GeoLatent: Geometry-Guided Latent Structuring with Routed Optimization for 3D Reasoning
GeoLatent reports stronger 3D reasoning by forcing visual answers to pass through structured geometry latents during training.
The paper separates spatial representation into position, direction, and global geometry, then uses CR-GEO to keep the geometry state from collapsing into one dominant direction. Its routed optimization temporarily channels answer learning through those latents before restoring full image attention. In the reported tests, CR-GEO raised geometry effective rank from 1.00 to 3.87, and blocking latent readout sharply reduced direction accuracy. GeoLatent scores 73.0% on SPAR-Bench and 72.1% on SPBench, above the previously reported methods cited by the authors. ArXiv · AI/CL/LG's note
The paper separates spatial representation into position, direction, and global geometry, then uses CR-GEO to keep the geometry state from collapsing into one dominant direction. Its routed optimization temporarily channels answer learning through those latents before restoring full image attention. In the reported tests, CR-GEO raised geometry effective rank from 1.00 to 3.87, and blocking latent readout sharply reduced direction accuracy. GeoLatent scores 73.0% on SPAR-Bench and 72.1% on SPBench, above the previously reported methods cited by the authors. ArXiv · AI/CL/LG's note
score 4