Researchers have developed a new framework called Open-MOPD to address the issue of capability imbalance in multi-teacher on-policy distillation. This technique combines multiple reinforcement learning experts into a single generalist student. However, previous methods have been limited by their inability to effectively integrate capabilities from different domains. The authors of this paper created a controlled benchmark and found that standard M-OPD captures only 35.6% of the available headroom relative to an oracle ensemble. They attribute this failure to misallocation of optimization budget due to structural sequence-length disparities, dynamic convergence drift, and multi-step reward staleness. Open-MOPD addresses these issues by incorporating token-share balancing, gap-aware dynamic budget allocation, and student reward refresh. --- Why it matters: This matters because it provides a more effective way to integrate capabilities from multiple domains in reinforcement learning, which can lead to better performance on complex tasks. By addressing the capability imbalance issue, researchers can create more generalizable and robust AI models. Source: https://arxiv.org/abs/2608.19098