huggingface.co web signal

M³OPD-Reasoner Lifts Qwen3.5-9B 7.1% Across 21 VLM Benchmarks

TL;DR

  • The 9B student, built from three GSPO teachers distilled via MOPD then topped with mixed RL, averages 77.1% across 21 benchmarks, up from 70.0%.
  • Video reasoning sees the biggest jump at +10.7% average, with LongVideoBench rising 12.3 points and Video-Holmes 10.3.
  • VinUniversity and A*STAR release the full M³OPD-SFT-1.35M corpus, the M³OPD-RL-193K verifiable-reward set, training code, and every checkpoint.

A nine-billion-parameter student, distilled from three domain experts and then pushed through a shared RL pass, averages 77.1% across 21 vision-language reasoning benchmarks, up from Qwen3.5-9B's 70.0% baseline. The recipe comes from VinUniversity and A*STAR in an open paper on Hugging Face covering image STEM, image perception and video.

The authors first train one Qwen3.5-9B teacher per domain with Group Sequence Policy Optimization, then distill all three into a single student via Multi-Teacher On-Policy Distillation. On top of that student they run a mixed-RL pass on the same prompts the mixed baseline uses. "MOPD and MixRL can therefore work as consecutive stages that get more out of the same data," the abstract reports, a staging the authors say "to our knowledge, no prior VLM recipe has studied."

Video absorbs the biggest lift at +10.7% average, with LongVideoBench climbing 12.3 points and Video-Holmes 10.3. Image STEM gains +5.1%, with EMMA jumping 14.5 points and LogicVista 12.0. Image perception moves +3.0%, a narrower gain the paper does not reframe. MOPD alone, before the added MixRL, already clears the mixed-training baseline, scoring 76.6% to MixRL's 74.9%.

Alongside the paper, the authors release the M³OPD-SFT-1.35M corpus, the M³OPD-RL-193K verifiable-reward set, the training code, and every intermediate checkpoint. It sits in a crowded autumn for distillation-style recipes our research tracker has been following.