arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2606.05976cs.AIcs.CL
收藏

自我修正错觉:LLM 纠正他人但不纠正自己

The Self-Correction Illusion: Role Relabeling Gates Explicit Error Flagging in Large Language Models

National Taiwan University(国立台湾大学)

Kuan-Yen Chen, Fang-Yi Su, Shih-Yen Lin, Bao Li, Jung-Hsien Chiang

AI总结:

本文通过保持错误声明字节一致仅改变角色标签,发现 LLM 无法自我修正并非能力缺陷,而是聊天模板角色标签的人为产物,并提出无需训练或模型修改的提示结构干预方法。

摘要:

近期研究表明,LLM 智能体难以纠正自身推理轨迹中的错误,但当相同声明出现在外部来源时,其修正率显著更高。我们探究这种不对称性反映的是能力缺陷还是角色标签的人为产物:智能体纠正错误声明的意愿是否因果地依赖于承载该声明的聊天模板角色,而非声明内容本身?我们的实验设置在所有条件下保持错误声明的字节完全一致(SHA-256 验证),仅改变其包装角色:智能体自身的 \role{<thought>}、\role{user} 消息、\role{tool} 响应或 \role{system <memory>} 块。在覆盖七个模型家族和三个领域的 13 个模型-领域单元(每个单元 n=30 对任务)中,将声明从 \role{<thought>} 重新标记为外部角色后,显式修正率提升了 23 到 93 个百分点,其中 13 个单元中有 10 个达到 p<0.001。进一步实验证实该效应是不对称的、机制上可分解的,并且跨领域稳健。自我修正失败并非认知缺陷,而是聊天模板的人为产物。我们利用这一人为产物设计了一种仅涉及提示结构、无需训练和模型修改的干预方法,其最强角色标签依赖于领域:在数学上 \role{<memory>} 占主导,而在逻辑推理上普通 \role{user} 消息占主导。

英文摘要:

Recent works show that LLM agents struggle to correct errors in their own reasoning traces, despite their ability to correct errors from external sources. We ask whether this reflects a capability deficit or an artifact of the role labeling. To test this, we design a training-free intervention, source-conditioned role relabeling, that keeps the erroneous claim byte-identical and varies only its message role. The claim is presented inside the agent's "<thought>", a user message, a tool response, or a system "<memory>" block. We test 12 model-domain combinations spanning closed-weight APIs and open-weight models from 70B-class down to smaller families. Relabeling "<thought>" to an external role increases the explicit-correction rate by 23 to 93 percentage points, significant in 10 of 12 experimental settings. This suggests that these models' failure to detect a self-generated error is largely an artifact of how the claim is role-labeled in the chat template, rather than a pure cognitive deficit. The most effective role label is domain-dependent: "<memory>" dominates in most math experiments, while a user message dominates in logical deduction. Recognizing role-label handling as a key experimental variable in instruction tuning presents a more direct path to closing the self-correction gap.d

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑