MMVistaReason open recipe trains multimodal reasoners with less data
Original title: MMVistaReason: Toward Open-Data and Post-Training Recipes for Multimodal Reasoning
MVR-4B averages 72.8 on 15 benchmarks, beating Qwen3.5-9B Instruct and MMFineReason-8B with about 70% fewer samples.
The paper introduces MMVistaReason, an open-data post-training recipe for multimodal reasoning. It builds MVR-SFT-528K and MVR-RL-63K, trains complementary reinforcement-learning experts, and merges them with multi-teacher on-policy distillation. Scaling to 9B raises the average to 74.4, above Qwen3.5-35B-A3B Instruct.
- SFT data use difficulty-aware distillation and answer-likelihood trajectory selection.
- Mixed-domain RL shows benchmark-level negative transfer; MOPD integrates capabilities more consistently.
- Smaller students gain more from selected supervision than larger ones.
- The preferred KL direction in distillation varies by model scale.
From the original article
1 min read at huggingface.co
Open multimodal reasoning models have benefited from large-scale reasoning supervision, yet reliable post-training remains challenging due to uneven data quality, inefficient supervision construction, imbalanced difficulty, and cross-domain interference.
By Juekai Lin, Honglin Lin, Yuqian Yuan · 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.