Elamur Research⌕
☰
home / papers / 2026 / CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models

CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models

Paper · 2026 · Shengli Zhou, Xiangchen Wang, Guanhua Chen et al.

Summary

LLM-based systems for 3D vision-language (3D-VL) tasks often describe a scene to the model as a graph of objects and their pairwise spatial relations. The number of relations grows quadratically with the object count, so the full graph quickly becomes too long for the model's input. The usual fix, keeping each object's few nearest neighbours as in 3DGraphLLM, often discards long-range relations the query actually needs and can split the graph into disconnected pieces.

The authors present CAPruner, a lightweight learned pruner. It scores each edge using two inputs:

  • Semantic relevance: a fuzzy match between each endpoint object's category name and the words of the task description.
  • Spatial proximity: the distance between the two objects.

Datasets mark only target objects, not individual relations. So training aggregates the scores of the edges around each node and supervises that node score with the target labels. At pruning time, every node keeps its highest-scoring edges.

On ScanRefer, ScanQA and SQA3D, LLMs given CAPruner-pruned graphs consistently beat those given proximity-pruned graphs, while using fewer tokens and adding negligible compute. The work shows that query-aware pruning is a practical way to scale scene-graph reasoning to large scenes.

Key points

  • Motivating analysis: Removing task-critical relations hurts LLM grounding accuracy on ScanRefer. Proximity-based KNN pruning leaves over 67% of ScanNet scenes disconnected.
  • Main gain: With Llama-3.2-1B, ScanRefer Acc@0.25 rose from 52.5 (proximity KNN) to 55.0. Smaller gains appeared on ScanQA and SQA3D metrics.
  • Token efficiency: At about 582 tokens on average, CAPruner beats the baseline running at about 882 tokens, a saving of roughly 34%.
  • Overhead: Pruning takes 0.75 ms per sample, compared with about 1731 ms for inference with the 8B LLM.
  • Ablations:
    • Fuzzy name matching works better than BERT embeddings and strict matching. Combined with proximity, it reaches 24.57 target-identification accuracy, versus 7.73 with no semantic cue.
    • Applying KNN selection to the learned scores beats an MST-based policy that preserves connectivity.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Southern University of Science and TechnologyShenzhen · China22 works · 1,337 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksComputer Vision and Pattern Recognition · 3,486 worksEngineering · 188 worksControl and Systems Engineering · 62 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksReasoning · 1,685 works

Cite

Cite as Shengli Zhou, Xiangchen Wang, Guanhua Chen and Feng Zheng. 2026. CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.23

@inproceedings{zhou2026capruner,
  title = {{CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models}},
  author = {Shengli Zhou, Xiangchen Wang, Guanhua Chen and Feng Zheng},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.23},
  url = {https://doi.org/10.18653/v1/2026.acl-long.23}
}

References

39 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. 2022. Scanqa: 3d question answering for spatial scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). find
  2. Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. 2025. Scene-llm: Extending language model for 3d visual reasoning. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 2195–2206. find
  3. Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. 2024. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295– 21304. find
  4. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, and 1 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  5. H. Paul Grice. 1975. Logic and conversation. Syntax and Semantics, 3:41–58. find
  6. Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, and 1 others. 2024. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conference on Robot… find
  7. Haonan Chang, Kowndinya Boyalakuntla, Shiyang Lu, Siwei Cai, Eric Jing, Shreesh Keskar, Shijie Geng, Adeeb Abbas, Lifeng Zhou, Kostas Bekris, and 1 others. 2023. Context-aware entity grounding with open-vocabulary 3d scene graphs. arXiv preprint arXiv:2309.15940. — [arXiv:2309.15940](https://arxiv.org/abs/2309.15940) find
  8. Wenxuan Guo, Xiuwei Xu, Ziwei Wang, Jianjiang Feng, Jie Zhou, and Jiwen Lu. 2025. Text-guided sparse voxel pruning for efficient 3d visual grounding. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 3666–3675. find
  9. Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. 2020. Scanrefer: 3d object localization in rgb-d scans using natural language. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 202–221. Springer. find
  10. Souhail Hadgi, Luca Moschella, Andrea Santilli, Diego Gomez, Qixing Huang, Emanuele Rodolà, Simone Melzi, and Maks Ovsjanikov. 2025. Escaping plato’s cave: Towards the alignment of 3d and text latent spaces. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19825– 19835… find
  11. Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. 2017. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE. find
  12. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  13. Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, Xiang-Dong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. 2021. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3722–3… find
  14. Daniel Honerkamp, Martin Buchner, Fabien De-spinoy, Tim Welschehold, and Abhinav Valada. 2024. Language-grounded dynamic scene graphs for interactive object search with mobile manipulation. IEEE Robotics and Automation Letters, 9:8298–8305. find
  15. Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 2023. 3d-llm: Injecting the 3d world into large language models. NeurIPS. find
  16. Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, and 1 others. 2024a. Chat-scene: Bridging 3d scene and large language models with object identifiers. Proceedings of the Advances in Neural Information Processing Systems, Va… find
  17. Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. 2024b. An embodied generalist agent in 3d world. In Proceedings of the International Conference on Machine Learning (ICML). find
  18. Language-Grounded Robot Navigation. In Proceedings of Robotics: Science and Systems, Delft, Netherlands. find
  19. Haoyuan Li, Yanpeng Zhou, Yufei Gao, Tao Tang, Jianhua Han, Yujie Yuan, Dave Zhenyu Chen, Jiawang Bian, Hang Xu, and Xiaodan Liang. 2025a. Does your 3d encoder really work? when pretrain-sft from 2d vlms meets 3d vlms. arXiv preprint arXiv:2506.05318. — [arXiv:2506.05318](https://arxiv.org/abs/2506.05318) find
  20. Zizhao Wu, Haohan Li, Gongyi Chen, Zhou Yu, Xiaoling Gu, and Yigang Wang. 2024. 3d question answering with scene graph reasoning. In ACM Multimedia 2024. find
  21. Rong Li, Shijie Li, Lingdong Kong, Xulei Yang, and Junwei Liang. 2025b. Zero-shot 3d visual grounding from vision-language models. ArXiv, abs/2505.22429. — [arXiv:2505.22429](https://arxiv.org/abs/2505.22429) find
  22. Sergey Linok, Tatiana Zemskova, Svetlana Ladanova, Roman Titkov, Dmitry A. Yudin, Maxim Monastyrny, and Aleksei Valenkov. 2024. Beyond bare queries: Open-vocabulary object grounding with 3d scene graph. 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 13582– 13589. find
  23. Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. 2023. Sqa3d: Situated question answering in 3d scenes. In International Conference on Learning Representations. find
  24. Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. 2012. Indoor segmentation and support inference from rgbd images. In ECCV. find
  25. Maxime Oquab, Timothée Darcet, Théo Moutakanni, and 1 others. 2024. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research. Featured Certification.
  26. Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. 2023. Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. International Conference on Robotics and Automation (ICRA). find
  27. Manycore Tech Inc. SpatialVerse Research Team. 2025. Interiorgs: A 3d gaussian splatting dataset of semantically labeled indoor scenes. https://huggingface.co/datasets/spatialverse/InteriorGS. ↗
  28. Xinyi Wang, Na Zhao, Zhiyuan Han, Dan Guo, and Xun Yang. 2025. Augrefer: Advancing 3d visual grounding via cross-modal augmentation and spatial relation-based referring. CoRR, abs/2501.09428. — [arXiv:2501.09428](https://arxiv.org/abs/2501.09428) find
  29. Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. 2023. Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21560– 21569. find
  30. Can Xu, Yuehui Han, Rui Xu, Le Hui, Jin Xie, and Jian Yang. 2024. Multi attributes interactions matters for 3d visual grounding. In CVPR. find
  31. Hanxun Yu, Wentong Li, Song Wang, Junbo Chen, and Jianke Zhu. 2025. Inst3d-lmm: Instance-aware 3d scene understanding with multi-modal instruction tuning. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14147– 14157. find
  32. Qihao Yuan, Kailai Li, and Jiaming Zhang. 2025. Solving zero-shot 3d visual grounding as constraint satisfaction problems. In 36th British Machine Vision Conference 2025, BMVC 2025, Sheffield, UK, November 24-27, 2025. BMVA. find
  33. Tatiana Zemskova and Dmitry Yudin. 2024. 3dgraphllm: Combining semantic graphs and large language models for 3d scene understanding. Preprint, arXiv:2412.18450. — [arXiv:2412.18450](https://arxiv.org/abs/2412.18450) find
  34. Guangyao Zhai, Evin Pınar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. 2024. Commonscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36. find
  35. Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. 2024. Uni3d: Exploring unified 3d representation at scale. In International Conference on Learning Representations (ICLR). find
  36. Shengli Zhou, Minghang Zheng, Feng Zheng, and Yang Liu. 2026. Scalable object relation encoding for better 3d spatial reasoning in large language models. Preprint, arXiv:2603.24721. — [arXiv:2603.24721](https://arxiv.org/abs/2603.24721) find
  37. Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 2023. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2911– 2921. find
  38. Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. 2025. Unifying 3d vision-language understanding via promptable queries. In Computer Vision – ECCV 2024, pages 188–206, Cham. Springer Nature Switzerland. find
  39. Abdelrhman Werby, Chenguang Huang, Martin Büch-ner, Abhinav Valada, and Wolfram Burgard. 2024. Hierarchical Open-Vocabulary 3D Scene Graphs for find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2026-capruner-conceptual-adjacent-scene-graph/