610","atlas_url":"https://app.syntology.ai/?focus=2210.11610","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}