Files
myWiki/concepts/overthinking.md

38 lines
1.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
title: "过度思考 (Overthinking)"
created: 2026-06-18
updated: 2026-06-18
type: concept
tags: [reasoning, efficiency, cot, thinking]
sources:
- gan-thinking-based-non-thinking-2026
---
# 过度思考 (Overthinking)
Overthinking 指[[large-reasoning-models|大推理模型]]对**所有查询**(包括简单查询)都产生冗长、重复的[[chain-of-thought|思维链]]CoT导致推理开销和延迟大幅增加的问题Sui et al., 2025; Qu et al., 2025
## 问题本质
LRM 的 CoT 包含持续的探索、反思和自我验证——这是它们在数学竞赛等复杂任务上成功的关键。但这种"不分难易一律思考"的策略导致:
- **推理时延**:简单问题也经历完整的多步推理
- **Token 浪费**:大量"Wait... Let me check..."等无建设性内容
- **计算成本**:每个请求的 FLOPs 和内存消耗显著增加
## 与混合推理的关系
Overthinking 是[[hybrid-reasoning-models|混合推理模型]]的**核心动机**——目标不是消除思考,而是让模型学会"简单问题直接答,复杂问题才思考"。
## TNT 的解决方案
TNT 让模型**自主选择**思考/非思考模式,最终在数学基准上:
- Token 使用量减少约 **50%**
- 准确率**同时提升** 4.1%
- Reward hacking 率 < 10%
## 参考
- [[hybrid-reasoning-models|混合推理模型]]
- [[token-efficiency|Token 效率]]
- [[gan-thinking-based-non-thinking-2026|TNT 论文]]