You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Large Language Models (LLMs) significantly benefit from Chain-of-Thought(CoT) prompting in performing various reasoning tasks. While CoT allows modelsto produce more comprehensive reasoning processes, its emphasis on intermediatereasoning steps can inadvertently introduce hallucinations and accumulatederrors, thereby limiting models' ability to solve complex reasoning tasks.Inspired by how humans engage in careful and meticulous deductive logicalreasoning processes to solve tasks, we seek to enable language models toperform explicit and rigorous deductive reasoning, and also ensure thetrustworthiness of their reasoning process through self-verification. However,directly verifying the validity of an entire deductive reasoning process ischallenging, even with advanced models like ChatGPT. In light of this, wepropose to decompose a reasoning verification process into a series ofstep-by-step subprocesses, each only receiving their necessary context andpremises. To facilitate this procedure, we propose Natural Program, a naturallanguage-based deductive reasoning format. Our approach enables models togenerate precise reasoning steps where subsequent steps are more rigorouslygrounded on prior steps. It also empowers language models to carry outreasoning self-verification in a step-by-step manner. By integrating thisverification process into each deductive reasoning stage, we significantlyenhance the rigor and trustfulness of generated reasoning steps. Along thisprocess, we also improve the answer correctness on complex reasoning tasks.Code will be released at https://github.com/lz1oceani/verify_cot.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
人間がタスクを解決するために注意深く緻密な演繹的論理推論プロセスに従事する方法に着想を得て、言語モデルが明示的で厳密な演繹的推論を実行し、自己検証を通じて推論プロセスの信頼性を確保できるようにすることを目指しています。しかし、ChatGPTなどの高度なモデルでも、演繹的推論プロセス全体の妥当性を直接検証することは困難です。このため、推論検証プロセスを一連のステップごとのサブプロセスに分解し、必要な文脈と前提条件のみを受け取るようにすることを提案します。この手順を容易にするために、自然言語ベースの演繹的推論形式であるNatural Programを提案します。このアプローチにより、モデルは、後続のステップが前のステップにより厳密に基礎づけられるように、正確な推論ステップを生成できるようになります。また、言語モデルは、ステップごとに推論自己検証を実行できるようになります。この検証プロセスを各演繹的推論段階に統合することで、生成された推論ステップの厳密性と信頼性を大幅に向上させます。このプロセスに沿って、複雑な推論タスクの回答の正確性も向上させます。
コードはhttps://github.com/lz1oceani/verify_cotで公開されます。
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: