Enregistré dans:
Détails bibliographiques
Auteurs principaux: Liang, Zhenwen, Zhou, Yujun, Lu, Sidi, Zhang, Xiangliang, Mi, Haitao, Yu, Dong
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:https://arxiv.org/abs/2604.18493
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917423126413312
author Liang, Zhenwen
Zhou, Yujun
Lu, Sidi
Zhang, Xiangliang
Mi, Haitao
Yu, Dong
author_facet Liang, Zhenwen
Zhou, Yujun
Lu, Sidi
Zhang, Xiangliang
Mi, Haitao
Yu, Dong
contents Reinforcement Learning (RL) enhances LLM reasoning, yet a paradox emerges as models scale: strong base models saturate standard benchmarks (e.g., MATH), yielding correct but homogeneous solutions. In such environments, the lack of failure cases causes the advantage signal in group-relative algorithms (e.g., GRPO) to vanish, driving policies into mode collapse. To address this, we propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration. Unlike standard sampling that follows model biases, CUTS flattens the local optimization landscape by sampling uniformly from constrained high-confidence candidates. We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance. Experiments on Qwen3 models demonstrate that our approach prevents policy degeneration and significantly boosts out-of-domain generalization. Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO, validating that maintaining diversity within the semantic manifold is critical for rigorous reasoning.
format Preprint
id arxiv_https___arxiv_org_abs_2604_18493
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data
Liang, Zhenwen
Zhou, Yujun
Lu, Sidi
Zhang, Xiangliang
Mi, Haitao
Yu, Dong
Machine Learning
Reinforcement Learning (RL) enhances LLM reasoning, yet a paradox emerges as models scale: strong base models saturate standard benchmarks (e.g., MATH), yielding correct but homogeneous solutions. In such environments, the lack of failure cases causes the advantage signal in group-relative algorithms (e.g., GRPO) to vanish, driving policies into mode collapse. To address this, we propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration. Unlike standard sampling that follows model biases, CUTS flattens the local optimization landscape by sampling uniformly from constrained high-confidence candidates. We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance. Experiments on Qwen3 models demonstrate that our approach prevents policy degeneration and significantly boosts out-of-domain generalization. Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO, validating that maintaining diversity within the semantic manifold is critical for rigorous reasoning.
title Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data
topic Machine Learning
url https://arxiv.org/abs/2604.18493