GeoLatent: Geometry-Guided Latent Structuring with Routed Optimization for 3D Reasoning
Abstract Overview
GeoLatent addresses 3D spatial reasoning in vision-language models by improving decomposed latent representations for position, direction, and global geometry. The paper identifies two limitations in prior geometry-supervised latent methods: global geometry states can collapse into redundant representations, and downstream answer prediction may bypass the latents through direct access to image tokens. To address these issues, the authors introduce Common-Residual Geometry Alignment (CR-GEO), which separates shared geometry from residual variation, and a routed optimization schedule that temporarily forces answer-relevant visual information through the latent pathway before restoring full attention. The study evaluates both benchmark accuracy and mechanistic evidence about representation diversity, latent usage, and visual grounding.
Novelty
The main novelty is the combination of CR-GEO and routed optimization within a decomposed spatial latent framework for 3D reasoning. Rather than only adding geometric supervision, the method explicitly targets redundancy in global geometry latents and tests whether latent intermediates are actually used during answer learning.
Results
GeoLatent increases the effective rank of the geometry representation from 1.00 to 3.87 in matched comparisons, indicating substantially less redundancy. It achieves 73.0% on SPAR-Bench and 72.1% on SPBench, exceeding the reported GeoAnchor results on both benchmarks. Intervention results also show strong bottleneck-stage dependence on latent readout, while recovered models retain a usable latent-mediated route and remain strongly dependent on image evidence.
Key Points
- CR-GEO restructures global geometry supervision by aligning shared geometry at the set level and encouraging individual GEO states to encode complementary residual information.
- Routed optimization uses joint, bottleneck, and recovery stages to promote latent-mediated answer learning without permanently removing direct image access.
- The paper supports its claims with both benchmark gains and diagnostic analyses, including effective-rank measurements, latent-readout interventions, and image blanking or mismatching tests.