Thinking with Geometry: Active Geometry Integration for Spatial Reasoning

Recent progress in spatial reasoning with Multimodal Large Language Models (MLLMs) increasingly leverages geometric priors from 3D encoders. However, most existing integration strategies remain passive: geometry is exposed as a global stream and fused in an indiscriminate manner, which often induces semantic-geometry misalignment and redundant signals. We propose GeoThinker, a framework that shifts the paradigm from passive fusion to active perception. Instead of feature mixing, GeoThinker enables the model to selectively retrieve geometric evidence conditioned on its internal reasoning demands. GeoThinker achieves this through Spatial-Grounded Fusion applied at carefully selected VLM layers, where semantic visual priors selectively query and integrate task-relevant geometry via frame-strict cross-attention, further calibrated by Importance Gating that biases per-frame attention toward task-relevant structures. Comprehensive evaluation results show that GeoThinker sets a new state-of-the-art in spatial intelligence, achieving a peak score of 72.6 on the VSI-Bench. Furthermore, GeoThinker demonstrates robust generalization and significantly improved spatial perception across complex downstream scenarios, including embodied referring and autonomous driving. Our results indicate that the ability to actively integrate spatial structures is essential for next-generation spatial intelligence.

Artificial intelligence systems are increasingly capable of understanding the physical world by combining visual images with 3D structural data. However, most current systems process this 3D information passively, trying to absorb every detail of an environment at once. This often overwhelms the AI with irrelevant information and causes confusion. To solve this, we introduce GeoThinker. Instead of blindly accepting all 3D data, GeoThinker actively searches for the specific geometric clues it needs based on the task at hand. It filters out distractions and focuses only on the physical structures that actually matter. Our tests show that this selective approach makes GeoThinker significantly better at understanding complex spaces, achieving record-breaking performance. It also proves highly effective in real-world scenarios, such as helping autonomous cars navigate or assisting robots in interacting with their surroundings. Our findings suggest that teaching AI to actively choose what to focus on is a crucial step toward advanced spatial intelligence.