frontier models
The Spatial Model Dismisses Its 3D Encoder at Runtime
GeoThinker-F learns predictive geometric and future-semantic latents from an offline expert, then removes that expert for inference.
Summary
GeoThinker-F learns predictive geometric and future-semantic latents from an offline expert, then removes that expert for inference.
External 3D encoders provide useful geometry but add a second runtime path. GeoThinker-F first trains lightweight heads to predict geometry-expert features and future semantics while the main multimodal model is frozen. It then removes the expert and teaches the backbone to fuse its own predictions. The authors report up to a 3.6-times speedup, cutting latency from 1.45 seconds to 0.40 seconds while retaining competitive performance across their spatial benchmarks. The efficiency claim is tied to the evaluated stack, but the handoff from teacher to internal representation is explicit.
Why it matters
GeoThinker-F learns predictive geometric and future-semantic latents from an offline expert, then removes that expert for inference.
Limits and context
No additional limitation was separately recorded.
Key claims
GeoThinker-F learns predictive geometric and future-semantic latents from an offline expert, then removes that expert for inference.
Evidence: source-2026-10-15-010
Sources
- arXiv preprint 2610.20592, original v1arXiv / submitting research authors · primary research
Corrections
: Publication note: This October 15 edition was prepared before its scheduled 9:00 AM America/Detroit reveal, but publication was delayed by a production build failure. The original reporting and image-generation timestamps are preserved.