MMVistaReason: Toward Open-Data and Post-Training Recipes for Multimodal Reasoning

MMVistaReason. Toward Open-Data and Post-Training Recipes for Multimodal Reasoning. By Juekai Lin and nine others, from Zhejiang University, Ant Group, and two other institutions. Published on the twenty-second of October. Its code repository is public and real: about fifty source files.

Open multimodal reasoning models have been helped along by large-scale reasoning supervision. Post-training them reliably is another matter, and the authors name four reasons it stays hard. Open data is uneven in quality. Building supervision from it is inefficient. Difficulty is imbalanced across examples. And training on several domains at once lets them interfere with one another. None of these is about the model being too small. They are about what goes into it and in what order.

MMVistaReason, shortened to MVR, is an open-data post-training recipe built to meet those four problems with three components. The first is coverage. It splits capability into two complementary groups: Analytical reasoning, which leans on structured reasoning, and Real-World reasoning, which leans on visual perception and spatial grounding. The second is efficient data construction for both supervised fine-tuning and reinforcement learning. The third is a training order the authors call specialize-then-integrate.

On the data side, heterogeneous open data is standardized through staged cleaning and annotation. Supervised data is then built by combining difficulty-aware cascaded teacher distillation with trajectory selection based on answer likelihood, which yields MVR-SFT-528K. For reinforcement learning, a scale-specific frontier filter produces MVR-RL-63K.

On the training side, rather than mixing every domain into one run, complementary RL experts are trained separately. Their capabilities are then consolidated into one model through multi-teacher on-policy distillation, called MOPD.

The analyses find that supervision difficulty, trajectory quality and model capacity interact, and that the interaction depends on capacity. Smaller students gain more from carefully selected supervision. Larger students are robust both to variation in trajectories and to mixed-domain interference. Mixed-domain RL produced negative transfer at the level of individual benchmarks, while MOPD integrated capabilities consistently, though which KL direction worked best changed with model scale.

Across fifteen multimodal benchmarks, MVR-4B averages seventy-two point eight. That puts it ahead of Qwen3.5-9B in its Instruct version and of MMFineReason-8B, while using about seventy percent fewer samples than MMFineReason. Scaling up to nine billion parameters lifts the average to seventy-four point four, which surpasses Qwen3.5-35B-A3B, also Instruct.

What this recipe asks of anyone copying it is a decision per model size. A small student wants its supervision filtered hard, a large one can absorb messier trajectories, and even the preferred KL direction in MOPD is a per-scale setting. The authors read that as a practical and scalable path to reliable multimodal reasoning built entirely on open data, provided the post-training is tuned to the capacity of the model receiving it.