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

arXiv: https://arxiv.org/abs/2510.17034 · HTML: https://mlsanity.com/paper/2510.17034/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Spatial and embodied reasoning, one of 17 directions in LLM reasoning.
- Builds on 18 of its 19 references that are in the map.
- First posted on arXiv 19 Oct 2025.

## Placed in

- LLM reasoning / [Spatial and embodied reasoning](https://mlsanity.com/lr/spatial-and-embodied-reasoning/)
- Multimodal LLMs / vision-language models / [Embodied and 3D vision-language models](https://mlsanity.com/mllm/embodied-and-3d-vision-language-models/)

## Builds on (most cited first, 18 of 18)

- [VGGT: Visual Geometry Grounded Transformer](https://mlsanity.com/paper/2503.11651/) (2025)
- [OpenScene: 3D Scene Understanding with Open Vocabularies](https://mlsanity.com/paper/2211.15654/) (2022)
- [LERF: Language Embedded Radiance Fields](https://mlsanity.com/paper/2303.09553/) (2023)
- [ScanRefer: 3D Object Localization in RGB-D Scans using Natural Language](https://mlsanity.com/paper/1912.08830/) (2019)
- [LangSplat: 3D Language Gaussian Splatting](https://mlsanity.com/paper/2312.16084/) (2023)
- [ScanQA: 3D Question Answering for Spatial Scene Understanding](https://mlsanity.com/paper/2112.10482/) (2021)
- [ULIP: Learning a Unified Representation of Language, Images, and Point Clouds for 3D Understanding](https://mlsanity.com/paper/2212.05171/) (2022)
- [Scan2Cap: Context-aware Dense Captioning in RGB-D Scans](https://mlsanity.com/paper/2012.02206/) (2020)
- [ULIP-2: Towards Scalable Multimodal Pre-training for 3D Understanding](https://mlsanity.com/paper/2305.08275/) (2023)
- [Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D Understanding, Generation, and Instruction Following](https://mlsanity.com/paper/2309.00615/) (2023)
- [ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection](https://mlsanity.com/paper/2106.01178/) (2021)
- [Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors](https://mlsanity.com/paper/2505.24625/) (2025)
- [Beyond Semantics: Rediscovering Spatial Awareness in Vision-Language Models](https://mlsanity.com/paper/2503.17349/) (2025)
- [Unveiling the Mist over 3D Vision-Language Understanding: Object-centric Evaluation with Chain-of-Analysis](https://mlsanity.com/paper/2503.22420/) (2025)
- [Can 3D Vision-Language Models Truly Understand Natural Language?](https://mlsanity.com/paper/2403.14760/) (2024)
- [SURPRISE3D: A Dataset for Spatial Understanding and Reasoning in Complex 3D Scenes](https://mlsanity.com/paper/2507.07781/) (2025)
- [Does Your 3D Encoder Really Work? When Pretrain-SFT from 2D VLMs Meets 3D VLMs](https://mlsanity.com/paper/2506.05318/) (2025)
- [Revisiting 3D LLM Benchmarks: Are We Really Testing 3D Capabilities?](https://mlsanity.com/paper/2502.08503/) (2025)

## Abstract (arXiv)

Multimodal 3D grounding has garnered considerable interest in Vision-Language Models (VLMs) citeyin2025spatial for advancing spatial reasoning in complex environments. However, these models suffer from a severe "2D semantic bias" that arises from over-reliance on 2D image features for coarse localization, largely disregarding 3D geometric inputs and resulting in suboptimal fusion performance. In this paper, we propose a novel training framework called What-Where Representation Re-Forming (W2R2) to tackle this issue via disentangled representation learning and targeted shortcut suppression. Our approach fundamentally reshapes the model's internal space by designating 2D features as semantic beacons for "What" identification and 3D features as spatial anchors for "Where" localization, enabling precise 3D grounding without modifying inference architecture. Key components include a dual-objective loss function with an Alignment Loss that supervises fused predictions using adapted cross-entropy for multimodal synergy, and a Pseudo-Label Loss that penalizes overly effective 2D-dominant pseudo-outputs via a margin-based mechanism. Experiments conducted on ScanRefer and ScanQA demonstrate the effectiveness of W2R2, with significant gains in localization accuracy and robustness, particularly in cluttered outdoor scenes.