Elamur Research⌕
☰
home / papers / 2026 / CANDICE: Agentic Causal Disentanglement with Class Conditional Knowledge Integration for Long Tailed Domain Generalization

CANDICE: Agentic Causal Disentanglement with Class Conditional Knowledge Integration for Long Tailed Domain Generalization

Paper · 2026 · Midhat Urooj, Ayan Banerjee and Sandeep Gupta

Summary

Medical imaging classifiers must generalize across hospitals, scanners and patient groups (domain generalization, DG). They must also recognize rare pathologies (long-tailed recognition, LT). The authors argue these goals conflict: enforcing domain invariance tends to erase the fine-grained cues that rare classes need.

Their proposed framework, CANDICE, treats diagnosis as building an LLM-driven pipeline rather than training one monolithic predictor. Three agents divide the work:

  • a Clinical/Conceptual Reasoning Agent (CRA) gathers diagnostic rules and biomarkers from sources such as PubMed;
  • a Causal Disentanglement Agent (CDA) plans which factors and models are usable;
  • a Code Generation Agent (CGA) writes and self-debugs executable extraction code.

A decision-level learner, CoCKTAIL, then combines knowledge-derived and deep-learning features class by class. Head classes rely mainly on data-driven signals, and tail classes rely on knowledge-driven ones.

Across 10 datasets in four modalities, CANDICE improves performance by about 10.3% on average. It also yields auditable reasoning traces. The work suggests that the robustness–rare-class tension can be handled by orchestrating decisions rather than redesigning representations.

Key points

  • Three agents (CRA, CDA, CGA) feed the CoCKTAIL integrator, which uses Expected Information Gain and a Gini/purity index to guide tree-based, class-conditional partitioning.
  • Four applications were tested: rs-fMRI seizure onset zone (SOZ) localization, fundus diabetic retinopathy grading, echocardiographic LVEF estimation (EchoNet-Dynamic), and exercise-stress ECG CAD detection (Mayo MISC data).
  • For rare-class SOZ localization, accuracy rose by 38.5 points. Expert review effort fell by up to 84% for SOZ and by an estimated 70% for LVEF.
  • For CAD detection (726 training subjects, 227 test subjects), CANDICE reached PPV/NPV/AUC of 91.2/93.0/92.2. The ViT baseline scored 79.0/81.8/82.0.
  • On EchoNet-Dynamic, CANDICE fell slightly short of the optimized original pipeline (HF AUC 0.94 vs 0.97). The stated limitations are added inference latency, dependence on knowledge quality, and evaluation restricted to medical tasks.

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

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Midhat Urooj, Ayan Banerjee and Sandeep Gupta. 2026. CANDICE: Agentic Causal Disentanglement with Class Conditional Knowledge Integration for Long Tailed Domain Generalization. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.2018

@inproceedings{urooj2026candice,
  title = {{CANDICE: Agentic Causal Disentanglement with Class Conditional Knowledge Integration for Long Tailed Domain Generalization}},
  author = {Midhat Urooj, Ayan Banerjee and Sandeep Gupta},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2018},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2018}
}

References

60 references, 15 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Stanford AIMI. 2020. Echonet-dynamic dataset. https://stanfordaimi.azurewebsites.net/datasets/ 834e1cd1-92f7-4268-9daa-d359198b310a. Accessed: 2025-11-11. ↗
  2. Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. 2016. Neural module networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 39–48.
  3. Asia Pacific Tele-Ophthalmology Society. 2019. find
  4. AP-TOS 2019 blindness detection. Kaggle competition. find
  5. Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. 2010. A theory of learning from different domains. Machine Learning, 79(1–2):151–175. find
  6. Carlo Biffi, Ozan Oktay, Giacomo Tarroni, Wenjia Bai, Antonio De Marvao, Georgia Doumou, Martin Rajchl, Reem Bedair, Sanjay Prasad, Stuart Cook, Declan O’Regan, and Daniel Rueckert. 2018. Learning interpretable anatomical features through deep generative models: Application to cardiac remodeling. In… find
  7. Leo Breiman, Jerome H. Friedman, Richard A. Olshen, and Charles J. Stone. 1984. Classification and Regression Trees. Wadsworth International Group, Belmont, CA. find
  8. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, and 1 others. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
  9. Xinyun Chen, Maxwell Lin, Nathanael Schärli, and Denny Zhou. 2024. Teaching large language models to self-debug. In Proceedings of the International Conference on Learning Representations (ICLR). find
  10. Jorge Cuadros and George Bresnick. 2009. Eye-PACS: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509–516. find
  11. Etienne Decencière, Xiwei Zhang, Guy Cazuguel, Bruno Laÿ, Béatrice Cochener, Caroline Trone, Philippe Gain, John-Richard Ordóñez-Varela, Pascale Massin, Ali Erginay, Béatrice Charton, and Jean-Claude Klein. 2014. Feedback on a publicly distributed image database: The Messidor database. Image Analysi… find
  12. Early Treatment Diabetic Retinopathy Study Research Group. 1991. Grading diabetic retinopathy from stereoscopic color fundus photographs – an extension of the modified airlie house classification: ETDRS report number 10. Ophthalmology, 98(5 Suppl):786– 806. find
  13. American Association for Pediatric Ophthalmology and Strabismus. 2023. Proliferative diabetic retinopathy. https://aapos.org/glossary/ proliferative-diabetic-retinopathy. ↗
  14. Robert N. Frank. 2004. Diabetic retinopathy. New England Journal of Medicine, 350(1):48–58. find
  15. Corrado Gini. 1912. Variabilità e Mutabilità (Variability and Mutability). Tipografia di Paolo Cuppini, Bologna, Italy. find
  16. Xiao Gu, Yao Guo, Zeju Li, Jianing Qiu, Qi Dou, Yuxuan Liu, Benny Lo, and Guang-Zhong Yang. 2022. Tackling long-tailed category distribution under domain shifts. arXiv preprint arXiv:2207.10150. — [arXiv:2207.10150](https://arxiv.org/abs/2207.10150) find
  17. Rahul Gupta, Soham Pal, Aditya Kanade, and Shirish Shevade. 2017. DeepFix: Fixing common C language errors by deep learning. In Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI), pages 1345–1351. find
  18. Ziwei Ji, Nayeon Lee, Rita Frieske, Tianyu Yu, and 1 others. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1–38.
  19. Payal Kamboj, Ankit Banerjee, Varina L. Boerwinkle, and S. K. S. Gupta. 2024. The expert’s knowledge combined with ai outperforms ai alone in seizure onset zone localization using resting-state fmri. Frontiers in Neurology, 14:1324461. find
  20. Nan Rosemary Ke, Anirudh Goyal, Stefan Bauer, Hugo Larochelle, and Yoshua Bengio. 2019. Learning neural causal models from unknown interventions. Advances in Neural Information Processing Systems. find
  21. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 399…
  22. Tushar Khot. 2023. Adapt: As-needed decomposition and planning with language models. arXiv preprint arXiv:2311.05772.
  23. StatPearls Publishing. 2024. Diabetic retinopathy. https://www.ncbi.nlm.nih.gov/books/ NBK560805/. ↗
  24. Jing Lei, Yaping Huang, Jundong Gao, and Hao Chen. 2022. Cross-domain deep learning in medical imaging: A survey. IEEE Transactions on Medical Imaging, 41(10):2732–2749. find
  25. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, and 1 others. 2023. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789.
  26. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems (NeurIPS).
  27. J. Ross Quinlan. 1993. C4.5: Programs for Machine Learning. Morgan Kaufmann, San Mateo, CA.
  28. Lusi Li, Haibo He, and Jie Li. 2020. Entropy-based sampling approaches for multi-class imbalanced problems. IEEE Transactions on Knowledge and Data Engineering, 32(11):2159–2170. find
  29. Xiaoxiao Li, Nicha C. Dvornek, Juntang Zhuang, Pamela Ventola, and James S. Duncan. 2018. Brain biomarker interpretation in ASD using deep learning and fMRI. In MICCAI 2018, volume 11072 of LNCS, pages 206–214. find
  30. Mingquan Lin, Gregory Holste, Song Wang, Yiliang Zhou, Yishu Wei, Imon Banerjee, Pengyi Chen, Tianjie Dai, Yuexi Du, Nicha C. Dvornek, Yuyan Ge, Zuwei Guo, Shouhei Hanaoka, Dongkyun Kim, Pablo Messina, Yang Lu, Denis Parra, Donghyun Son, Álvaro Soto, and 14 others. 2025. find
  31. CXR-LT 2024: A MICCAI challenge on long-tailed, multi-label, and zero-shot disease classification from chest X-ray. Medical Image Analysis, 106:103739. find
  32. Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. 2016. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779–788. find
  33. Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418–5426. find
  34. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 9351:234–241.
  35. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761.
  36. Wei-Yin Loh. 2011. Classification and regression trees. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 1(1):14–23. find
  37. Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. 2021. Toward causal representation learning. Proceedings of the IEEE. find
  38. American Academy of Ophthalmology. 2023. Diabetic retinopathy preferred practice pattern, 2023. https://www.aao.org/preferred-practice-pattern/ diabetic-retinopathy-ppp. ↗
  39. Claude Elwood Shannon. 1948. A mathematical theory of communication. Bell System Technical Journal, 27(3–4):379–423, 623–656. find
  40. Daniel Ouyang, Benjamin J He, Amirreza Ghorbani, Lan Xia, Hua Li, Muhammad Alfakir, Lisa Hou, Roy K Chen, Erqou Ding, Aaron D. Aguirre, and 1 others. 2020. Video-based AI for beat-to-beat assessment of cardiac function. Nature, 580(7802):252– 256. find
  41. Judea Pearl. 2009. Causality: Models, Reasoning, and Inference. Cambridge University Press. find
  42. Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. 2017. Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press. find
  43. Archiki Prasad, Alexander Koller, Mareike Hartmann, Peter Clark, Ashish Sabharwal, Mohit Bansal, and Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366.
  44. Unnati V. Shukla and Koushik Tripathy. 2025. Diabetic retinopathy. Updated August 25, 2023, https://www.ncbi.nlm.nih.gov/books/NBK560805/. ↗
  45. Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2023. Progprompt: Generating situated robot task plans using large language models. In Proceedings of the IEEE International Conference on Robotics and Automation (…
  46. R. Singh, K. Ramasamy, C. Abraham, V. Gupta, and A. Gupta. 2008. Diabetic retinopathy: An update. Indian Journal of Ophthalmology, 56(3):179–188. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2636123/. ↗
  47. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, and 1 others. 2023b. React: Synergizing reasoning and acting in language models. In Proceedings of the International Conference on Learning Representations (ICLR).
  48. Peter Spirtes, Clark Glymour, and Richard Scheines. 2000. Causation, Prediction, and Search. MIT Press. find
  49. Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. 2021. Deep long-tailed learning: A survey. arXiv preprint, 2110.04596. find
  50. Oyvind Tafjord and Peter Clark. 2021. Proofwriter: Generating implications, proofs, and abductive explanations. In Proceedings of the Association for Computational Linguistics (ACL), pages 3621–3637. find
  51. Xun Zheng, Bryon Aragam, Pradeep Ravikumar, and Eric Xing. 2018. Dags with no tears: Continuous optimization for structure learning. Advances in Neural Information Processing Systems. find
  52. Midhat Urooj, Ayan Banerjee, and Sandeep Gupta. 2026. Agentic causal disentanglement (ACD) framework: Reversing the generalization–tail trade-off via clinical knowledge integration in medical AI. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. T… find
  53. Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Wenjun Zeng, and Tao Qin. 2021. Generalizing to unseen domains: A survey on domain generalization. arXiv preprint arXiv:2103.03097. — [arXiv:2103.03097](https://arxiv.org/abs/2103.03097) find
  54. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems.
  55. Charles P. Wilkinson, Frederick L. Ferris, Ronald E. Klein, Peter P. Lee, Carl-David Agardh, Mark Davis, and Hans-Peter Hammes. 2003. Proposed international clinical diabetic retinopathy and diabetic macular edema disease severity scales. Ophthalmology, 110(9):1677–1682. find
  56. Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. 2019. Detectron2. https://github.com/facebookresearch/ detectron2. ↗
  57. John Yang, Akshara Prabhakar, Karthik Narasimhan, and Shunyu Yao. 2023. Intercode: Standardizing and benchmarking interactive coding with execution feedback. In Advances in Neural Information Processing Systems.
  58. Yuzhe Yang, Hao Wang, and Dina Katabi. 2022. On multi-domain long-tailed recognition, imbalanced domain generalization and beyond. In Computer Vision – ECCV 2022, volume 13680 of Lecture Notes in Computer Science, pages 57–75, Cham. Springer. find
  59. Myron Yanoff and Jay S. Duker. 2019. Ophthalmology, 5th edition. Elsevier. find
  60. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. 2023a. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/urooj-2026-candice-agentic-causal-disentanglement-class/