ICLR 2026Best Paper Runner-Up @ CVPR 2026 3D-LLM/VLA Workshop

Do 3D Large Language Models Really Understand 3D Spatial Relationships?

Xianzheng Ma^*, Tao Sun^*, Shuai Chen, Yash Bhalgat, Jindong Gu, Angel X Chang, Iro Armeni, Iro Laina, Songyou Peng, Victor Adrian Prisacariu

* Equal contribution

International Conference on Learning Representations (ICLR) 2026

Paper Code Project BibTeX

Real-3DQA creates spatial reasoning questions by filtering language shortcuts and varying the viewpoint of an indoor 3D scene. — Constructing Real-3DQA through question filtering and viewpoint augmentation.

Abstract

Text-only models can match or surpass 3D language models on SQA3D, suggesting that language shortcuts can obscure weaknesses in spatial reasoning. Real-3DQA filters easy-to-guess questions and introduces a taxonomy of 3D reasoning tasks. A 3D-reweighted training objective encourages models to use visual geometry and improves spatial reasoning performance.

BibTeX

@inproceedings{ma2026real3dqa,
  title={Do 3D Large Language Models Really Understand 3D Spatial Relationships?},
  author={Ma, Xianzheng and Sun, Tao and Chen, Shuai and Bhalgat, Yash and Gu, Jindong and Chang, Angel X and Armeni, Iro and Laina, Iro and Peng, Songyou and Prisacariu, Victor Adrian},
  booktitle={International Conference on Learning Representations},
  year={2026},
  pages={32591--32610},
  url={https://proceedings.iclr.cc/paper_files/paper/2026/hash/3739050312e637d158e39a2859fa8b01-Abstract-Conference.html}
}