Structure Enables Effective Self-Localization of Errors in LLMs

TL;DR

This paper presents a structured prompting method and a self-correction framework called Thought-ICS that enable language models to localize and correct errors in reasoning by breaking down thought processes into discrete steps.

Contribution

The paper introduces a novel structured reasoning approach and the Thought-ICS framework, improving error localization and self-correction in language models compared to unstructured methods.

Findings

Thought-ICS reliably localizes errors within structured reasoning steps.

It achieves 20-40% self-correction improvement with oracle verification.

Outperforms existing self-correction baselines in autonomous settings.

Abstract

Self-correction in language models remains elusive. In this work, we explore whether language models can explicitly localize errors in incorrect reasoning, as a path toward building AI systems that can effectively correct themselves. We introduce a prompting method that structures reasoning as discrete, semantically coherent thought steps, and show that models are able to reliably localize errors within this structure, while failing to do so in conventional, unstructured chain-of-thought reasoning. Motivated by how the human brain monitors errors at discrete decision points and resamples alternatives, we introduce Iterative Correction Sampling of Thoughts (Thought-ICS), a self-correction framework. Thought-ICS iteratively prompts the model to generate reasoning one discrete and complete thought at a time--where each thought represents a deliberate decision by the model--creating natural boundaries for precise error localization. Upon verification, the model localizes the first erroneous step, and the system backtracks to generate alternative reasoning from the last correct point. When asked to correct reasoning verified as incorrect by an oracle, Thought-ICS achieves 20-40% self-correction lift. In a completely autonomous setting without external verification, it outperforms contemporary self-correction baselines.