ReDit: Reward Dithering for Improved LLM Policy Optimization
This addresses optimization stability issues in LLM training, offering a practical improvement for researchers and practitioners, though it is incremental as it builds on existing reward-based methods.
The paper tackles the problem of discrete rewards causing gradient anomalies and slow convergence in LLM policy optimization by proposing ReDit, which adds random noise to dither the reward signal, achieving comparable performance to vanilla GRPO with about 10% of the training steps and a 4% improvement when trained similarly.
DeepSeek-R1 has successfully enhanced Large Language Model (LLM) reasoning capabilities through its rule-based reward system. While it's a ''perfect'' reward system that effectively mitigates reward hacking, such reward functions are often discrete. Our experimental observations suggest that discrete rewards can lead to gradient anomaly, unstable optimization, and slow convergence. To address this issue, we propose ReDit (Reward Dithering), a method that dithers the discrete reward signal by adding simple random noise. With this perturbed reward, exploratory gradients are continuously provided throughout the learning process, enabling smoother gradient updates and accelerating convergence. The injected noise also introduces stochasticity into flat reward regions, encouraging the model to explore novel policies and escape local optima. Experiments across diverse tasks demonstrate the effectiveness and efficiency of ReDit. On average, ReDit achieves performance comparable to vanilla GRPO with only approximately 10% the training steps, and furthermore, still exhibits a 4% performance improvement over vanilla GRPO when trained for a similar duration. Visualizations confirm significant mitigation of gradient issues with ReDit. Moreover, theoretical analyses are provided to further validate these advantages.