MindCube: Spatial Mental Modeling from Limited Views | VLM Spatial Reasoning Benchmark

MindCube: a spatial reasoning benchmark of 21,154 questions on 3,268 multi-view images, testing whether vision-language models build spatial mental models of unseen space and reason from limited views. VLMs are near-random; map-then-reason training lifts accuracy from 37.8% to 57.8%, and to 61.3% w…

关键词：MindCube, spatial mental models, spatial mental modeling, spatial reasoning, vision-language models, VLM, benchmark, cognitive map, perspective taking, mental simulation, map-then-reason, multi-view,…