---
id: 20260729-T0-04
title: "LLM推理失败新发现：过早自我怀疑导致「上下文焦虑」"
title_en: "LLMs Fail Due to Premature Self-Doubt, Not Lack of Ability"
url: https://ai.daily.yangsir.net/daily/20260729-T0-04
issue_date: 2026-07-29
publish_date: 2026-07-28T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2607.21616
---

# LLM推理失败新发现：过早自我怀疑导致「上下文焦虑」

arXiv论文「Lost in Context」发现，前沿推理模型并非因能力不足而答错题，而是在推理过程中过早产生自我怀疑（即「上下文焦虑」），从而放弃正确思路。这与传统认知——模型超出能力范围就会失败——不同。论文提出识别和干预这种焦虑的方法，有望提升模型推理的鲁棒性。

## English Version

**LLMs Fail Due to Premature Self-Doubt, Not Lack of Ability**

A paper titled 'Lost in Context' shows that frontier reasoning models fail not because they lack capability, but due to premature self-doubt during inference — termed 'context anxiety.' Unlike conventional wisdom, models may possess needed skills but yield incorrect answers. The paper suggests interventions to improve reasoning robustness.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2607.21616)

**详情页**：https://ai.daily.yangsir.net/daily/20260729-T0-04

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*