---
id: 20260424-T0-16
title: "大模型不确定性与正确性由不同机制控制"
title_en: "LLM Uncertainty and Correctness Use Distinct Mechanisms"
url: https://ai.daily.yangsir.net/daily/20260424-T0-16
issue_date: 2026-04-24
publish_date: 2026-04-23T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2604.19974
---

# 大模型不确定性与正确性由不同机制控制

UC Berkeley研究发现，大模型的不确定性判断和实际正确性由不同的内部特征控制。团队通过稀疏自编码器分离了相关神经表征，发现“自信错误”和“谨慎正确”存在功能分离。该研究解释了为何有时模型会过度自信地输出错误答案，为提升LLM可靠性提供了新路径。

## English Version

**LLM Uncertainty and Correctness Use Distinct Mechanisms**

UC Berkeley research reveals LLMs' uncertainty and correctness are controlled by distinct neural features. Using sparse autoencoders, the team separated 'confident errors' from 'cautious correct' activations. This explains why models sometimes overconfidently produce wrong answers, offering new paths to improve reliability.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2604.19974)

**详情页**：https://ai.daily.yangsir.net/daily/20260424-T0-16

---

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