Elamur Research⌕
☰
home / papers / 2026 / Can LLMs Learn to Map the World from Local Descriptions?

Can LLMs Learn to Map the World from Local Descriptions?

Paper · 2026 · Sirui Xia, Aili Chen, Xintao Wang et al.

Summary

This paper asks whether a language model can assemble a consistent global map when it only ever sees piecemeal, relative statements about places, much as people build mental maps from directions. The authors built a synthetic 100 × 100 km grid city with weighted roads and randomly scattered POIs. They then continually pre-trained Qwen2.5-0.5B in two stages. The first stage used templated sentences giving the distance and azimuth between POI pairs. The second stage used natural-language route instructions derived from shortest paths. Evaluation had three parts: explicit predictions for POI pairs never seen in training, probes of hidden states for coordinate information, and perturbation tests during navigation. The model generalised to unseen relations and planned routes between previously unconnected locations. Its internal representations encoded absolute positions and tracked its location along a route. However, it recovered poorly from route disruptions, especially at turning points that were rare in training. This suggests its road knowledge is patchy rather than truly global. The work matters because it shows structured spatial knowledge can emerge from text alone, while also exposing where that knowledge breaks down.

Key points

  • Environment: a 100 × 100 grid, with each unit equal to 1 km. Roads run along every integer line, with traversal weights drawn from [0.8, 1.2]. There are 1024 anonymised POIs.
  • Training data: two templated datasets with varied phrasing, used in two-stage continual pre-training (CPT) on Qwen2.5-0.5B:
    • a Relational Spatial Dataset of Euclidean distance and azimuth between POI pairs;
    • a Trajectory Dataset of Dijkstra shortest paths written as step-by-step instructions.
  • Perception results: distance and angle consistency on unseen pairs was measured with Spearman and Pearson correlations under 8:2, 6:4 and 4:6 train/test splits. The reported values range from about 0.65 to 0.91.
  • Robustness findings: recovery after a route perturbation improved when the new starting point was a frequently seen turning point. The authors read this as evidence of localised rather than comprehensive road-network knowledge.
  • Limitations and resources:
    • The training caused some catastrophic forgetting of general language ability.
    • Real-world POI experiments are given in Appendix C.3, and model size and architecture comparisons in Appendix E.2.
    • Code and data are available at github.com/pdxthree/Map.

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

🎓Fudan UniversityShanghai · China202 works · 22,509 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

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

Cite

Cite as Sirui Xia, Aili Chen, Xintao Wang, Tinghui Zhu, Yikai Zhang, Jiangjie Chen and Yanghua Xiao. 2026. Can LLMs Learn to Map the World from Local Descriptions?. 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.128

@inproceedings{xia2026can,
  title = {{Can LLMs Learn to Map the World from Local Descriptions?}},
  author = {Sirui Xia, Aili Chen, Xintao Wang, Tinghui Zhu, Yikai Zhang, Jiangjie Chen and Yanghua Xiao},
  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.128},
  url = {https://doi.org/10.18653/v1/2026.acl-long.128}
}

References

29 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  2. AI@Meta. 2024. Llama 3 model card. find
  3. Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.
  4. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
  5. E. W. Dijkstra. 1959. A note on two problems in connexion with graphs. Numerische Mathematik, 1(1). find
  6. Jie Feng, Yuwei Du, Tianhui Liu, Siqi Guo, Yuming Lin, and Yong Li. 2024a. Citygpt: Empowering urban spatial cognition of large language models. arXiv preprint arXiv:2406.13948. — [arXiv:2406.13948](https://arxiv.org/abs/2406.13948) find
  7. Jie Feng, Jun Zhang, Junbo Yan, Xin Zhang, Tianjian Ouyang, Tianhui Liu, Yuwei Du, Siqi Guo, and Yong Li. 2024b. Citybench: Evaluating the capabilities of large language model as world model. CoRR. find
  8. Wes Gurnee and Max Tegmark. Language models represent space and time. In The Twelfth International Conference on Learning Representations. Dean Hazineh, Zechen Zhang, and Jeffrey Chiu. Linear latent world models in simple transformers: A case study on othello-gpt. In Socially Responsible Language Mo… find
  9. Charles Jin and Martin Rinard. 2024. Emergent representations of program semantics in language models trained on programs. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 22160–22184. PMLR. find
  10. Nicolas Martorell. 2025. From text to space: Mapping abstract spatial models in llms during a grid-world navigation task. find
  11. Roshanak Mirzaee, Hossein Rajaby Faghihi, Qiang Ning, and Parisa Kordjamshidi. 2021. SPARTQA: A textual question answering benchmark for spatial reasoning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologie…
  12. Ida Momennejad, Hosein Hasanbeig, Felipe Vieira Frujeri, Hiteshi Sharma, Nebojsa Jojic, Hamid Palangi, Robert Ness, and Jonathan Larson. 2023. Evaluating cognitive maps and planning in large language models with cogeval. Advances in Neural Information Processing Systems, 36:69736–69751. find
  13. Neel Nanda, Andrew Lee, and Martin Wattenberg. 2023. Emergent linear representations in world models of self-supervised sequence models. In Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 16–30, Singapore. Association for Computational Linguisti…
  14. Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. 2011. Scikit-learn: Machine learning in python. the Journal of machine Learning research, 12:2825–2830. find
  15. Santhosh Kumar Ramakrishnan, Erik Wijmans, Philipp Kraehenbuehl, and Vladlen Koltun. 2024. Does spatial cognition emerge in frontier models? arXiv preprint arXiv:2410.06468. — [arXiv:2410.06468](https://arxiv.org/abs/2410.06468) find
  16. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199– 22213.
  17. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  18. Kenneth Li, Aspen K. Hopkins, David Bau, Fernanda B. Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Emergent world representations: Exploring a sequence model trained on a synthetic task. In Proceedings of the 11th International Conference on Learning Representations (ICLR). find
  19. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  20. Wenbin Li, Di Yao, Ruibo Zhao, Wenjie Chen, Zijie Xu, Chengxue Luo, Chang Gong, Quanliang Jing, Haining Tan, and Jingping Bi. 2024. Stbench: Assessing the ability of large language models in spatiotemporal analysis. arXiv preprint arXiv:2406.19065. — [arXiv:2406.19065](https://arxiv.org/abs/2406.19065) find
  21. Johannes Treutlein, Dami Choi, Jan Betley, Samuel Marks, Cem Anil, Roger B Grosse, and Owain Evans. 2024. Connecting the dots: Llms can infer and verbalize latent structure from disparate training data. Advances in Neural Information Processing Systems, 37:140667–140730. find
  22. Bastien Liétard, Mostafa Abdou, and Anders Søgaard. 2021. Do language models know the way to rome? In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 510–517. find
  23. Keyon Vafa, Justin Chen, Ashesh Rambachan, Jon Kleinberg, and Sendhil Mullainathan. 2024. Evaluating the world model implicit in a generative model. Advances in Neural Information Processing Systems, 37:26941–26975. find
  24. Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. 2024. Mind’s eye of llms: Visualization-of-thought elicits spatial reasoning in large language models. In Neural Information Processing Systems. find
  25. An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, et al. 2024a. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122. — Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement
  26. Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. 2024b. Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171.
  27. Qwen An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxin Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei… find
  28. Dazhou Yu, Riyang Bao, Gengchen Mai, and Liang Zhao. 2025. Spatial-rag: Spatial retrieval augmented generation for real-world spatial reasoning questions. arXiv preprint arXiv:2502.18470. — [arXiv:2502.18470](https://arxiv.org/abs/2502.18470) find
  29. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400–410, Bangkok, Thail…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/xia-2026-can-llms-learn-map-world/