Entropy-Preserving Reinforcement Learning

TL;DR

This paper highlights the importance of actively monitoring and controlling entropy in policy gradient reinforcement learning to maintain diversity, improve performance, and enhance adaptability in language model reasoning tasks.

Contribution

It introduces explicit entropy control mechanisms, including REPO and ADAPO, to prevent entropy reduction and preserve exploration during training.

Findings

Entropy naturally decreases in policy gradient training without intervention.

Entropy-preserving methods improve policy diversity and final performance.

Models with entropy control retain adaptability in new environments.

Abstract

Policy gradient algorithms have driven many recent advancements in language model reasoning. An appealing property is their ability to learn from exploration on their own trajectories, a process crucial for fostering diverse and creative solutions. As we show in this paper, many policy gradient algorithms naturally reduce the entropy -- and thus the diversity of explored trajectories -- as part of training, yielding a policy increasingly limited in its ability to explore. In this paper, we argue that entropy should be actively monitored and controlled throughout training. We formally analyze the contributions of leading policy gradient objectives on entropy dynamics, identify empirical factors (such as numerical precision) that significantly impact entropy behavior, and propose explicit mechanisms for entropy control. These include REPO, a family of algorithms that modify the advantage function to regulate entropy, and ADAPO, an adaptive asymmetric clipping approach. Models trained with our entropy-preserving methods maintain diversity throughout training, yielding final policies that are more performant and retain their trainability for sequential learning in new environments.