A*STAR open-sources M3OPD: multi-teacher on-policy distillation recipe for multimodal reasoning
Nguyen Quang Trung · hf · 2026-11-18
ASTAR and VinUniversity release M³OPD-Reasoner, a fully open post-training recipe for multimodal reasoning with Multi-Teacher On-Policy Distillation (MOPD).
- A single student distills three domain teachers (image STEM, image perception, video) by learning from its own generated responses; teachers are trained with GSPO directly from Qwen3.5-9B.
- Two curated datasets: M3OPD-SFT-1.35M for SFT and M3OPD-RL-193K for RL with verifiable rewards, decontaminated against 21 benchmarks.
- The MOPD student matches each teacher on its domain, beats mixed-training baselines on average, and serves as a stronger RL starting point — adding MixRL on top still improves it without collapse. All data, code and per-stage checkpoints are open-sourced.
More from Multimodal
- Former voice actor launches Voyager 1, an AI voice you can direct line by line — Scobleizer · 2026-11-18
- Recreating Whiplash's tense rehearsal room with MiniMax H3 and ComfyUI — hucun16 · 2026-11-18
- Claude Plus Unreal Engine 5.9 MCP Builds a Game Prototype in One Day — TomLikesRobots · 2026-11-18
- Suno's new Visuals feature pairs AI-generated music with AI-generated video — The Decoder · 2026-11-18
- Osmo claims human-level AI smell: 10M labels trained world's first olfaction foundation model — MIT Tech Review AI · 2026-11-18
- AURA 2.0: a streaming video LLM that speaks first, calls tools, and grounds visually — Shuhan Li · 2026-11-18