GeoUG — Learning and Reasoning with Geometry

How Generation Helps Spatial Understanding. A unified model that reasons about space and can decide to generate a metric depth map as evidence mid-reasoning before answering. Point coordinates use the <point>x y</point> convention normalized to 0–1000 per axis.

📄 Paper · 💻 Code · 🤗 Checkpoint

Decision mode
adaptive: the model decides at <DECIDE> · force_depth / force_answer override the branch
Official example (GeoUG repo README)