The Self-Correction Illusion: LLMs Correct Others but Not Themselves

2026-06-04

AI Summary

This study investigates why large language models struggle to correct their own reasoning errors yet effectively fix identical errors when attributed to external roles. By fixing erroneous content and systematically varying only the role labels (e.g., <thought>, user, tool, <memory>) in dialogue templates, the authors isolate the effect of role attribution. Experiments across 13 model–domain combinations employ SHA-256 checksums to ensure content consistency and include cross-model, cross-task comparisons with statistical significance testing. Results show that relabeling errors from <thought> to an external role increases explicit correction rates by 23–93 percentage points (10 comparisons, p<0.001), with <memory> most effective in mathematical tasks and user in logical reasoning. The findings indicate that self-correction failure stems from representational bias induced by role labels—not inherent cognitive limitations—and propose a training-free, prompt-structure intervention to mitigate this issue.

Abstract

Recent work shows that LLM agents struggle to correct errors in their own reasoning traces yet show markedly higher correction rates when identical claims appear under external sources. We ask whether this asymmetry reflects a capability deficit or a role-label artifact: does an agent's willingness to correct a wrong claim depend causally on the chat-template role that carries it, rather than on the claim's content? Our setup keeps the erroneous claim byte-identical across all conditions (SHA-256 verified) and varies only its wrapping role: the agent's own \role{<thought>}, a \role{user} message, a \role{tool} response, or a \role{system <memory>} block. Across 13 model-domain cells covering seven model families and three domains ($n{=}30$ paired tasks per cell), relabeling the claim from \role{<thought>} to an external role lifts the explicit-correction rate by 23 to 93 percentage points, with 10 of 13 cells reaching $p{<}0.001$. Further experiments confirm that the effect is asymmetric, mechanistically decomposable, and robust across domains. The failure to self-correct is not a cognitive deficit; it is a chat-template artifact. We exploit this artifact by designing a prompt-structure-only intervention that requires no training and no model modification, with its strongest role label being domain-dependent: \role{<memory>} dominates on math, while a plain \role{user} message dominates on logical deduction.

Problem

• self-correction
• large language models
• role labeling
• reasoning errors
• chat template

Innovation

• self-correction illusion
• role-label artifact
• chat-template
• prompt engineering
• large language models