The Self-Correction Illusion: LLMs correct others but not themselves

Category: ML research
Status: developing
Sources: arXiv:2606.05976
Summary: Chen, Su, and Chiang (submitted 2026-06-04) tested seven LLM families on 30 paired tasks across three domains, keeping erroneous claims byte-identical and varying only the wrapping role: the model's own thought, a user message, a tool response, or a system memory block. Relabeling a claim from the model's own thought to any external role lifted the explicit-correction rate by 23 to 93 percentage points. 10 of 13 model-domain cells reached maximum correction rates when the claim appeared as a tool or system memory entry.
Why it matters: Multi-agent architectures where independent agents review each other's outputs are substantially more reliable than single-agent self-correction loops; agent system designers should treat self-correction as unreliable by default.