Research map / Large language models / Reasoning and chain of thought

Structure Enables Effective Self-Localization of Errors in LLMs

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Thought-ICS, which generates reasoning thought-by-thought in a Thought MDP so models can self-localize the first error and backtrack-resample, yielding gains under oracle verification and a net-positive autonomous variant (Thought-ICS-A).

Problem

LLMs struggle to self-localize errors in their own reasoning, and self-correction often fails because verification is unreliable and monolithic chain-of-thought regeneration confounds resampling with genuine correction.

Approach

Structure reasoning as discrete, model-delimited thoughts (a Thought MDP) so errors can be localized to a step and the trace backtracked to the last verified-correct prefix and resampled; add a confidence safeguard that reverts to the initial answer when verification is low-confidence.

Key results

• Thought-ICS under oracle verification gives 20-40% lift over CoT-based self-correction across models and benchmarks (e.g., OSS-120B on AMC23 reaches 92.5 vs 57.5 initial).
• Resampling from clean prefixes improves correction probability 2-4x over erroneous prefixes.
• Thought-ICS-A achieves +3.9% net lift from verified-accuracy exits, while Single self-verification breaks 10.9% of correct answers and fixes 6.2% of incorrect ones.

Surprising findings

• Self-verification specificity collapses over iterations, so self-correction under self-verification breaks more correct answers than it fixes, even with 9-sample majority or unanimous gating.
• Thought-MDP generation improves initial accuracy by +6.5% on average over token-level CoT, with gains of up to +41 points on AIME for 20B+ models.

Limitations

• Localization oracle is a consensus of frontier LLM judges rather than ground-truth annotation, so localization metrics are relative.
• Evaluation limited to 3-120B models and to math, science, and commonsense benchmarks with verifiable answers.
• Smaller models struggle with formatting instructions during localization, so clean-parse analysis covers only about a quarter of localizations.

Experimental setup

Datasets: AMC23 AIME MATH500-L5 MathQA CSQA GPQA

Benchmarks: AMC23 AIME MATH500-L5 MathQA CSQA GPQA

Metrics: accuracy localization clean-prefix rate edit accuracy recall specificity net correction lift

Baselines: Self-Refine CoVe Token-ICS MV@10 Iterative-CoT

Builds on: Self-Refine Chain-of-Verification Reflexion vLLM

Closest papers by meaning

• Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.89)
• Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.88)
• Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.88)
• A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · similarity 0.87)
• Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.87)
• S^3cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners (AAAI 2025 · similarity 0.87)