**Title: Structured Reasoning Facilitates Accurate Error Localization in Large Language Models

Abstract: While self-correction in language models has long been a challenge, this study investigates the potential for models to explicitly pinpoint errors within flawed reasoning processes, aiming to pave the way for AI systems capable of robust self-improvement. We propose a prompting strategy that organizes reasoning into distinct, semantically unified thought steps. Our findings demonstrate that this structured approach allows models to identify errors with greater reliability compared to traditional, unstructured chain-of-thought methods.

Inspired by the human brain’s ability to monitor mistakes at specific decision junctures and re-evaluate alternatives, we present Iterative Correction Sampling of Thoughts (Thought-ICS), a novel self-correction framework. Thought-ICS guides the model to produce reasoning sequentially, one discrete and complete thought at a time. Since each thought constitutes a deliberate choice by the model, these segments establish clear boundaries that facilitate precise error detection. Once an error is identified, the system backtracks to the most recent verified correct step and generates alternative reasoning paths.

Experimental results show that when prompted to correct reasoning flagged as incorrect by an oracle, Thought-ICS yields a 20-40% improvement in self-correction rates. Furthermore, in fully autonomous scenarios lacking external verification, the framework surpasses current self-correction baseline methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC