M3OPD-Reasoner open recipe distills multi-domain VLM teachers
Original title: M3OPD-Reasoner: An Open Recipe for Post-Training Multimodal Reasoning with Multi-Teacher On-Policy Distillation
Distilling GSPO teachers from Qwen3.5-9B beat mixed training, and a later MixRL stage produced M3OPD-Reasoner-9B.
Researchers present M3OPD-Reasoner, an open post-training recipe for vision-language models across image STEM, image perception and video. Starting from Qwen3.5-9B, they compare mixed SFT and RL with per-domain teachers trained by Group Sequence Policy Optimization and distilled by multi-teacher on-policy distillation. On 21 decontaminated benchmarks, MOPD of teachers trained directly with GSPO was best, and adding MixRL on the same prompts further improved the student without collapse. All data, training code and stage checkpoints are released.
- SFT set is M3OPD-SFT-1.35M; RL set is M3OPD-RL-193K.
- The student matches each teacher on its domain and averages above mixed-training baselines.
- Authors say no prior VLM recipe studied MixRL after MOPD on the same prompts.
From the original article
1 min read at huggingface.coMulti-Teacher On-Policy Distillation (MOPD) enables a single student to absorb the skills of several domain teachers by learning from its own generated responses. We present Multimodal Model Multi-Teacher On-Policy Distillation Reasoner (M3OPD-Reasoner), a fully open recipe for post-training vision-language models (VLMs) over three reasoning domains: image STEM, image perception, and video.
By Nguyen Quang Trung, Hoang Phan, Shuo Sun · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.