Structure Enables Effective Self-Localization of Errors in LLMs

Ankur Samanta, Akshayaa Magesh, Ayush Jain, Kavosh Asadi, Youliang Yu, Daniel R. Jiang, Boris Vidolov, Kaveh Hassani, Paul Sajda, Jalaj Bhandari, Yonathan Efroni
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:107412-107440, 2026.

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 can localize errors more reliably within this structure than 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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-samanta26a, title = {Structure Enables Effective Self-Localization of Errors in {LLM}s}, author = {Samanta, Ankur and Magesh, Akshayaa and Jain, Ayush and Asadi, Kavosh and Yu, Youliang and Jiang, Daniel R. and Vidolov, Boris and Hassani, Kaveh and Sajda, Paul and Bhandari, Jalaj and Efroni, Yonathan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {107412--107440}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/samanta26a/samanta26a.pdf}, url = {https://proceedings.mlr.press/v306/samanta26a.html}, 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 can localize errors more reliably within this structure than 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.} }
Endnote
%0 Conference Paper %T Structure Enables Effective Self-Localization of Errors in LLMs %A Ankur Samanta %A Akshayaa Magesh %A Ayush Jain %A Kavosh Asadi %A Youliang Yu %A Daniel R. Jiang %A Boris Vidolov %A Kaveh Hassani %A Paul Sajda %A Jalaj Bhandari %A Yonathan Efroni %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-samanta26a %I PMLR %P 107412--107440 %U https://proceedings.mlr.press/v306/samanta26a.html %V 306 %X 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 can localize errors more reliably within this structure than 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.
APA
Samanta, A., Magesh, A., Jain, A., Asadi, K., Yu, Y., Jiang, D.R., Vidolov, B., Hassani, K., Sajda, P., Bhandari, J. & Efroni, Y.. (2026). Structure Enables Effective Self-Localization of Errors in LLMs. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:107412-107440 Available from https://proceedings.mlr.press/v306/samanta26a.html.

Related Material