LLM Self-Correction: When It Helps vs. Hurts

A new study published on arXiv (2604.22273) discusses the benefits of iterative self-correction in agentic LLM systems. The researchers view self-correction as a feedback loop, where one model functions as both the controller and the system. They apply a two-state Markov model with the states {Correct, Incorrect} to set a criterion for deployment: self-correct only when the ratio ECR/EIR is greater than Acc/(1 - Acc). Here, EIR represents a stability margin, while prompting is a straightforward controller approach. Tests on 7 models across 3 datasets (GSM8K, MATH, StrategyQA) uncovered a critical EIR threshold of 0.5% that differentiates effective from ineffective self-correction. Only o3-mini, Claude Opus 4.6, and o4-mini maintained performance, whereas GPT-5 declined by 1.8 pp. An additional study showed that prompting can affect this threshold.