Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Li, Yuan, Wang, Bo, Gao, Yufei, Yao, Yuqian, Wang, Xinyuan, Yin, Zhangyue, Qiu, Xipeng
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2603.04918
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910042223017984
author Li, Yuan
Wang, Bo
Gao, Yufei
Yao, Yuqian
Wang, Xinyuan
Yin, Zhangyue
Qiu, Xipeng
author_facet Li, Yuan
Wang, Bo
Gao, Yufei
Yao, Yuqian
Wang, Xinyuan
Yin, Zhangyue
Qiu, Xipeng
contents Proximal constraints are fundamental to the stability of the Large Language Model reinforcement learning. While the canonical clipping mechanism in PPO serves as an efficient surrogate for trust regions, we identify a critical bottleneck: fixed bounds strictly constrain the upward update margin of low-probability actions, disproportionately suppressing high-advantage tail strategies and inducing rapid entropy collapse. To address this, we introduce Band-constrained Policy Optimization (BandPO). BandPO replaces canonical clipping with Band, a unified theoretical operator that projects trust regions defined by f-divergences into dynamic, probability-aware clipping intervals. Theoretical analysis confirms that Band effectively resolves this exploration bottleneck. We formulate this mapping as a convex optimization problem, guaranteeing a globally optimal numerical solution while deriving closed-form solutions for specific divergences. Extensive experiments across diverse models and datasets demonstrate that BandPO consistently outperforms canonical clipping and Clip-Higher, while robustly mitigating entropy collapse.
format Preprint
id arxiv_https___arxiv_org_abs_2603_04918
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning
Li, Yuan
Wang, Bo
Gao, Yufei
Yao, Yuqian
Wang, Xinyuan
Yin, Zhangyue
Qiu, Xipeng
Machine Learning
Artificial Intelligence
Proximal constraints are fundamental to the stability of the Large Language Model reinforcement learning. While the canonical clipping mechanism in PPO serves as an efficient surrogate for trust regions, we identify a critical bottleneck: fixed bounds strictly constrain the upward update margin of low-probability actions, disproportionately suppressing high-advantage tail strategies and inducing rapid entropy collapse. To address this, we introduce Band-constrained Policy Optimization (BandPO). BandPO replaces canonical clipping with Band, a unified theoretical operator that projects trust regions defined by f-divergences into dynamic, probability-aware clipping intervals. Theoretical analysis confirms that Band effectively resolves this exploration bottleneck. We formulate this mapping as a convex optimization problem, guaranteeing a globally optimal numerical solution while deriving closed-form solutions for specific divergences. Extensive experiments across diverse models and datasets demonstrate that BandPO consistently outperforms canonical clipping and Clip-Higher, while robustly mitigating entropy collapse.
title BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2603.04918