When Does LLM Self-Correction Help? A Control-Theoretic Markov Diagnostic and Verify-First Intervention
A simple control-theoretic test can predict if iterating will degrade performance.
A new paper on arXiv (2604.22273) by Aofan Liu and Jingxiang Meng tackles a critical question in AI agent design: when does letting an LLM refine its own output actually help? Framing iterative self-correction as a cybernetic feedback loop where the model acts as both controller and plant, the authors propose a simple two-state Markov model over {Correct, Incorrect} states. The key diagnostic is the ratio ECR/EIR (Error Correction Rate / Error Introduction Rate) compared to Acc/(1 - Acc). EIR acts as a stability margin, and the paper finds that a near-zero EIR threshold (<= 0.5%) cleanly separates beneficial from harmful self-correction across 7 models and 3 datasets (GSM8K, MATH, StrategyQA).
Only o3-mini (+3.4 pp, EIR = 0%), Claude Opus 4.6 (+0.6 pp, EIR ~ 0.2%), and o4-mini (+/-0 pp) remained non-degrading, while GPT-5 degraded by -1.8 pp. A verify-first prompt ablation on GPT-4o-mini reduced EIR from 2% to 0%, turning a -6.2 pp degradation into +0.2 pp (paired McNemar p < 10^-4). The paper also examines ASC (adaptive stopping criteria), which halts harmful refinement but incurs a 3.8 pp confidence-elicitation cost. The core takeaway: self-correction should be a control decision governed by measurable error dynamics, not a default behavior.
- Simple Markov diagnostic (ECR/EIR ratio) predicts when self-correction helps vs harms accuracy
- Sharp EIR threshold of ~0.5% separates beneficial from harmful iteration across 7 models
- Verify-first prompt on GPT-4o-mini cut EIR from 2% to 0%, turning -6.2 pp degradation into +0.2 pp
Why It Matters
Gives AI engineers a practical, math-backed rule to decide when to let agents self-correct.