Efficient and Stable Reinforcement Learning for Diffusion Language Models
This work addresses a domain-specific problem for researchers and practitioners using diffusion language models, offering incremental improvements in RL efficiency and stability.
The paper tackles the efficiency and stability challenges of applying reinforcement learning to diffusion-based large language models by proposing Spatio-Temporal Pruning (STP), which compresses redundancy in the generative process and reduces variance in log-likelihood estimation, leading to improved performance over state-of-the-art baselines.
Reinforcement Learning (RL) is crucial for unlocking the complex reasoning capabilities of Diffusion-based Large Language Models (dLLMs). However, applying RL to dLLMs faces unique challenges in efficiency and stability. To address these challenges, we propose Spatio-Temporal Pruning (STP), a framework designed to simultaneously improve the efficiency and stability of RL for dLLMs. STP compresses the redundancy in the generative process through: (1) \textit{spatial pruning}, which constrains the exploration space using static priors; and (2) \textit{temporal pruning}, which bypasses redundant late-stage refinement steps. Our theoretical analysis demonstrates that STP strictly reduces the variance of the log-likelihood estimation, thereby ensuring more stable policy updates. Extensive experiments demonstrate that STP surpasses state-of-the-art baselines in both efficiency and accuracy. Our code is available at https://github.com/Lolo1222/STP.