Researchers have proposed a new method called SMOPD for improving model performance in multi-reward reinforcement learning. The issue with existing methods is that they can struggle to balance reward signals of different granularities. For example, a dense reward might assign fine-grained scores, while a sparse reward only provides binary feedback. To address this limitation, the authors propose a two-stage training method called SMOPD. In the first stage, multiple teachers are trained using reward-priority configurations, allowing each reward to be learned under conditions where its signal can effectively drive optimization. In the second stage, online policy distillation is used to combine the reward-specialized capabilities of these teachers into a single student policy. The authors claim that SMOPD outperforms existing methods in experiments on two multi-reward settings. --- Why it matters: This matters because it provides an improvement over existing methods for handling complex reward signals in reinforcement learning, which can be challenging to balance. Engineers and researchers may find this useful for developing more effective policies in real-world applications. Source: https://arxiv.org/abs/2608.03092