Multimodal reasoning improves with open data and tailored post-training
MMVistaReason: Toward Open-Data and Post-Training Recipes for Multimodal Reasoning
Computer Vision and Pattern Recognition
Summary
Multimodal reasoning models combine visual and analytical skills but face challenges training effectively with diverse data sources. The authors introduced MMVistaReason, a method that uses carefully cleaned open data and a special training recipe to improve model reasoning. They found that smaller models benefit from selected, high-quality training data, while larger models handle more complex, mixed data better. MMVistaReason performs well across many reasoning tests, even matching or beating bigger models using less data.
What this means in practice
- •For ai developers: Create multimodal AI systems that reason reliably with less data by following MMVistaReason’s data cleaning and training approach.
- •For product managers: Develop AI products with stronger multimodal reasoning capabilities while using fewer training resources by applying MVR’s structured recipe.$Commercial implications: Enables building competitive AI services with multimodal reasoning efficiency to reduce costs and improve performance.
Authors
Juekai Lin, Honglin Lin, Yuqian Yuan, Xiaolong Wu, Jie Cao, Liang Liang, Yunqi Cao, Yun Zhu, Wenqiao Zhang, Lijun Wu
Abstract
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. We introduce MMVistaReason (MVR), an open-data post-training recipe with three components: (1) broader capability coverage across complementary Analytical and Real-World reasoning groups, emphasizing structured reasoning versus visual perception and spatial grounding; (2) efficient SFT and RL data construction, standardizing heterogeneous open data through staged cleaning and annotation, combining difficulty-aware cascaded teacher distillation with answer-likelihood-based trajectory selection to construct MVR-SFT-528K, and applying scale-specific frontier filtering for MVR-RL-63K; and (3) specialize-then-integrate training, which trains complementary RL experts and consolidates their capabilities through multi-teacher on-policy distillation (MOPD). Our analyses reveal a capacity-dependent interaction between supervision difficulty, trajectory quality, and model capacity: smaller students benefit more from selected supervision, while larger students are robust to trajectory variation and mixed-domain interference. Mixed-domain RL introduces benchmark-level negative transfer, whereas MOPD provides consistent capability integration, with the preferred KL direction varying across model scales. Across 15 multimodal benchmarks, MVR-4B achieves an average score of 72.8, outperforming Qwen3.5-9B (Instruct) and MMFineReason-8B while using about 70% fewer samples than MMFineReason. Scaling to 9B improves the average to 74.4, surpassing Qwen3.5-35B-A3B (Instruct). Overall, MMVistaReason demonstrates that systematic open-data construction and capacity-aware post-training provide a practical and scalable path toward reliable multimodal reasoning.