Saved in:
Bibliographic Details
Main Authors: Fan, Mingyuan, Han, Weiguang, Wang, Daixin, Chen, Cen, Zhang, Zhiqiang, Zhou, Jun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.15609
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911398821363712
author Fan, Mingyuan
Han, Weiguang
Wang, Daixin
Chen, Cen
Zhang, Zhiqiang
Zhou, Jun
author_facet Fan, Mingyuan
Han, Weiguang
Wang, Daixin
Chen, Cen
Zhang, Zhiqiang
Zhou, Jun
contents Reinforcement Learning with Verifiable Rewards (RLVR) is a central paradigm for turning large language models (LLMs) into reliable problem solvers, especially in logic-heavy domains. Despite its empirical success, it remains unclear whether RLVR elicits novel capabilities or merely sharpens the distribution over existing knowledge. We study this by formalizing over-sharpening, a phenomenon where the policy collapses onto limited modes, suppressing valid alternatives. At a high level, we discover finite-batch updates intrinsically bias learning toward sampled modes, triggering a collapse that propagates globally via semantic coupling. To mitigate this, we propose inverse-success advantage calibration to prioritize difficult queries and distribution-level calibration to diversify sampling via a memory network. Empirical evaluations validate that our strategies can effectively improve generalization.
format Preprint
id arxiv_https___arxiv_org_abs_2601_15609
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle When Sharpening Becomes Collapse: Sampling Bias and Semantic Coupling in RL with Verifiable Rewards
Fan, Mingyuan
Han, Weiguang
Wang, Daixin
Chen, Cen
Zhang, Zhiqiang
Zhou, Jun
Machine Learning
Computation and Language
Reinforcement Learning with Verifiable Rewards (RLVR) is a central paradigm for turning large language models (LLMs) into reliable problem solvers, especially in logic-heavy domains. Despite its empirical success, it remains unclear whether RLVR elicits novel capabilities or merely sharpens the distribution over existing knowledge. We study this by formalizing over-sharpening, a phenomenon where the policy collapses onto limited modes, suppressing valid alternatives. At a high level, we discover finite-batch updates intrinsically bias learning toward sampled modes, triggering a collapse that propagates globally via semantic coupling. To mitigate this, we propose inverse-success advantage calibration to prioritize difficult queries and distribution-level calibration to diversify sampling via a memory network. Empirical evaluations validate that our strategies can effectively improve generalization.
title When Sharpening Becomes Collapse: Sampling Bias and Semantic Coupling in RL with Verifiable Rewards
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2601.15609