DGRO: Enhancing LLM Reasoning via Exploration-Exploitation Control and Reward Variance Management
This work addresses a key bottleneck in LLM reasoning for AI researchers, offering an incremental improvement over existing RL methods.
The paper tackles the challenge of balancing exploration and exploitation in reinforcement learning for LLM reasoning by proposing DGRO, a method that decouples regularization parameters and manages reward variance, achieving state-of-the-art performance with 96.9% average accuracy on the Logic dataset.
Inference scaling further accelerates Large Language Models (LLMs) toward Artificial General Intelligence (AGI), with large-scale Reinforcement Learning (RL) to unleash long Chain-of-Thought reasoning. Most contemporary reasoning approaches usually rely on handcrafted rule-based reward functions. However, the tarde-offs of exploration and exploitation in RL algorithms involves multiple complex considerations, and the theoretical and empirical impacts of manually designed reward functions remain insufficiently explored. In this paper, we propose Decoupled Group Reward Optimization (DGRO), a general RL algorithm for LLM reasoning. On the one hand, DGRO decouples the traditional regularization coefficient into two independent hyperparameters: one scales the policy gradient term, and the other regulates the distance from the sampling policy. This decoupling not only enables precise control over balancing exploration and exploitation, but also can be seamlessly extended to Online Policy Mirror Descent (OPMD) algorithms in Kimi k1.5 and Direct Reward Optimization. On the other hand, we observe that reward variance significantly affects both convergence speed and final model performance. We conduct both theoretical analysis and extensive empirical validation to assess DGRO, including a detailed ablation study that investigates its performance and optimization dynamics. Experimental results show that DGRO achieves state-of-the-art performance on the Logic dataset with an average accuracy of 96.9\%, and demonstrates strong generalization across mathematical benchmarks.