One-shot Entropy Minimization for Language Model Reasoning

Zitian Gao, Yilong Chen, Haoming Luo, Joey Zhou, Bryan Dai
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:33069-33080, 2026.

Abstract

In this work, we propose One-shot Entropy Minimization (EM), a simple and fully unsupervised post-training approach that significantly improves reasoning and generation performance using only a single unlabeled data and approximately ten gradient steps. To avoid data contamination, we pretrain a 7-billion-parameter language model from scratch with strictly decontaminated data. Despite its extreme simplicity, one-shot EM yields substantial performance gains and improves reasoning abilities across a broad range of domains, including mathematical reasoning, logical reasoning, and coding. We further show that entropy minimization induces a characteristic right-skewed logit shift, amplifying high-probability tokens while suppressing low-probability tails, in contrast to reinforcement learning. Our findings suggest that entropy minimization primarily acts as a distribution shaping mechanism rather than a conventional learning process, offering an efficient and practical algorithm for post-training large language models.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-gao26d, title = {One-shot Entropy Minimization for Language Model Reasoning}, author = {Gao, Zitian and Chen, Yilong and Luo, Haoming and Zhou, Joey and Dai, Bryan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {33069--33080}, 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/gao26d/gao26d.pdf}, url = {https://proceedings.mlr.press/v306/gao26d.html}, abstract = {In this work, we propose One-shot Entropy Minimization (EM), a simple and fully unsupervised post-training approach that significantly improves reasoning and generation performance using only a single unlabeled data and approximately ten gradient steps. To avoid data contamination, we pretrain a 7-billion-parameter language model from scratch with strictly decontaminated data. Despite its extreme simplicity, one-shot EM yields substantial performance gains and improves reasoning abilities across a broad range of domains, including mathematical reasoning, logical reasoning, and coding. We further show that entropy minimization induces a characteristic right-skewed logit shift, amplifying high-probability tokens while suppressing low-probability tails, in contrast to reinforcement learning. Our findings suggest that entropy minimization primarily acts as a distribution shaping mechanism rather than a conventional learning process, offering an efficient and practical algorithm for post-training large language models.} }
Endnote
%0 Conference Paper %T One-shot Entropy Minimization for Language Model Reasoning %A Zitian Gao %A Yilong Chen %A Haoming Luo %A Joey Zhou %A Bryan Dai %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-gao26d %I PMLR %P 33069--33080 %U https://proceedings.mlr.press/v306/gao26d.html %V 306 %X In this work, we propose One-shot Entropy Minimization (EM), a simple and fully unsupervised post-training approach that significantly improves reasoning and generation performance using only a single unlabeled data and approximately ten gradient steps. To avoid data contamination, we pretrain a 7-billion-parameter language model from scratch with strictly decontaminated data. Despite its extreme simplicity, one-shot EM yields substantial performance gains and improves reasoning abilities across a broad range of domains, including mathematical reasoning, logical reasoning, and coding. We further show that entropy minimization induces a characteristic right-skewed logit shift, amplifying high-probability tokens while suppressing low-probability tails, in contrast to reinforcement learning. Our findings suggest that entropy minimization primarily acts as a distribution shaping mechanism rather than a conventional learning process, offering an efficient and practical algorithm for post-training large language models.
APA
Gao, Z., Chen, Y., Luo, H., Zhou, J. & Dai, B.. (2026). One-shot Entropy Minimization for Language Model Reasoning. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:33069-33080 Available from https://proceedings.mlr.press/v306/gao26d.html.

Related Material