VinUniversity-A*STAR M³OPD-Reasoner Stacks Multi-Teacher Distillation With RL, Lifts Qwen3.5-9B 6.5% Across 21 Benchmarks
Summary
Researchers from VinUniversity and A*STAR release M³OPD-Reasoner, a fully open recipe that distills three domain-expert VLM teachers (image STEM, image perception, video) into a Qwen3.5-9B student via on-policy multi-teacher distillation then tops it with mixed RL. The resulting 9B model beats mixed-training baselines by +7.1% on 9 STEM benchmarks, +10.7% on 5 video benchmarks and +6.5% overall across 21 evals, with the full M³OPD-SFT-1.35M + M³OPD-RL-193K datasets and every checkpoint released.
Originally reported by huggingface.co
Read the original article →Original headline: VinUniversity-A*STAR M³OPD-Reasoner Stacks Multi-Teacher Distillation With RL, Lifts Qwen3.5-9B 6.5% Across 21 Benchmarks