D3-MOPD is a zero-overhead scheduler for multi-teacher on-policy distillation that adjusts per-domain training mixture ratios online, reusing the reverse-KL signal already produced during training instead of relying on a fixed data mixture. An asynchronous watcher tracks each domain’s convergence rate and remaining headroom to reallocate the sampling ratio accordingly. On a Qwen3.6-35B-A3B student distilled from four domain-expert teachers, D3-MOPD closes 97% of the student-to-teacher performance gap versus 63% for a fixed-mixture baseline, while needing roughly 3x fewer rollout steps to reach the same peak performance.
