我们为何思考

Why We Think

翁荔 Lilian Weng · Thinking Machines Lab · 2025-05-01 · Lil'Log ↗

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

日期：2025 年 5 月 1 日 | 预计阅读时间：40 分钟 | 作者：Lilian Weng * 模型是否忠实反映其思考 * 思维链上的优化压力：好还是坏？

Date: May 1, 2025 | Estimated Reading Time: 40 min | Author: Lilian Weng * Does the Model Tell What it Thinks Faithfully * Optimization Pressure on CoT: Good or Bad?

核心贡献 · Key contributions

• 测试时计算和思维链推理显著提升模型性能，尤其在复杂任务上。
  Test-time compute and chain-of-thought reasoning significantly improve model performance, especially on complex tasks.
• 在推理任务上使用强化学习（如 DeepSeek-R1），无需监督微调即可涌现反思和回溯能力。
  Reinforcement learning on reasoning tasks, as in DeepSeek-R1, enables emergent reflection and backtracking without supervised fine-tuning.
• 并行采样和顺序修正是利用测试时计算的两种主要方法，混合策略对难题最优。
  Parallel sampling and sequential revision are two main approaches for utilizing test-time compute, with hybrid strategies optimal for hard problems.
• 思维链提供了可解释性，但并非总是忠实的；推理模型比非推理模型表现出更忠实的思维链。
  Chain-of-thought provides interpretability but is not always faithful; reasoning models show more faithful CoT than non-reasoning ones.
• 对于简单和中等难度问题，优化测试时计算比扩展模型参数更有效。
  Optimizing test-time compute can be more effective than scaling model parameters for easy and medium questions.
• 潜变量模型和 EM 算法为理解和改进思维链推理提供了理论框架。
  Latent variable models and EM algorithm provide a theoretical framework for understanding and improving chain-of-thought reasoning.

局限 · Limitations

• 当基础模型能力差距较大时，测试时计算对难题效果有限。
  Test-time compute is less effective for hard problems when the base model capability gap is large.
• 在强化学习训练中直接优化思维链可能导致奖励黑客行为和推理模糊化。
  Direct optimization on chain-of-thought during RL training can lead to reward hacking and obfuscated reasoning.
• 无外部反馈的自我修正常常失败，导致幻觉或行为崩溃。
  Self-correction without external feedback often fails, causing hallucination or behavior collapse.
• 过程奖励模型难以定义每一步的评分标准，且易受奖励黑客攻击。
  Process reward models are hard to define per-step rubrics and vulnerable to reward hacking.
• 通过拒绝采样进行预算强制显示出反向缩放，即更长的思维链会降低性能。
  Budget forcing via rejection sampling shows reversed scaling, where longer CoTs degrade performance.

论文章节 · Sections（共 22）

• 概述 Overview
• 动机 Motivation#(https://lilianweng.github.io/posts/2025-05-01-thinking/#motivation)
• 与心理学的类比 Analogy to Psychology#(https://lilianweng.github.io/posts/2025-05-01-thinking/#analogy-to-psychology)
• 算力作为一种资源 Computation as a Resource#(https://lilianweng.github.io/posts/2025-05-01-thinking/#computation-as-a-resource)
• 潜在变量建模 Latent Variable Modeling#(https://lilianweng.github.io/posts/2025-05-01-thinking/#latent-variable-modeling)
• 用 Token 思考 Thinking in Tokens#(https://lilianweng.github.io/posts/2025-05-01-thinking/#thinking-in-tokens)
• 分支与编辑 Branching and Editing#(https://lilianweng.github.io/posts/2025-05-01-thinking/#branching-and-editing)
• 并行采样 Parallel Sampling#(https://lilianweng.github.io/posts/2025-05-01-thinking/#parallel-sampling)
• 顺序修订 Sequential Revision#(https://lilianweng.github.io/posts/2025-05-01-thinking/#sequential-revision)
• 用于更好推理的强化学习 RL for Better Reasoning#(https://lilianweng.github.io/posts/2025-05-01-thinking/#rl-for-better-reasoning)
• 外部工具使用 External Tool Use#(https://lilianweng.github.io/posts/2025-05-01-thinking/#external-tool-use)
• 忠实思考 Thinking Faithfully#(https://lilianweng.github.io/posts/2025-05-01-thinking/#thinking-faithfully)
• 模型是否忠实地表达其思考 Does the Model Tell What it Thinks Faithfully#(https://lilianweng.github.io/posts/2025-05-01-thinking/#does-the-model-tell-what-it-thinks-faithfully)
• 对思维链的优化压力：好还是坏？ Optimization Pressure on CoT: Good or Bad?#(https://lilianweng.github.io/posts/2025-05-01-thinking/#optimization-pressure-on-cot-good-or-bad)
• 连续空间中的思考 Thinking in Continuous Space#(https://lilianweng.github.io/posts/2025-05-01-thinking/#thinking-in-continuous-space)
• 循环架构 Recurrent Architecture#(https://lilianweng.github.io/posts/2025-05-01-thinking/#recurrent-architecture)
• 思考令牌 Thinking Tokens#(https://lilianweng.github.io/posts/2025-05-01-thinking/#thinking-tokens)
• 思维作为潜在变量 Thinking as Latent Variables#(https://lilianweng.github.io/posts/2025-05-01-thinking/#thinking-as-latent-variables)
• 期望最大化 Expectation-Maximization#(https://lilianweng.github.io/posts/2025-05-01-thinking/#expectation-maximization)
• 迭代学习 Iterative Learning#(https://lilianweng.github.io/posts/2025-05-01-thinking/#iterative-learning)
• 思考时间的缩放定律 Scaling Laws for Thinking Time#(https://lilianweng.github.io/posts/2025-05-01-thinking/#scaling-laws-for-thinking-time)
• 未来展望 What’s for Future#(https://lilianweng.github.io/posts/2025-05-01-thinking/#whats-for-future)

阅读逐段中英对照全文 →