Flexible Entropy Control in RLVR with Gradient-Preserving Perspective
This work addresses entropy collapse in RLVR for LLMs, an incremental improvement in training stability and output diversity.
The paper tackled the problem of policy entropy collapse in Reinforcement Learning with Verifiable Rewards (RLVR) for Large Language Models, which causes premature overconfidence and reduced diversity, by introducing dynamic clipping threshold strategies to manage entropy, resulting in effective mitigation and superior performance across benchmarks.
Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a critical method for enhancing the reasoning capabilities of Large Language Models (LLMs). However, continuous training often leads to policy entropy collapse, characterized by a rapid decay in entropy that results in premature overconfidence, reduced output diversity, and vanishing gradient norms that inhibit learning. Gradient-Preserving Clipping is a primary factor influencing these dynamics, but existing mitigation strategies are largely static and lack a framework connecting clipping mechanisms to precise entropy control. This paper proposes reshaping entropy control in RL from the perspective of Gradient-Preserving Clipping. We first theoretically and empirically verify the contributions of specific importance sampling ratio regions to entropy growth and reduction. Leveraging these findings, we introduce a novel regulation mechanism using dynamic clipping threshold to precisely manage entropy. Furthermore, we design and evaluate dynamic entropy control strategies, including increase-then-decrease, decrease-increase-decrease, and oscillatory decay. Experimental results demonstrate that these strategies effectively mitigate entropy collapse, and achieve superior performance across multiple benchmarks.