BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning
This work provides an incremental improvement to the PPO algorithm for large language model reinforcement learning, specifically addressing the problem of entropy collapse and exploration for researchers and practitioners in this domain.
This paper addresses the issue of fixed clipping bounds in PPO, which restrict the update margin for low-probability actions and lead to entropy collapse in LLM reinforcement learning. They propose BandPO, a method that uses dynamic, probability-aware clipping intervals derived from f-divergences, which they show consistently outperforms canonical clipping and Clip-Higher across various models and datasets.
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.