2026-10-26 · America/Los_Angeles · 论文 · #14

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

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…

热度 41.6 / 100；排名与评分保留该期记录。