Elamur Research⌕
☰
home / papers / 2026 / Self-Reinforcing Controllable Synthesis of Rare Relational Data via Bayesian Calibration

Self-Reinforcing Controllable Synthesis of Rare Relational Data via Bayesian Calibration

Paper · 2026 · Chongsheng Zhang, Hao Wang, Zelong Yu et al.

Summary

Rare classes in tabular (relational) datasets hurt classifier performance. LLM-based synthesis of such data has received little study, and current prompt-based generators get no signal about how good their output is while they run. The authors present RDDG (Relational Data generator with Dynamic Guidance), a training-free in-context learning pipeline built from staged chain-of-thought steps. First, it picks a compact, class-balanced core set of representative training rows so that the examples fit in the context window. Next, it asks the LLM to infer dependencies and correlations between attributes and turns them into explicit generation constraints. Finally, it generates data in batches. After each batch, several fidelity measures compare the output with real data, and the scores become feedback prompts for the next batch. The authors cast this loop as Bayesian calibration and give theoretical optimality arguments for it. On eight real and synthetic datasets, RDDG beats the in-context baseline EPIC in downstream imbalanced classification and in how faithfully it reproduces the data. This makes it a practical, fine-tuning-free route to minority-class augmentation.

Key points

  • Core set selection: An MLP is trained, and each sample's L2 prediction error is tracked over early and late epochs. Within every class, the Top-K samples whose error variance is largest are retained. This adapts a method by Hong et al. (2024) that was originally built for medical data.
  • Classification results: Against state-of-the-art in-context approaches, RDDG gains on average more than 2% in weighted Macro-F1 and more than 1% in Balanced Accuracy.
  • Fidelity results versus EPIC:
    • Overall KL divergence drops from 4.55 to 3.72, and RDDG wins on 6 of 8 datasets. On Thyroid it falls from 2.83 to 0.278.
    • Correlation preservation improves on every metric: Frobenius norm 0.827 vs 2.40, MAE 0.048 vs 0.115, RMSE 0.065 vs 0.163.
  • Ablations:
    • On Travel and Thyroid, the core set outperforms randomly sampled subsets, and the feedback loop clearly improves results.
    • On UCI segmentation with varied imbalance ratios, RDDG's lead over EPIC widens as imbalance increases.
  • Limitations and code: RDDG depends on an LLM accessed through an API or run locally. Batch-wise generation can weaken consistency across batches. Code is available at github.com/cszhangLMU/RDDG.

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

🎓Ludwig-Maximilians-Universität MünchenMunich · Germany113 works · 56,817 citations🎓Munich Center for Machine LearningMunich · Germany62 works · 1,167 citations🔬Helmholtz Center for Information SecuritySaarbrücken · Germany20 works · 1,172 citations🎓Henan UniversityKaifeng · China2 works

 Fields, topics and concepts

📚 Shelves 4The 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 works

Cite

Cite as Chongsheng Zhang, Hao Wang, Zelong Yu, Esteban Garces Arias, Julian Rodemann, Zhanshuo Zhang, Qilong Li, Gaojuan Fan, Krikamol Muandet and Christian Heumann. 2026. Self-Reinforcing Controllable Synthesis of Rare Relational Data via Bayesian Calibration. 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.423

@inproceedings{zhang2026self,
  title = {{Self-Reinforcing Controllable Synthesis of Rare Relational Data via Bayesian Calibration}},
  author = {Chongsheng Zhang, Hao Wang, Zelong Yu, Esteban Garces Arias, Julian Rodemann, Zhanshuo Zhang, Qilong Li, Gaojuan Fan, Krikamol Muandet and Christian Heumann},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.423},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.423}
}

References

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

  1. Sigrún Andradóttir and Vicki M Bier. 2000. Applying bayesian ideas in simulation. Simulation Practice and Theory, 8(3-4):253–280. find
  2. James O Berger. 2013. Statistical decision theory and Bayesian analysis. Springer Science & Business Media. find
  3. Vadim Borisov, Kathrin Seßler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. 2022. Language models are realistic tabular data generators. In Proceedings of the International Conference on Learning Representations (ICLR). find
  4. Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, and W. Philip Kegelmeyer. 2002. SMOTE: Synthetic minority over-sampling technique. Preprint, arXiv:1106.1813.
  5. Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794. ACM. find
  6. Anna Veronika Dorogush, Vasily Ershov, and Andrey Gulin. 2018. CatBoost: Gradient boosting with categorical features support. In Proceedings of the 2018 IEEE International Conference on Data Mining Workshops (ICDMW), pages 1021–1028. IEEE. find
  7. Xun Liang, Hanyu Wang, Yezhaohui Wang, Shichao Song, Jiawei Yang, Simin Niu, Jie Hu, Dan Liu, Shunyu Yao, Feiyu Xiong, and 1 others. 2024. Controllable text generation for large language models: A survey. arXiv preprint arXiv:2408.12599.
  8. Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. 2020. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition (CVPR 2020), pages 2970–2979. find
  9. Markus Mueller, Kathrin Gruber, and Dennis Fok. 2025. Continuous diffusion for mixed-type tabular data. In The Thirteenth International Conference on Learning Representations (ICLR 2025). find
  10. Josh Gardner, Juan C. Perdomo, and Ludwig Schmidt. 2024. Large scale transfer learning for tabular data via language modeling. In Advances in Neural Information Processing Systems (NeurIPS 2024), pages 45155–45205. find
  11. Noseong Park, Mahmoud Mohammadi, Kshitij Gorde, Sushil Jajodia, Hongkyu Park, and Youngmin Kim. 2018. Data synthesis based on generative adversarial networks. Proceedings of the VLDB Endowment, 11:1071–1083. find
  12. Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. 2023. Tabpfn: A transformer that solves small tabular classification problems in a second. In The Eleventh International Conference on Learning Representations (ICLR 2023). find
  13. David Poole and Adrian E Raftery. 2000. Inference for deterministic simulation models: the bayesian melding approach. Journal of the American Statistical Association, 95(452):1244–1255. find
  14. Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. 2025. Accurate predictions on small data with a tabular foundation model. Nature, 637(8045):319–326. find
  15. Yuxin Hong, Xiao Zhang, Xin Zhang, and Joey Tianyi Zhou. 2024. Evolution-aware variance (EVA) coreset selection for medical image classification. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 301–310. find
  16. Haque Ishfaq, Assaf Hoogi, and Daniel L. Rubin. 2018. TVAE: Triplet-based variational autoencoder using metric learning. Preprint, arXiv:1802.04403. — [arXiv:1802.04403](https://arxiv.org/abs/1802.04403) find
  17. Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems, pages 3146–3154. Curran Associates Inc. find
  18. Jinhee Kim, Taesung Kim, and Jaegul Choo. 2024. EPIC: Effective prompting for imbalanced-class data synthesis in tabular data classification via large language models. In Advances in Neural Information Processing Systems. find
  19. Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. 2022. TabDDPM: Modelling tabular data with diffusion models. Preprint, arXiv:2209.15421. — [arXiv:2209.15421](https://arxiv.org/abs/2209.15421) find
  20. Herbert Robbins and Sutton Monro. 1951. A stochastic approximation method. The annals of mathematical statistics, pages 400–407. find
  21. Aivin V. Solatorio and Olivier Dupriez. 2023. REaLTabFormer: Generating realistic relational and tabular data using transformers. arXiv preprint, arXiv 2302.02041. — [arXiv:2302.02041](https://arxiv.org/abs/2302.02041) find
  22. Boris van Breugel and Mihaela van der Schaar. 2024. Position: Why tabular foundation models should be a research priority. In Forty-first International Conference on Machine Learning (ICML 2024). find
  23. Stephen Wade, Marianne F Weber, Peter Sarich, Pavla Vaneckova, Silvia Behar-Harpaz, Preston J Ngo, Sonya Cressman, Coral E Gartner, John M Murray, and Tony A Blakely. 2022. Bayesian calibration of simulation models: a tutorial and an australian smoking behaviour model. arXiv preprint arXiv:2202.0292… — [arXiv:2202.02923](https://arxiv.org/abs/2202.02923) find
  24. Chaozheng Wang, Shuzheng Gao, Pengyun Wang, Cuiyun Gao, Wenjie Pei, Lujia Pan, and Zenglin Xu. 2022. Label-aware distribution calibration for long-tailed classification. IEEE Transactions on Neural Networks and Learning Systems, 35(5):6963–6975. find
  25. Wentao Wang, Suhang Wang, Wenqi Fan, Zitao Liu, and Jiliang Tang. 2020. Global-and-local aware data generation for the class imbalance problem. In Proceedings of the 2020 SIAM International Conference on Data Mining (SDM 2020), pages 307–315. find
  26. 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, volume 35, pages 24824–24837.
  27. Xumeng Wen, Han Zhang, Shun Zheng, Wei Xu, and Jiang Bian. 2024. From supervised to generative: A novel paradigm for tabular deep learning with large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2024), pages 3323–3333. find
  28. Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. 2019. Modeling tabular data using conditional GAN. In Advances in Neural Information Processing Systems. find
  29. Jinsung Yoon, Louise N. Drumright, and Mihaela van der Schaar. 2020. Anonymization through data synthesis using generative adversarial networks (ADS-GAN). IEEE Journal of Biomedical and Health Informatics, 24(8):2378–2388. find
  30. Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. 2023a. A survey of controllable text generation using transformer-based pre-trained language models. ACM Computing Surveys, 56(3):1– 37. find
  31. Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan, Zhengyuan Shen, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. 2023b. Mixed-type tabular data synthesis with score-based diffusion in latent space. arXiv preprint arXiv:2310.09656. — [arXiv:2310.09656](https://arxiv.org/abs/2310.09656) find
  32. Qihao Zhao, Yalun Dai, Hao Li, Wei Hu, Fan Zhang, and Jun Liu. 2024a. LTGC: Long-tail recognition via leveraging LLMs-driven generated content. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19510–19520. find
  33. Zilong Zhao, Aditya Kunar, Robert Birke, Hiek van der Scheer, and Lydia Y. Chen. 2024b. CTAB-GAN+: Enhancing tabular data synthesis. Frontiers in Big Data, 6:1296508. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-self-reinforcing-controllable-synthesis-rare/