DyJR: Preserving Diversity in Reinforcement Learning with Verifiable Rewards via Dynamic Jensen-Shannon Replay

TL;DR

DyJR introduces a dynamic replay buffer and Jensen-Shannon divergence regularization to enhance diversity and efficiency in reinforcement learning for language models, outperforming existing methods on reasoning and Text-to-SQL tasks.

Contribution

The paper proposes DyJR, a novel regularization framework with a dynamic buffer and distributional constraint to preserve diversity and improve training efficiency in RL for language models.

Findings

DyJR outperforms GRPO, RLEP, and Ex-GRPO on benchmarks.

DyJR maintains training efficiency comparable to GRPO.

DyJR enhances diversity and reduces over-reliance on top tokens.

Abstract

While Reinforcement Learning (RL) enhances Large Language Model reasoning, on-policy algorithms like GRPO are sample-inefficient as they discard past rollouts. Existing experience replay methods address this by reusing accurate samples for direct policy updates, but this often incurs high computational costs and causes mode collapse via overfitting. We argue that historical data should prioritize sustaining diversity rather than simply reinforcing accuracy. To this end, we propose Dynamic Jensen-Shannon Replay (DyJR), a simple yet effective regularization framework using a dynamic reference distribution from recent trajectories. DyJR introduces two innovations: (1) A Time-Sensitive Dynamic Buffer that uses FIFO and adaptive sizing to retain only temporally proximal samples, synchronizing with model evolution; and (2) Jensen-Shannon Divergence Regularization, which replaces direct gradient updates with a distributional constraint to prevent diversity collapse. Experiments on mathematical reasoning and Text-to-SQL benchmarks demonstrate that DyJR significantly outperforms GRPO as well as baselines such as RLEP and Ex-GRPO, while maintaining training efficiency comparable to the original GRPO. Furthermore, from the perspective of Rank-$k$ token probability evolution, we show that DyJR enhances diversity and mitigates over-reliance on Rank-1 tokens, elucidating how specific sub-modules of DyJR influence the training dynamics.