Iterative Chain-of-Thought: Self-Correcting Reasoning via Verifier-Guided Feedback Loop

SUMMARY & CONCLUSIONSLarge language models often produce logically flawed reasoning despite generating correct final answers, which limits their trustworthiness in high-stakes applications. This limitation is problematic in reliability-critical systems where both accuracy and verifiable reasoning are essential. This paper presents Iterative Chain-of-Thought Refinement, a lightweight framework employing a fine-tuned verifier to detect and provide targeted feedback on logic errors, guiding the model through iterative self-correction, without modifying the frozen base generator.This implementation uses a fine-tuned Gemma-2B [10] verifier to evaluate a LLaMA-3B [12] generator's reasoning chains, creating a feedback loop that terminates when reasoning becomes fully valid or reaches maximum iterations. Evaluations on the GSM8K [7] and REVEAL [11] benchmarks showed strong improvements in the proportion of valid reasoning chains, rising from 71.7% to 93.5% and from 58.3% to 89.6%, respectively, achieved with minimal computational overhead averaging less than two refinement iterations per problem. These gains were accompanied by improvements in coherence, readability, and conciseness.This is particularly relevant in reliability engineering contexts, where flawed diagnostic inference can compromise system safety and lead to suboptimal maintenance decisions. In reliability-critical domains, practitioners must trace through complex failure mode analyses, root cause investigations, and risk assessments where each logical step must be verifiable and defensible. For instance, in failure mode and effects analysis (FMEA), incorrect reasoning about failure propagation can lead to underestimated risk ratings, inadequate mitigation strategies, and potential safety incidents. Similarly, predictive maintenance systems increasingly rely on AI-driven diagnostic reasoning to recommend maintenance actions, where erroneous logical steps could result in unnecessary downtime or, more critically, missed failure predictions that compromise system integrity.