Elamur Research⌕
☰
home / papers / 2026 / Can LLMs Self-Correct Table Reasoning Errors?

Can LLMs Self-Correct Table Reasoning Errors?

Paper · 2026 · Farseen Shaikh

Summary

The paper asks whether a single LLM, without external tools, can find and fix its own mistakes in table reasoning. In this setting, errors come from misread cells, faulty arithmetic, broken logic or invented values. Five models from five providers (Gemini 3.1 Pro, Kimi K2.5, GLM 5, Qwen 3.5+, MiniMax M2.5) were tested on 1,000 examples each from WikiTableQuestions and TabFact. The protocol was paired and multi-seed, with bootstrap CIs and McNemar tests. The author proposes Structured Self-Correction (SSC), an inference-time prompt chain in which the model reviews cells, calculations, logic and completeness in turn.

The Accuracy-Correction Paradox, first reported for math, also appears on tables. Models starting roughly in the 65–75% accuracy range gain a little from self-correction. Stronger models lose accuracy because they rewrite answers that were already right. SSC reduces this over-correction in most settings. The study matters because it turns these findings into practical guidance on when a lightweight self-review step is worth adding to a table QA system.

Key points

  • Self-Correction Gain (SCG, averaged over seeds) peaks at about +1.3% for mid-range models. It falls to about −1.3% for stronger models, with 95% bootstrap CIs below zero.
  • Compared with naive self-correction, SSC lowers over-correction in 9 of 10 conditions, cutting it by 38–69% on TabFact. For example, Qwen 3.5+ on TabFact drops from 2.0% to 0.9%.
  • Turning reasoning mode off leaves Qwen 3.5+'s results stable. GLM 5's SCG drops sharply, to −3.4 on WikiTQ and −4.7 on TabFact, so robustness to inference mode depends on the model.
  • A FinQA transfer probe shows a capability floor: at 21.5% base accuracy, self-correction does not help.
  • Suggested rule: use answer-aware, single-round SSC for base accuracy of about 65–75%. Avoid naive self-correction above 80%. Code, prompts and outputs are at github.com/FarseenSh/table-self-correction.

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 5The 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 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 works

Cite

Cite as Farseen Shaikh. 2026. Can LLMs Self-Correct Table Reasoning Errors?. 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. DOI 10.18653/v1/2026.surgellm-1.19

@inproceedings{shaikh2026can,
  title = {{Can LLMs Self-Correct Table Reasoning Errors?}},
  author = {Farseen Shaikh},
  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-1.19},
  url = {https://doi.org/10.18653/v1/2026.surgellm-1.19}
}

References

25 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2020. TabFact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations.
  2. Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. 2021. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural…
  3. Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. 2023. Binding language models in symbolic languages. In International Conference on Learning Representations.
  4. Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2024. Chain-of-verification reduces hallucination in large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 3563–3578.
  5. Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large language models cannot self-correct reasoning yet. In International Conference on Learning Representations.
  6. Atin Sakkeer Hussain. 2024. ARTEMIS-DA: An advanced reasoning and transformation engine for multi-step insight synthesis in data analytics. arXiv preprint arXiv:2412.14146. — [arXiv:2412.14146](https://arxiv.org/abs/2412.14146) find
  7. Chuang Jiang, Mingyue Cheng, Xiaoyu Tao, Qingyang Mao, Jie Ouyang, and Qi Liu. 2026. Table-Mind: An autonomous programmatic agent for tool-augmented table reasoning. In Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining, pages 260–270. find
  8. Ruochun Jin, Dong Wang, Xiyue Wang, and Haoqi Zheng. 2025. TableCritic: Refine table reasoning via self-criticism and tool library. In Proceedings of the 47th Annual Meeting of the Cognitive Science Society. find
  9. Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, and Rui Zhang. 2024. When can LLMs actually correct their own mistakes? A critical survey of self-correction of LLMs. Transactions of the Association for Computational Linguistics, 12:1417–1440.
  10. Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, and 1 others. 2025. Training language models to self-correct via reinforcement learning. In International Conference on Learning Representations. Oral.
  11. Yin Li. 2025. Decomposing LLM self-correction: The accuracy-correction paradox and error depth hypothesis. arXiv preprint arXiv:2601.00828. — [arXiv:2601.00828](https://arxiv.org/abs/2601.00828) find
  12. Circulated December 2025; posted arXiv January 2026. find
  13. Aofan Liu and Jingxiang Meng. 2026. Self-correction as feedback control: Error dynamics, stability thresholds, and prompt interventions in LLMs. arXiv preprint arXiv:2604.22273. — [arXiv:2604.22273](https://arxiv.org/abs/2604.22273) find
  14. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems 36. find
  15. Panupong Pasupat and Percy Liang. 2015. Compositional semantic parsing on semi-structured tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1470–…
  16. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems 36, pages 8634–8652.
  17. Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. 2024. Table meets LLM: Can large language models understand structured table data? A benchmark and empirical study. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 645–654. find
  18. Gladys Tyen, Hassan Mansoor, Victor Carbune, Peter Chen, and Tony Mak. 2024. LLMs cannot find reasoning errors, but can correct them given the error location. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13894–13908.
  19. Lukas Voss. 2026. Calibrated confidence estimation for tabular question answering. arXiv preprint arXiv:2604.12491. — [arXiv:2604.12491](https://arxiv.org/abs/2604.12491) find
  20. Self-correction distillation for structured data question answering. In Proceedings of the Fortieth AAAI Conference on Artificial Intelligence, pages 16566– 16574. find
  21. Zilong Wang, Hao Zhang, Chun-Liang Li, Julian Martin Eisenschlos, Vincent Perot, Zifeng Wang, Lesly Miculicich, Yasuhisa Fujii, Jingbo Shang, Chen-Yu Lee, and Tomas Pfister. 2024. Chain-of-table: Evolving tables in the reasoning chain for table understanding. In International Conference on Learning… find
  22. Zhe Yang, Yichang Zhang, Yudong Wang, Ziyao Xu, Junyang Lin, and Zhifang Sui. 2025a. Confidence v.s. critique: A decomposition of self-correction capability for LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3998– 4014.
  23. Zhen Yang, Ziwei Du, Minghan Zhang, Wei Du, Jie Chen, Zhen Duan, and Shu Zhao. 2025b. Triples as the key: Structuring makes decomposition and verification easier in LLM-based TableQA. In International Conference on Learning Representations. find
  24. Yunhu Ye, Binyuan Hui, Min Yang, Binhua Li, Fei Huang, and Yongbin Li. 2023. Large language models are versatile decomposers: Decomposing evidence and questions for table-based reasoning. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retriev… find
  25. Peiying Yu, Guoxin Chen, and Jingjing Wang. 2025. Table-critic: A multi-agent framework for collaborative criticism and refinement in table reasoning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 17432–17451. *Figure 3: Sel…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shaikh-2026-can-llms-self-correct-table/