Chain-of-Thought
Chain-of-Thought (CoT) is a prompting technique in large language model inference that guides models to generate answers through stepwise intermediate reasoning by incorporating sequential derivation steps into the input prompt. This technique significantly improves model performance on complex tasks such as mathematical reasoning, logical judgment, and multi-step problem solving, offering greater interpretability and accuracy compared to standard prompting approaches that elicit direct answers.
Core Facts
Reasoning models perform better in information synthesis tasks requiring multi-step logical deduction, but have higher latency and should be selected appropriately based on task characteristics.
Chain-of-thought reasoning was proposed by Wei et al. in 2022, referring to guiding a model to output step-by-step intermediate reasoning processes before giving a final answer
Reasoning models, represented by OpenAI o1, DeepSeek-R1, and QwQ, have chain-of-thought expansion as their core mechanism, generating a large number of intermediate reasoning steps before producing a final answer.
Timeline (last 90 days)
作者在五个推理基准上验证了 RCD，结果显示相比标准 CoT，RCD 提升了答案准确率。
In CoT reasoning, models often continue generating after the answer has become reachable, extra steps drive up serving cost, and subsequent steps may overwrite originally correct intermediate results, lowering accuracy.
Long chain-of-thought reasoning (such as Chain-of-Thought) excels at complex problems, but incurs high token overhead and high latency.
CoT prompting, without changing the model architecture, acts as an implicit working memory buffer by explicitly writing the intermediate reasoning process into the output sequence, thereby improving reasoning depth; but its limitation is that once the reasoning chain is generated it cannot be modified, and errors propagate and amplify along the chain.
A single forward pass is relative to multiple sampling (such as Self-Consistency) and chain-of-thought reasoning, eliminating the sampling and aggregation steps, reducing inference latency and compute overhead, and making the output deterministic.
The effectiveness of CoT stems from the externalized intermediate steps providing explicit context for subsequent tokens and allowing each step to be sampled, checked, and self-corrected.
The cost of CoT is that reasoning depth is strictly bound to the length of the generated sequence, and every additional token written incurs the cost and latency of one forward computation.
Traditional chain-of-thought prompting has three major pain points: poor verifiability, high generation overhead, and questionable faithfulness.
OpenAI's new method pushes CoT down from the prompt layer to the training and runtime mechanism layer: the model internally generates multiple candidate reasoning paths, filters them through self-consistency checks or a process reward model, and then outputs the final answer.
Academic research shows that adjusting the order of question phrasing, adding role settings, or asking the model to think step by step (chain-of-thought prompting) can cause accuracy on the same task to fluctuate by tens of percentage points.
40 more timeline events
All Facts (20)
Reasoning models perform better in information synthesis tasks requiring multi-step logical deduction, but have higher latency and should be selected appropriately based on task characteristics.
90%VerifiedChain-of-thought reasoning was proposed by Wei et al. in 2022, referring to guiding a model to output step-by-step intermediate reasoning processes before giving a final answer
90%VerifiedReasoning models, represented by OpenAI o1, DeepSeek-R1, and QwQ, have chain-of-thought expansion as their core mechanism, generating a large number of intermediate reasoning steps before producing a final answer.
90%VerifiedThe Chain-of-Thought prompting paper published in 2022 by Jason Wei et al. of Google Brain was the first to systematically demonstrate that adding step-by-step reasoning examples significantly improves the accuracy of large models such as PaLM on mathematical reasoning tasks
90%VerifiedHigh-intensity reasoning configurations typically entail higher costs and longer response times, trading response speed for logical rigor.
90%VerifiedChain-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%VerifiedO3 belongs to OpenAI's o-series reasoning models and introduces a Chain-of-Thought mechanism for multi-step internal reasoning.
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%VerifiedDeepSeek's deep thinking mode is essentially an enhanced implementation of the Chain-of-Thought mechanism.
90%VerifiedThe Chain-of-Thought prompting technique was formally proposed by the Google Brain team in the 2022 paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models".
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%VerifiedThe role of testers is shifting from executors to AI reviewers, and the prerequisite for harnessing AI is possessing foundational skills such as automated testing and performance testing.
80%VerifiedThe core idea of ReAct is to interweave the language model's thinking and actions, forming an iterative loop of Think → Act → Observe, which differs from the pure reasoning chain of Chain-of-Thought.
80%VerifiedReAct fuses the Chain-of-Thought and direct tool-calling research lines, forming an iterative loop of "Think → Act → Observe."
80%VerifiedDeepSeek-R1 is a model focused on reasoning capabilities and possesses chain-of-thought ability through reinforcement learning training.
80%VerifiedOpenAI's o1 series represents a new paradigm called "test-time compute"
80%VerifiedResearch by Google in 2022 showed that adding 'Let's think step by step' to a prompt can significantly improve model performance on math and logical reasoning tasks.
80%VerifiedChain-of-Thought technology was first systematically proposed by Jason Wei and others from the Google Brain team in a 2022 paper.
80%VerifiedThe ReAct (Reasoning + Acting) framework was proposed by Yao et al. in 2022, alternating Chain-of-Thought reasoning with external tool interactions.
80%