Denoising Iterative Self-Correction: Structured Verification Loops for Reliable LLM Reasoning

The paper introduces Denoising Iterative Self-Correction (DISC), a novel test-time procedure designed to enhance multi-step reasoning in large language models by using verification outputs as noisy signals to progressively reduce errors. DISC employs a binary judgment gate to maintain the integrity of correct answers while iteratively correcting mistakes, achieving an accuracy of 81.6% on the BIG-Bench Mistake benchmark and outperforming existing methods like Chain-of-Verification and Self-Refine in precision-recall metrics. This approach is significant for practitioners as it offers a structured method to improve the reliability of LLM outputs, particularly in complex reasoning tasks.