---
title: "TabGuard — paper by Unnikrishnan and Ma (2026)"
description: "TabGuard tackles column-level anomaly detection in heterogeneous tables, which mix identifiers, numbers, and free text. 21 references."
authors: ["Srihari Unnikrishnan", "Minghua Ma"]
year: 2026
venue: "Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026)"
source: "https://aclanthology.org/2026.surgellm-1.9/"
doi: "10.18653/v1/2026.surgellm-1.9"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/unnikrishnan-2026-tabguard-agentic-llm-orchestration-adaptive/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/unnikrishnan-2026-tabguard-agentic-llm-orchestration-adaptive/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/51/) / TabGuard: Agentic LLM Orchestration for Adaptive Tabular Anomaly Detection via Dynamic Validator Selection and Generation

# TabGuard: Agentic LLM Orchestration for Adaptive Tabular Anomaly Detection via Dynamic Validator Selection and Generation

**Authors:** Srihari Unnikrishnan and [Minghua Ma](https://md.research.elamur.ai/free/en/authors/minghua-ma/) · **Year:** 2026 · **Venue:** [Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026)](https://md.research.elamur.ai/free/en/venues/surgellm/) ([Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era 2026](https://md.research.elamur.ai/free/en/venues/surgellm/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.surgellm-1.9/) · [PDF](https://aclanthology.org/2026.surgellm-1.9.pdf) · [DOI 10.18653/v1/2026.surgellm-1.9](https://doi.org/10.18653/v1/2026.surgellm-1.9) · [OpenAlex](https://openalex.org/W7166869942)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
TabGuard tackles column-level anomaly detection in heterogeneous tables, which mix identifiers, numbers, and free text. Rule-based checkers need manual setup for every schema. Statistical detectors miss semantic errors. Running an LLM on every cell costs too much at enterprise scale.

The framework uses a GPT-4o agent with function calling. For each column, it samples a handful of structurally diverse values using farthest-point sampling over sentence embeddings. It then picks one of three validators:
- a regex validator for syntactic formats,
- a code-generation validator that writes procedural checks such as Luhn checksums,
- an embedding-similarity validator for distributional outliers.

The chosen validator runs deterministically on all values in the column. This means LLM calls grow with the number of columns, not cells, and each flag comes with an explanation for human review.

On the TABARD benchmark (FeTaQA, WikiTQ, Spider+BEAVER), TabGuard achieves the highest F1 on all three datasets compared with the strongest earlier TABARD LLM configurations. The work suggests that agentic routing can make LLM-based data validation both semantically aware and cheap enough for production.

## Key points
- **Results on TABARD:** F1 is 59.2 on FeTaQA, 62.5 on Spider+BEAVER, and 58.1 on WikiTQ. The best prior configurations reached roughly 53.4, 55.8, and 55.6 respectively. Recall reaches 67.1 on Spider+BEAVER and 63.4 on WikiTQ without a drop in precision.
- **Cost:** the LLM is called once per column (O(m)) instead of once per cell (O(n·m)). The authors' example is a table with 100k rows and 50 columns, which would otherwise need 5 million calls.
- **Why diverse sampling:** with a 95/5 pattern split, random sampling of 15 values catches minority values only about 54% of the time. This motivates the diversity-based sampling step.
- **Validator details:** the embedding validator flags values whose cosine similarity to the column centroid falls below μ − 2σ. Removing the code-generation validator lowers F1 on Spider+BEAVER by an estimated 6–8 points.
- **Limitations:**
  - It detects anomalies within single columns only; cross-column constraints are out of scope.
  - It depends on GPT-4o and on how prompts are written.
  - The two-sigma threshold is fixed.
  - Generated code runs in a sandbox and should still be audited.

*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

- [Microsoft](https://md.research.elamur.ai/free/en/companies/microsoft/) — Company, Redmond, United States · 378 works · 161,630 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎯 Decision Sciences](https://md.research.elamur.ai/free/en/shelves/decision-sciences/) · 249 works · [🎯 Management Science and Operations Research](https://md.research.elamur.ai/free/en/shelves/decision-sciences/management-science-and-operations-research/) · 134 works · [🌐 Computer Networks and Communications](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-networks-and-communications/) · 100 works

## Cite

Cite as: Srihari Unnikrishnan and Minghua Ma. 2026. TabGuard: Agentic LLM Orchestration for Adaptive Tabular Anomaly Detection via Dynamic Validator Selection and Generation. Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{unnikrishnan2026tabguard,
  title = {{TabGuard: Agentic LLM Orchestration for Adaptive Tabular Anomaly Detection via Dynamic Validator Selection and Generation}},
  author = {Srihari Unnikrishnan and Minghua Ma},
  booktitle = {Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.surgellm-1.9},
  url = {https://doi.org/10.18653/v1/2026.surgellm-1.9}
}
```

- [← previous T: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learning](https://md.research.elamur.ai/free/en/papers/xia-2026-t-star-progressive-block-scaling/)
- [next → Table-R1: Region-based Reinforcement Learning for Table Understanding](https://md.research.elamur.ai/free/en/papers/wu-2026-table-r1-region-based-reinforcement/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/51/)

## References

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

1. Tommaso Bendinelli, Artur Dox, and Christian Holz. 2025. Exploring LLM agents for cleaning tabular machine learning datasets. Preprint, arXiv:2503.06664. — [arXiv:2503.06664](https://arxiv.org/abs/2503.06664) [find](https://scholar.google.com/scholar?q=Tommaso+Bendinelli%2C+Artur+Dox%2C+and+Christian+Holz%2E+2025%2E+Exploring+LLM+agents+for+cleaning+tabular+machine+learning+datasets%2E+Preprint%2C+arXiv%3A2503%2E06664%2E+%E2%80%94+%5BarXiv%3A2503%2E06664%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F250)
2. Aditya Bhaskara, Sharvaree Vadgama, and Hong Xu. 2019. Greedy sampling for approximate clustering in the presence of outliers. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc. [find](https://scholar.google.com/scholar?q=Aditya+Bhaskara%2C+Sharvaree+Vadgama%2C+and+Hong+Xu%2E+2019%2E+Greedy+sampling+for+approximate+clustering+in+the+presence+of+outliers%2E+In+Advances+in+Neural+Information+Processing+Systems%2C+volume+32%2E+Curran+A)
3. Fabian Biester, Mohamed Abdelaal, and Daniel Del Gaudio. 2024. LLMClean: Context-aware tabular data cleaning via LLM-generated OFDs. Preprint, arXiv:2404.18681. — [arXiv:2404.18681](https://arxiv.org/abs/2404.18681) [find](https://scholar.google.com/scholar?q=Fabian+Biester%2C+Mohamed+Abdelaal%2C+and+Daniel+Del+Gaudio%2E+2024%2E+LLMClean%3A+Context%2Daware+tabular+data+cleaning+via+LLM%2Dgenerated+OFDs%2E+Preprint%2C+arXiv%3A2404%2E18681%2E+%E2%80%94+%5BarXiv%3A2404%2E18681%5D%28https%3A%2F%2Farxiv%2Eorg%2F)
4. Yunkang Cao, Xiaohao Xu, Chen Sun, Xiaonan Huang, and Weiming Shen. 2023. Towards generic anomaly detection and understanding: Large-scale visual-linguistic model (GPT-4V) takes the lead. Preprint, arXiv:2311.02782. — [arXiv:2311.02782](https://arxiv.org/abs/2311.02782) [find](https://scholar.google.com/scholar?q=Yunkang+Cao%2C+Xiaohao+Xu%2C+Chen+Sun%2C+Xiaonan+Huang%2C+and+Weiming+Shen%2E+2023%2E+Towards+generic+anomaly+detection+and+understanding%3A+Large%2Dscale+visual%2Dlinguistic+model+%28GPT%2D4V%29+takes+the+lead%2E+Preprint%2C+ar)
5. Jaime Carbonell and Jade Goldstein. 1998. The use of MMR, diversity-based reranking for reordering documents and producing summaries. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’98, pages 335–336, New York, NY, USA… [find](https://scholar.google.com/scholar?q=Jaime+Carbonell+and+Jade+Goldstein%2E+1998%2E+The+use+of+MMR%2C+diversity%2Dbased+reranking+for+reordering+documents+and+producing+summaries%2E+In+Proceedings+of+the+21st+Annual+International+ACM+SIGIR+Conferen)
6. Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Lilian Weng, and Aleksander M adry. 2025. [MLE-bench: Evaluating machine learning agents on machine learning engineering](https://md.research.elamur.ai/free/en/papers/chan-2024-mle-bench-evaluating-machine/). Preprint, arXiv:2410.07095.
7. Qixu Chen, Yeye He, Raymond Chi-Wing Wong, Weiwei Cui, Song Ge, Haidong Zhang, Dongmei Zhang, and Surajit Chaudhuri. 2025. Auto-test: Learning semantic-domain constraints for unsupervised error detection in tables. Preprint, arXiv:2504.10762. — [arXiv:2504.10762](https://arxiv.org/abs/2504.10762) [find](https://scholar.google.com/scholar?q=Qixu+Chen%2C+Yeye+He%2C+Raymond+Chi%2DWing+Wong%2C+Weiwei+Cui%2C+Song+Ge%2C+Haidong+Zhang%2C+Dongmei+Zhang%2C+and+Surajit+Chaudhuri%2E+2025%2E+Auto%2Dtest%3A+Learning+semantic%2Ddomain+constraints+for+unsupervised+error+detect)
8. Manan Roy Choudhury, Anirudh Iyengar Kaniyar Narayana Iyengar, Shikhhar Siingh, Sugeeth Puranam, and Vivek Gupta. 2025. TABARD: A novel benchmark for tabular anomaly analysis, reasoning and detection. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 21783–21817, Suzhou… [find](https://scholar.google.com/scholar?q=Manan+Roy+Choudhury%2C+Anirudh+Iyengar+Kaniyar+Narayana+Iyengar%2C+Shikhhar+Siingh%2C+Sugeeth+Puranam%2C+and+Vivek+Gupta%2E+2025%2E+TABARD%3A+A+novel+benchmark+for+tabular+anomaly+analysis%2C+reasoning+and+detection%2E)
9. Danil. 2025. WikiTQ (revision a95e0b1). [find](https://scholar.google.com/scholar?q=Danil%2E+2025%2E+WikiTQ+%28revision+a95e0b1%29%2E)
10. Margarida Ferreira, Victor Nicolet, Luan Pham, Joey Dodds, Daniel Kroening, Ines Lynce, and Ruben Martins. 2025. Hypergraph-guided regex filter synthesis for event-based anomaly detection. Preprint, arXiv:2509.06911. [find](https://scholar.google.com/scholar?q=Margarida+Ferreira%2C+Victor+Nicolet%2C+Luan+Pham%2C+Joey+Dodds%2C+Daniel+Kroening%2C+Ines+Lynce%2C+and+Ruben+Martins%2E+2025%2E+Hypergraph%2Dguided+regex+filter+synthesis+for+event%2Dbased+anomaly+detection%2E+Preprint%2C+a)
11. Madelon Hulsebos, Kevin Hu, Michiel Bakker, Emanuel Zgraggen, Arvind Satyanarayan, Tim Kraska, Çağatay Demiralp, and César Hidalgo. 2019. Sherlock: A deep learning approach to semantic data type detection. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data… [find](https://scholar.google.com/scholar?q=Madelon+Hulsebos%2C+Kevin+Hu%2C+Michiel+Bakker%2C+Emanuel+Zgraggen%2C+Arvind+Satyanarayan%2C+Tim+Kraska%2C+%C3%87ag%CC%86atay+Demiralp%2C+and+C%C3%A9sar+Hidalgo%2E+2019%2E+Sherlock%3A+A+deep+learning+approach+to+semantic+data+type+dete)
12. Alex Kulesza. 2012. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2–3):123–286. [find](https://scholar.google.com/scholar?q=Alex+Kulesza%2E+2012%2E+Determinantal+point+processes+for+machine+learning%2E+Foundations+and+Trends%C2%AE+in+Machine+Learning%2C+5%282%E2%80%933%29%3A123%E2%80%93286%2E)
13. Hans P. Luhn. 1960. Computer for verifying numbers. Filed 1954-01-06. [find](https://scholar.google.com/scholar?q=Hans+P%2E+Luhn%2E+1960%2E+Computer+for+verifying+numbers%2E+Filed+1954%2D01%2D06%2E)
14. Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kryściński, Nick Schoelkopf, Riley Kong, Xiangru Tang, Murori Mutuma, Ben Rosand, Isabel Trindade, Renusree Bandaru, Jacob Cunningham, Caiming Xiong, and Dragomir Radev. 2022. [FeTaQA: Free-form table question…](https://md.research.elamur.ai/free/en/papers/nan-2022-fetaqa-free-form-table-question/)
15. Asara Senaratne, Peter Christen, Graham Williams, and Pouya Ghiasnezhad Omran. 2023. Rule-based knowledge discovery via anomaly detection in tabular data. CEUR Workshop Proceedings, 3433. [find](https://scholar.google.com/scholar?q=Asara+Senaratne%2C+Peter+Christen%2C+Graham+Williams%2C+and+Pouya+Ghiasnezhad+Omran%2E+2023%2E+Rule%2Dbased+knowledge+discovery+via+anomaly+detection+in+tabular+data%2E+CEUR+Workshop+Proceedings%2C+3433%2E)
16. Nouman Shamim, Muhammad Asim, Ali Ismail Awad, and Muhammad Khurram Khan. 2025. Anomaly detection in Internet of Things system calls using a centroid-based vector-space model. IEEE Internet of Things Journal, 12(14):26868–26881. [find](https://scholar.google.com/scholar?q=Nouman+Shamim%2C+Muhammad+Asim%2C+Ali+Ismail+Awad%2C+and+Muhammad+Khurram+Khan%2E+2025%2E+Anomaly+detection+in+Internet+of+Things+system+calls+using+a+centroid%2Dbased+vector%2Dspace+model%2E+IEEE+Internet+of+Things+)
17. Yoshihiko Suhara, Jinfeng Li, Yuliang Li, Dan Zhang, Çağatay Demiralp, Chen Chen, and Wang-Chiew Tan. 2022. Annotating columns with pre-trained language models. In Proceedings of the 2022 International Conference on Management of Data, pages 1493–1503, New York, NY, USA. Association for Computing M… [find](https://scholar.google.com/scholar?q=Yoshihiko+Suhara%2C+Jinfeng+Li%2C+Yuliang+Li%2C+Dan+Zhang%2C+%C3%87ag%CC%86atay+Demiralp%2C+Chen+Chen%2C+and+Wang%2DChiew+Tan%2E+2022%2E+Annotating+columns+with+pre%2Dtrained+language+models%2E+In+Proceedings+of+the+2022+Internation)
18. Che-Ping Tsai, Ganyu Teng, Phil Wallis, and Wei Ding. 2025. AnoLLM: Large language models for tabular anomaly detection. In The Thirteenth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Che%2DPing+Tsai%2C+Ganyu+Teng%2C+Phil+Wallis%2C+and+Wei+Ding%2E+2025%2E+AnoLLM%3A+Large+language+models+for+tabular+anomaly+detection%2E+In+The+Thirteenth+International+Conference+on+Learning+Representations%2E)
19. Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. 2021. MiniLMv2: Multi-head self-attention relation distillation for compressing pre-trained transformers. Preprint, arXiv:2012.15828. [find](https://scholar.google.com/scholar?q=Wenhui+Wang%2C+Hangbo+Bao%2C+Shaohan+Huang%2C+Li+Dong%2C+and+Furu+Wei%2E+2021%2E+MiniLMv2%3A+Multi%2Dhead+self%2Dattention+relation+distillation+for+compressing+pre%2Dtrained+transformers%2E+Preprint%2C+arXiv%3A2012%2E15828%2E)
20. Xiaoyin Wang and Dakai Zhu. 2024. Validating LLM-generated programs with metamorphic prompt testing. Preprint, arXiv:2406.06864. — [arXiv:2406.06864](https://arxiv.org/abs/2406.06864) [find](https://scholar.google.com/scholar?q=Xiaoyin+Wang+and+Dakai+Zhu%2E+2024%2E+Validating+LLM%2Dgenerated+programs+with+metamorphic+prompt+testing%2E+Preprint%2C+arXiv%3A2406%2E06864%2E+%E2%80%94+%5BarXiv%3A2406%2E06864%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2406%2E06864%29)
21. Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. [Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task](https://md.research.elamur.ai/free/en/papers/yu-2018-spider-large-scale-human-labeled/). In Proceedings of the 2018…

*Updated 1 November 2026*