DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning

Researchers introduced DVAO, a new multi-reward RL optimization method that adapts objective weights using empirical reward variance.

DVAO keeps advantage magnitudes bounded, aiming to reduce instability and rigidity in multi-objective training.

The paper reports stronger results than baseline methods on reasoning and tool-use benchmarks, including models like Qwen3 and Qwen2.5.

The approach is positioned as a practical step toward better Pareto-efficient training for language models with competing rewards.