MMVistaReason: Open Recipe Trains a 4B Model That Beats Qwen3.5-9B on 15 Benchmarks

Open multimodal reasoning models have long been trained on large collections of annotated trajectories, but the resulting quality often fluctuates: the data is heterogeneous, difficulty is skewed, and domains interfere with each other. The authors of MMVistaReason (MVR) propose a unified post-training recipe built on open data and back it with numbers. The MVR-4B model averages 72.8 on 15 multimodal benchmarks and beats Qwen3.5-9B (Instruct) and MMFineReason-8B. It did so with about 70% fewer training samples than MMFineReason. The 9B version raises the average to 74.4 and outperforms Qwen3.5-35B-A3B (Instruct).

Three layers of the recipe: coverage, data, specialization

The first layer is skill coverage. The authors split capabilities into two complementary groups, Analytical and Real-World. The emphasis is on structured reasoning, not just visual perception and spatial grounding.

The second layer is data construction. Heterogeneous open datasets are converted to a single format through staged cleaning and annotation. For SFT, the team uses cascaded distillation from teachers, taking task difficulty into account. From the pool of generated trajectories, they select those with a higher answer-likelihood for the correct answer. The result is the MVR-SFT-528K corpus. For RL, the authors built a separate set, MVR-RL-63K, filtered to the "frontier" of a given model scale's abilities: tasks should be neither too easy nor clearly out of reach.

The third layer is called specialize-then-integrate. First, complementary RL experts are trained, each in its own domain. Then their skills are merged into a single model via multi-teacher on-policy distillation (MOPD).

Small models need curation; large ones can get by without it

The most interesting part of the paper is its analysis of how results depend on model capacity. Example difficulty, trajectory quality, and student size all affect one another. Small models gain noticeably from carefully curated supervision. Larger models are more robust: they tolerate variance in trajectory quality and interference between mixed domains.

Domain-mixed RL causes negative transfer at the level of individual benchmarks: a gain in one place is paid for with a drop in another. MOPD, by contrast, integrates skills stably. The preferred direction of KL divergence in distillation also changes with model scale, so there is no universal setting.

What was compared and what came out

• MVR-4B: average of 72.8 on 15 multimodal benchmarks • MVR-9B: average of 74.4 on the same 15 benchmarks • Qwen3.5-9B (Instruct): trails MVR-4B • MMFineReason-8B: trails MVR-4B, while using about 70% more samples than MVR needs • Qwen3.5-35B-A3B (Instruct): trails MVR-9B • MVR-SFT-528K: corpus for the supervised stage • MVR-RL-63K: corpus for the RL stage, filtered by capability frontier • MOPD: method for merging RL experts via multi-teacher on-policy distillation

The comparison with Qwen3.5-35B-A3B should be read with care: it is a MoE model with a small number of active parameters, and the comparison is against its Instruct version. Still, a dense 9B model trained on open data outperforms a model that is considerably larger in total parameter count. That is the authors' point.

What to take into your pipeline

For teams fine-tuning open VLMs, the paper yields several practical rules. If the model is small, invest in trajectory selection and difficulty calibration: that is where the return is highest. If it is large, you can economize on cleaning, but not on separating domains in RL.

Mixing all domains in a single RL run is risky. It is safer to train narrow experts and merge them through distillation, choosing the KL direction to fit the model size. RL tasks should be filtered for the specific model scale: a set that works well for 9B may prove useless for 4B.

The project code is published on GitHub in the JackieForest/MMVistaReason repository. The paper came out on 22 October 2026; its authors are Juekai Lin, Honglin Lin, Yuqian Yuan, Xiaolong Wu, Jie Cao, Liang Liang, Yunqi Cao, Yun Zhu, Wenqiao Zhang, and Lijun Wu. The work was featured in Hugging Face Daily Papers, where it has so far collected three upvotes, so independent verification of the results is still to come.