Entropy Ratio Clipping as a Soft Global Constraint for Stable Reinforcement Learning

arXiv:2512.05591v2 Announce Type: replace-cross Abstract: Large language model post-training relies on reinforcement learning to improve model capability and alignment quality. However, the off-policy training paradigm introduces distribution shift, which often pushes the policy beyond the trust region, leading to training instabilities manifested as fluctuations in policy entropy and unstable gradients. Although PPO-Clip mitigates this issue through importance clipping, it still overlooks the global distributional shift of actions. To address these challenges, we propose using the entropy ratio between the current and previous policies as a new global metric that effectively quantifies the relative change in policy exploration throughout updates. Building on this metric, we introduce an extbf{Entropy Ratio Clipping} (ERC) mechanism that imposes bidirectional constraints on the entropy ratio. This stabilizes policy updates at the global distribution level and compensates for the inability of PPO-clip to regulate probability shifts of un-sampled actions. We integrate ERC into both DAPO and GPPO reinforcement learning algorithms. Experiments across multiple benchmarks show that ERC consistently improves performance.

Related

• CE-GPPO: Coordinating Entropy via Gradient-Preserving Clipping Policy Optimization in Reinforcement Learning
• Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning
• AAPO: Enhancing the Reasoning Capabilities of LLMs with Advantage Margin
• MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization
• Reinforcement-aware Knowledge Distillation for LLM Reasoning

Source: arXiv cs.CL | 2026-04-24