TL;DR of the paper, generated automatically.

Research paper: a VLM spatial-reasoning method that runs scenes in a physics engine.

• Method: a Perceiver builds a scene in MuJoCo, and a Reasoner tests viewpoint changes there.
• MindCube: Qwen3.5-27B goes from 52.70% to 71.14%, 7.52 points past pySpatial.
• Ground truth: with exact object and camera poses, it hits 98.25% versus 71.81% for model reasoning alone.