Megadose The AI curve, compressed.

M3OPD-Reasoner: An Open Recipe for Post-Training Multimodal Reasoning with Multi-Teacher On-Policy Distillation

· HF Daily Papers ·
Reusing the same training prompts for reinforcement learning improved the distilled model further, without collapse.

Starting from Qwen3.5-9B, the researchers trained specialist teachers for image STEM, image perception and video, then distilled their skills into one student. Across 21 benchmarks, the best distillation recipe closely matched each specialist in its domain and outperformed mixed-domain training baselines on average. Adding mixed-domain reinforcement learning produced M3OPD-Reasoner-9B. The authors release the datasets, training code and checkpoints for every stage.

HF Daily Papers' note

score 5

Categories: Research