GeoReader: Learning Latent Geometry for Spatial Reasoning in Vision-Language Models
We introduce GeoReader, a framework that learns geometric representations for the VLM that reads them.
ProofPaper ↗
Key points
- Vision-language models (VLMs) struggle with metric and multi-view relationships that remain implicit in RGB observations.
- Geometry foundation models recover this structure from RGB, but existing methods either fuse the estimated features with visual tokens and learn to use them from question answering alone, or align VLM latents with geometry features.
- Capturing geometry, however, does not ensure its use: zeroing the aligned latents of a released feature-alignment model leaves its accuracy unchanged.
- Geometry descriptions serve only as training targets: at inference, the VLM answers directly from the continuous geometric tokens without generating a geometry chain of thought.
Sources (1)
- [1]GeoReader: Learning Latent Geometry for Spatial Reasoning in Vision-Language ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 04:45 PM
We introduce GeoReader, a framework that learns geometric representations for the VLM that reads them.
Vision-language models (VLMs) struggle with metric and multi-view relationships that remain implicit in RGB observations.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Risk-Averse Inference-Time Diffusion Model Alignment with Finite-Sample Guarantees
- Oct 13, 2026Benchmarking Self-Supervised 3D Medical Vision Foundation Models and Transferability Estimation
- Oct 12, 2026UZHCL_for_the_win at Touche: Encoder Fine-Tuning, Synthetic Augmentation, and an Underpowered Encoder-LLM Comparison
- Oct 12, 2026Estimating Uncertainty of Omnimodal Large Language Models via External Omnimodal Embeddings
- Oct 12, 2026Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models
- Oct 12, 2026A Survey of Flow Matching in Reinforcement Learning