The authors introduce M3OPD-Reasoner, a fully open post-training recipe for vision-language models across image STEM, image perception, and video reasoning domains. Starting from Qwen3.5-9B, they curate M3OPD-SFT-1.35M for supervised fine-tuning and M3OPD-RL-193K for reinforcement learning with verifiable rewards, decontaminating both against 21 evaluation benchmarks. They compare mixed-data training (MixSFT and MixRL) against a multi-teacher approach: three domain-specific teachers trained with Group Sequence Policy Optimization (GSPO), either directly from the base model or after an SFT warm-up, then distilled into a single student via Multi-Teacher On-Policy Distillation (MOPD). MOPD of teachers trained with GSPO directly from Qwen3.5-9B proves optimal, with the student matching each teacher on its domain and outperforming mixed-training baselines on average. Treating MOPD as a stronger initialization, the authors then apply MixRL on the same prompts and settings to the MOPD student, demonstrating continued improvement without collapse — a consecutive MOPD-then-MixRL pipeline not previously studied for VLMs. This yields the final M3OPD-Reasoner-9B model. All datasets, training code, and intermediate checkpoints are released.

Read original