The appeal of iterative self-correction in agentic LLM systems is intuitive: if a model makes an error, let it review and refine. Yet empirical reality diverges sharply from this intuition. Recent work demonstrates that for most contemporary models, repeated self-refinement actively degrades performance—a phenomenon that has lacked principled explanation until now. A new framework grounded in control theory and Markov dynamics provides both diagnostic clarity and actionable interventions, challenging the default assumption that iteration improves outcomes.

This matters immediately for practitioners deploying LLM agents. The difference between beneficial and harmful self-correction hinges on a single measurable quantity: the error-injection rate (EIR). When this quantity exceeds approximately 0.5%, iteration becomes net-negative. For models like GPT-4o and GPT-5, this threshold violation translates to degradation ranging from -1.8 to -6.2 percentage points on standard benchmarks. Understanding why and how to reverse this degradation is critical for any system relying on multi-step reasoning or agentic loops.

The authors reframe self-correction as a cybernetic feedback control problem. In this view, the language model simultaneously acts as both the controller (the entity making refinement decisions) and the plant (the system being controlled). This dual role creates an inherent instability: the same model that generated the initial error must now diagnose and correct it, without access to external ground truth during deployment. The framework operationalizes this intuition through a two-state Markov chain over {Correct, Incorrect} states, where transitions depend on three empirical quantities: the error-correction rate (ECR), the error-injection rate (EIR), and the baseline accuracy (Acc).

The diagnostic criterion is elegantly simple: iterate when ECR/EIR > Acc/(1 - Acc). This inequality encodes a fundamental trade-off. The left side represents the benefit-to-risk ratio of iteration—how many errors the model fixes relative to how many it introduces. The right side represents the baseline stakes—how much room exists for degradation given current accuracy. When this inequality holds, iteration is expected to improve the distribution over many calls. When it fails, repeated refinement becomes a stability hazard, with EIR functioning as the critical margin. The authors argue that prompting design functions as "lightweight controller design"—adjusting the prompt can shift EIR without retraining, thereby moving systems across the stability boundary.

Empirical validation spans 7 models and 3 reasoning benchmarks (GSM8K, MATH, StrategyQA), revealing a sharp phase transition. Models with EIR ≤ 0.5% maintain non-degrading iteration; those exceeding this threshold suffer systematic decline. Only three models achieve sub-threshold EIR: o3-mini (EIR = 0%, +3.4 pp improvement), Claude Opus 4.6 (EIR ≈ 0.2%, +0.6 pp), and o4-mini (no degradation). Notably, GPT-5 violates the threshold, yielding -1.8 pp degradation—a sobering result for a frontier model. GPT-4o-mini exhibits particularly severe failure (-6.2 pp), with EIR at 2%.

The paper provides causal evidence that this threshold is actionable through prompting alone. A "verify-first" intervention—instructing the model to assess correctness before attempting refinement—reduces GPT-4o-mini's EIR from 2% to 0%, simultaneously converting its -6.2 pp degradation into +0.2 pp improvement (paired McNemar test, p < 10⁻⁴). This result is striking: a single prompt modification crosses the stability boundary, transforming harmful iteration into beneficial refinement. Crucially, the same intervention produces minimal change on already-sub-threshold models, suggesting the effect is precisely calibrated to the underlying error dynamics rather than a generic performance boost.

A secondary finding involves adaptive self-correction (ASC)—a mechanism that halts refinement when confidence signals suggest further iteration would be harmful. ASC successfully prevents degradation but incurs a 3.8 pp cost in confidence-elicitation accuracy, representing a trade-off between stability and information extraction. This cost-benefit calculation will vary across applications; for safety-critical domains, accepting the accuracy penalty may be justified by avoiding compounded errors.

CuraFeed Take: This work inverts the conventional narrative around LLM self-correction. Rather than treating iteration as a default behavior to enable—as most agentic frameworks currently do—it should be treated as a control decision governed by measurable error dynamics. The practical implication is immediate: before deploying iterative refinement, measure EIR on your task and model combination. If it exceeds 0.5%, either (a) redesign your prompt to reduce error injection, or (b) disable iteration entirely. The verify-first ablation demonstrates that prompt engineering can be surprisingly effective at shifting EIR, offering a low-cost intervention before considering model upgrades.

The frontier model asymmetry is noteworthy. o3-mini and Claude Opus have apparently learned to correct errors without introducing new ones—a capability absent in GPT-4 variants. This suggests that scaling and/or alignment techniques used in the latest models confer genuine robustness to self-correction, not merely higher baseline accuracy. For researchers, the Markov framework itself is valuable: it provides a principled language for analyzing iterative refinement across any LLM, decoupling the stability question from task-specific performance metrics. Watch for follow-up work exploring whether EIR correlates with model scale, training data composition, or specific architectural choices—answers would clarify whether self-correction robustness is an inevitable property of frontier models or a learnable skill.