Relabel an agent's own draft as 'external submission' before review — lifts self-correction 23–93 points

arXiv·high signal

A new paper shows LLMs barely correct errors in their own reasoning traces but readily correct the identical claim when it's attributed to an external source — the 'self-correction illusion.' Across model families and domains, simply relabeling a claim from the agent's own role to an external role raises the explicit-correction rate by 23 to 93 percentage points. Actionable for multi-agent pipelines: when an agent reviews its prior output, frame it as a third party's submission ('Critique this submission from another engineer'), not 'review your own work.'