Structure Enables Effective Self-Localization of Errors in LLMs

TL;DR

The paper investigates whether LLMs can explicitly localize errors in their own reasoning to enable self-correction. It introduces Thought-ICS and the Thought MDP, which structure reasoning into discrete thoughts and use verification, localization, and resampling to backtrack and revise incorrect steps. Under oracle verification, Thought-ICS achieves a 20-40% lift over unstructured baselines, with larger models showing stronger localization and correction performance. In autonomous settings, Thought-ICS-A demonstrates net positive correction lift when combined with confidence safeguards, though self-verification remains a key bottleneck. The work highlights the practical potential of structured thinking for robust self-correction while acknowledging limitations and directions for improving verification reliability and scalability.

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.

Figures (20)

• Figure 1: Illustration of the Thought-ICS framework. Generation: The model is tasked with sampling the next action/thought given a prefix of the original prompt and the previously sampled steps. We call this a Thought MDP, as it constructs a structured reasoning trace thought by thought ($a_1, a_2, \ldots$) until it deems it has finished the task. Verification: A verification signal (configurable: self or external verification) indicates whether the final answer is correct or not; this gates whether the system continues or exits. Localization: The model is provided the full reasoning trace and asked to analyze each thought to identify the first thought where an error has been made. If it still cannot find an error, then it exits the loop. Resampling: Having identified an erroneous thought, the inference framework then backtracks to the last verified correct step, and resumes generation thought by thought from that shared prefix. This loop repeats until reasoning is verified as correct or a termination condition is met. In this example, the model corrects errors at steps 10 and 6 over two iterations before arriving at the correct answer. See Appendix \ref{['app:tree_viz_example']} for the full correction example that this visualization corresponds to.
• Figure 2: Self-correction under an oracle verifier. Localization on structured reasoning (Thought-ICS, red) achieves higher self-correction lift than within unstructured chain-of-thought (Token-ICS, green), both operating on known incorrect solutions. Light red shows initial lift from thought-by-thought generation (see App. \ref{['app:initial_accuracy']}), solid red shows additional lift from self-correction (see Tab. \ref{['tab:l2_results']}, App. \ref{['app:l2_full_results']}).
• Figure 3: Self-localization within structure vs oracle localization. Clean prefix (left bar): self $\leq$ oracle, comprising exact matches (green) and earlier localizations (blue). Erroneous prefix (right bar, red): self $>$ oracle.
• Figure 4: Self-correction accuracy within structure when resampling from clean prefixes (green) is substantially higher than from erroneous prefixes (red). Localization quality directly impacts correction performance within structure.
• Figure 5: Distribution of localization deviation (self $-$ oracle). Larger models show tighter distributions centred at zero, indicating more precise error localization with structured generation rather than LLMs being more conservative.