Where, Not What: Compelling Video LLMs to Learn Geometric Causality for 3D-Grounding

Multimodal 3D grounding has attracted increasing interest in vision language models (VLMs) [1] for advancing spatial reasoning in complex environments, yet existing approaches often suffer from a strong 2D semantic bias. Such models tend to rely on appearance cues for coarse localization while underutilizing reconstructed 3D geometric information, leading to suboptimal grounding performance. We propose What Where Representation Reforming (W2R2), a training framework that reshapes internal representations without modifying the inference time architecture. W2R2 explicitly designates 2D features as semantic beacons for identifying "What" and 3D features as spatial anchors for localizing "Where", and employs a dual-objective pull push loss to suppress overly effective 2D shortcuts while preserving semantic alignment. Experiments on ScanRefer, ScanQA and Scan2Cap demonstrate consistent improvements in 3D grounding accuracy and robustness, particularly in cluttered scenes.