思维链什么时候有用When Chain-of-Thought Helps
微型指南 · 约 6 分钟Guide · about 6 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
「一步步想」对某些任务提升明显，对另一些几乎没用甚至帮倒忙。这篇讲怎么判断值不值得让模型展开推理。"Think step by step" helps some tasks a lot and barely helps — or actively hurts — others. How to tell whether extended reasoning is worth it.
先分清任务类型：算术、多步逻辑、规划类任务收益大；简单抽取、翻译、格式转换基本无收益。
「一步步想」这类触发语对老一代模型才关键；推理模型已内置长思维链，再手动催只是浪费 token。
要求模型先列步骤再作答，比只加一句「想一想」更有效，也更可控。
思维链输出不是可靠的思考记录：它可能事后编出说得通的过程，别把它当审计日志。
展开推理会成倍增加输出长度和延迟，实时场景先量代价再决定。
判断有没有用只有一个办法：拿你自己的任务对比测，别信「普遍提升」的说法。
Sort the task type first: arithmetic, multi-step logic, and planning gain a lot; simple extraction, translation, and format conversion gain almost nothing.
Trigger phrases like "think step by step" mattered for older models; reasoning models already have long chains built in, so prompting for one just wastes tokens.
Asking the model to list steps before answering works better — and is more controllable — than a bare "think about it".
Chain-of-thought output is not a reliable record of thinking: the model may fabricate a plausible-sounding process after the fact; don't treat it as an audit log.
Extended reasoning multiplies output length and latency; measure the cost before enabling it in real-time settings.
There is only one way to know if it helps: run a comparison on your own task, don't trust claims of "general improvement".
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
提取一段文字中的日期时，可以直接要求日期和原文位置；没必要要求长篇推理。多条件排期则可要求列出约束、给出方案并核对是否冲突。检查可验证的结果，不把解释越长当作越可靠。For date extraction, request the date and its source location rather than long reasoning. For scheduling, request constraints, a plan and a conflict check. Verify observable results instead of equating longer explanations with reliability.
思维链Chain of Thought (CoT) 推理模型Reasoning Model 温度Temperature 单次调用成本Cost per Call
- Wei 等《Chain-of-Thought Prompting Elicits Reasoning in Large Language Models》Wei et al., Chain-of-Thought Prompting Elicits Reasoning in LLMs ↗
- Zhou 等《Least-to-Most Prompting Enables Complex Reasoning in Large Language Models》Zhou et al., Least-to-Most Prompting ↗
- OpenAI 推理模型指南OpenAI Reasoning models guide ↗