ReDit: Reward Dithering for Improved LLM Policy Optimization

TL;DR

ReDit introduces reward dithering by adding noise to discrete reward signals in LLM policy optimization, resulting in smoother training, faster convergence, and better exploration, validated through experiments and theory.

Contribution

The paper proposes ReDit, a novel reward dithering technique that improves LLM policy optimization by mitigating gradient issues and enhancing exploration.

Findings

ReDit reduces training steps by approximately 90% to reach comparable performance.

ReDit achieves a 4% performance boost over vanilla GRPO with similar training duration.

Visualizations show significant mitigation of gradient anomalies with ReDit.

Abstract

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.