CoT
Chain of Thought (CoT) is a prompting method used to enhance the reasoning capabilities of large language models. Its core principle is to guide the model to produce intermediate reasoning steps before generating a final answer, decomposing complex problems into a sequence of ordered sub-steps. The method is widely applied to mathematical reasoning, logical inference, and multi-step question answering, and can significantly improve model accuracy and interpretability on complex tasks.
Core Facts
Chain-of-Thought (CoT) reasoning was proposed by the Google Brain team in 2022 and improves accuracy on complex tasks by guiding the model to expand its reasoning step by step.
Chain of Thought能显著提升模型在数学推理、逻辑判断等需要多步推导的任务上的表现
Agent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.
Timeline (last 90 days)
Adding the instruction "please calculate step by step" (chain of thought) to a request can raise the accuracy of a large model on complex arithmetic such as multiplying a six-digit number by a seven-digit number to above 90%.
The author ignored all CoT (chain-of-thought) models, considering them wasteful of tokens and slower.
In Agent development, CoT often serves as the underlying mechanism of the planning module, helping the model decompose complex goals into executable sub-steps, while ReAct is responsible for interacting with external tools and environments, and the two complement each other.
The core Agent architecture encompasses three core technologies: CoT/ReAct chain-of-thought, RAG retrieval-augmented generation, and model memory mechanisms.
Best practices for getting reliable scores from large models include using a scoring rubric, scoring multiple dimensions separately and then weighting them, calibrating with few-shot examples, requiring chain-of-thought analysis item by item, and optionally using multiple sampling with averaging or setting temperature=0.
All Facts (13)
Chain-of-Thought (CoT) reasoning was proposed by the Google Brain team in 2022 and improves accuracy on complex tasks by guiding the model to expand its reasoning step by step.
90%VerifiedChain of Thought能显著提升模型在数学推理、逻辑判断等需要多步推导的任务上的表现
90%VerifiedAgent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.
90%VerifiedThe core finding of CoT is that when models are asked to think step by step, accuracy on tasks such as arithmetic and commonsense reasoning improves substantially.
80%VerifiedChain-of-Thought (CoT), proposed by Wei et al., improves accuracy on complex mathematical and logical reasoning tasks by guiding the model to display its reasoning process step by step.
75%UnverifiedAdding the instruction "please calculate step by step" (chain of thought) to a request can raise the accuracy of a large model on complex arithmetic such as multiplying a six-digit number by a seven-digit number to above 90%.
50%UnverifiedThe author ignored all CoT (chain-of-thought) models, considering them wasteful of tokens and slower.
50%UnverifiedIn Agent development, CoT often serves as the underlying mechanism of the planning module, helping the model decompose complex goals into executable sub-steps, while ReAct is responsible for interacting with external tools and environments, and the two complement each other.
50%UnverifiedThe core Agent architecture encompasses three core technologies: CoT/ReAct chain-of-thought, RAG retrieval-augmented generation, and model memory mechanisms.
50%UnverifiedBest practices for getting reliable scores from large models include using a scoring rubric, scoring multiple dimensions separately and then weighting them, calibrating with few-shot examples, requiring chain-of-thought analysis item by item, and optionally using multiple sampling with averaging or setting temperature=0.
50%UnverifiedReAct交替进行推理与行动，CoT通过中间步骤提升单步推理质量，ToT通过树状搜索探索多条解题路径
50%UnverifiedReAct范式在CoT基础上引入行动维度，形成思考→行动→观察→再思考的闭环，显著减少大模型幻觉问题
50%UnverifiedCoT是静态推理链，而ReAct Agent是动态的、与环境交互的闭环系统
50%