Elamur Research⌕
☰
home / papers / 2026 / GeoGround: Uncertainty-Weighted Multi-Task Learning for Geo-Alignment and Address Defect Detection

GeoGround: Uncertainty-Weighted Multi-Task Learning for Geo-Alignment and Address Defect Detection

Paper · 2026 · Srinivas Virinchi, Aman Gulati and Anoop Saladi

Summary

In e-commerce logistics, free-form customer addresses must be converted into coordinates, and addresses likely to cause delivery problems such as loss, damage, or fraud must be flagged. Both tasks have to run in under 50 ms. Earlier systems handled them separately, even though a precise location is a strong signal for defect risk. GeoGround, from Amazon, trains one compact encoder (all-MiniLM-L6-v2) for both tasks. It combines four training signals:

  • Address–coordinate alignment: an InfoNCE objective matches address text to coordinates encoded with the Equal Earth projection and Random Fourier Features.
  • Hierarchical contrast: a text–text objective draws hard negatives from H3 hexagonal cells at resolutions 9–11.
  • Defect classification: Focal Loss handles defects, which make up fewer than 5% of labels.
  • Loss balancing: uncertainty-based weighting adjusts the tasks against each other.

The training data is augmented with LLM-generated noisy, incomplete, and obfuscated addresses. Compared with a RoBERTa baseline, GeoGround improves defect precision 5.86× and location accuracy up to 4.86×. It also runs about 75× faster than Qwen2-1.5B. The work shows that a joint, spatially grounded encoder can outperform decoder LLMs while staying within real-time serving limits.

Key points

  • Accuracy: GeoGround-MTL reaches 5.863× defect Prec@x1% and 4.863× location Prec@y2% relative to Roberta-General. It beats AddressBind, Llama-3.2-1B and Qwen2-1.5B, and has the best silhouette score (7.973×).
  • Ablations: Dropping the geolocation objective cuts defect precision from 5.86× to 2.08×. Removing Focal Loss or data augmentation also causes large drops.
  • Speed: On one NVIDIA L4 (batch size 64), latency is about 9.47 ms per batch, compared with 719.10 ms for Qwen2-1.5B.
  • Remaining errors:
    • about 45% come from missing unit or suite details;
    • about 35% come from cold-start (newly built) addresses;
    • about 20% come from extreme transliteration.
  • Production impact: A two-week A/B test showed a +50 bps gain in defect detection and +40 bps in location prediction. The estimated savings are $3.09M per year.

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

🏢AmazonSeattle · United States112 works · 40,565 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksDecision Sciences · 209 worksManagement Science and Operations Research · 108 works

Cite

Cite as Srinivas Virinchi, Aman Gulati and Anoop Saladi. 2026. GeoGround: Uncertainty-Weighted Multi-Task Learning for Geo-Alignment and Address Defect Detection. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.27

@inproceedings{virinchi2026geoground,
  title = {{GeoGround: Uncertainty-Weighted Multi-Task Learning for Geo-Alignment and Address Defect Detection}},
  author = {Srinivas Virinchi, Aman Gulati and Anoop Saladi},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.27},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.27}
}

References

28 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Isaac Brodsky. 2018a. H3: Uber’s hexagonal hierarchical spatial index. https://www.uber.com/en-SE/ blog/h3/. ↗
  2. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  3. Govind, Sayan Putatunda, and Saurabh Sohoney. 2025. Addressbind: Cross-modal alignment of addresses and geocodes for last-mile transportation systems. In Proceedings of the 33rd ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. ACM. find
  4. Govind and Saurabh Sohoney. 2022. Learning geolocations for cold-start and hard-to-resolve addresses via deep metric learning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing: Industry Track (EMNLP). find
  5. William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems.
  6. James Inman. 1835. Navigation and nautical astronomy: For the use of british seamen. find
  7. Isaac Brodsky. 2018b. H3: Uber’s hexagonal hierarchical spatial index. https://www.uber.com/blog/ h3/. Accessed: 2026-02-14. ↗
  8. Alex Kendall, Yarin Gal, and Roberto Cipolla. 2018. Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482–7491. find
  9. Isaac Brodsky, David Crain, Gergely Fekete, Hao Li, Nick Mattingly, and Ben Worsham. 2018. H3: Konstantin Klemmer, Nathan S. Safir, and Daniel B. Neill. 2023. Positional encoder graph neural networks for geographic data. In Proceedings of AIS-TATS 2023, volume 206, pages 1379–1389. find
  10. Govind Kothari and Saurabh Sohoney. 2022. Geospatially informed models for geocoding unstructured addresses. arXiv preprint arXiv:2203.01234. — [arXiv:2203.01234](https://arxiv.org/abs/2203.01234) find
  11. Uddeshya Singh, Gowtham Bellala, Ravi Shankar De-vanapalli, and Vikas Goel. 2025. Geo-spatially informed models for geocoding unstructured addresses. In Proceedings of the 31st International Conference on Computational Linguistics: Industry Track, pages 236–242. find
  12. Aaron Van den Oord, Yazhe Li, and Igor Babakhin. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748.
  13. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. Proceedings of the IEEE International Conference on Computer Vision (ICCV).
  14. Saining Liu and et al. 2019. End-to-end multi-task learning with attention. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). find
  15. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Li, Sanqiang Ko, Justin Mahoney, Yangfu Chen, Emily Glass, Xiaodong Le, Veselin Stoyanov, Jacob Eisenstein, Anca Rus, Mohit Vijayanarasimhan, Fei Deng, Ali Ghazvininejad, Mike Guler, Scott Johnson, Luke Lewis, and Luke Zettlemoyer. 2019. Roberta:…
  16. Guy Macdonald Morton. 1966. A computer oriented geodetic data base and a new technique in file sequencing. Technical report, IBM Ltd. Ottawa, Canada. find
  17. Vamsi Krishna Penumadu, Nitesh Methani, and Saurabh Sohoney. 2022. Learning geospatially aware place embeddings via weak supervision. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM), pages 4071–4080. find
  18. Hui Qian and et al. 2021. Seq2seq geocoding for chinese addresses with hierarchical attention. In Proceedings of the IEEE International Conference on Data Mining (ICDM). find
  19. Ali Rahimi and Benjamin Recht. 2007. Random features for large-scale kernel machines. Proceedings of the Neural Information Processing Systems (NeurIPS). find
  20. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982– 3992,…
  21. Sebastian Ruder. 2017. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098.
  22. David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. 1986. Learning representations by backpropagating errors. nature, 323(6088):533–536. find
  23. Eric Veach. 2017. S2 geometry library. https://s2geometry.io/. ↗
  24. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672.
  25. Jiachen Wu and et al. 2022. Exploring vision-language models for geographic location prediction. In Proceedings of the European Conference on Computer Vision (ECCV). find
  26. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, and . . . et al. 2024. Qwen2 technical report. arXiv preprint. find
  27. Jianing Yuan and et al. 2021. Place representation learning with graph neural networks for warm-start geocoding. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD). find
  28. Bojan Šavrič, Tom Patterson, and Bernhard Jenny. 2019. The equal earth map projection. International Journal of Geographical Information Science, 33(3):454– 465. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/virinchi-2026-geoground-uncertainty-weighted-multi-task/