Findings of EACL 2026

What Really Matters for Table LLMs?

A Meta-Evaluation of Model and Data Effects

Naihao Deng 1* , Sheng Zhang 2 , Henghui Zhu 3 , Shuaichen Chang 2 , Jiani Zhang 5 , Alexander Hanbo Li 4 , Chung-Wei Hang 2 , Hideo Kobayashi 2 , Yiqun Hu 2 , Patrick Ng 2

1University of Michigan  ·  2AWS AI Labs  ·  3Figma  ·  4OKX  ·  5Google

* Work done during Naihao's internship at AWS

📄 Paper (Anthology) → 💻 Code (GitHub) → 🤗 Models (HF) → 💾 Predictions dataset →
Conceptual overview: 3 base models × 4 training corpora → 12 models → 16 table benchmarks + 4 RQs
We replicate four table LLMs by instruction-tuning three foundation models on four existing training datasets, yielding 12 models, then evaluate across 16 table benchmarks against four research questions.

📝 Abstract

Table understanding has evolved over decades of NLP research. In this work, we revisit this trajectory and highlight emerging challenges in the LLM era — particularly the paradox of choice: the difficulty of attributing performance gains amid diverse base models and training sets in the context of table instruction tuning.

We replicate four table LLMs by instruction-tuning three foundation models on four existing datasets, yielding 12 models. We then evaluate these models across 16 table benchmarks. Our study is the first to quantitatively disentangle the effects of training data and base model selection, revealing that base model choice plays a more dominant role than the training data itself. Generalization and reasoning remain challenging, inviting future effort on table modeling.

⚡ TL;DR

📊 Results browser

Per-{base, training corpus, benchmark} metric. Toggle the chips to focus on a subset of base models, training corpora, or benchmarks. Cells are colour-scaled within each column (darker = higher).

Base model
Train corpus
Model wikitq fetaqa tabfact tabmwp totto infotabs tatqa hitab tablebench TLLM·wtq TLLM·fetaqa TLLM·tatqa TLLM·ottqa Beer DeepM Efthymiou Spreadsheets_CF Spreadsheets_DI Spreadsheets_ED_Real Spreadsheets_MVI_ColumnNoSep instruction_following
mistral-v0.3-tablellama 19.0 65.1 86.9 28.1 53.2 27.7 13.3 33.5 16.9 41.1 73.1 55.2 8.3 48.8 28.3 69.0 0.1
mistral-v0.3-tablellm 5.6 40.5 3.5 3.7 22.4 2.9 11.1 1.3 3.5 8.7 40.2 14.4 8.5 29.6 90.7 96.6 42.8 60.4 21.6 83.3 0.0
mistral-v0.3-tablebench 6.9 31.3 4.0 3.7 18.8 5.1 8.3 1.1 19.2 87.8 85.7 97.5 42.6 64.0 23.9 84.2 0.1
mistral-v0.3-tablegpt 15.8 42.3 3.2 13.3 31.1 4.5 16.1 6.7 11.0 25.5 93.6 97.2 45.5 69.1 44.9 83.8 0.1
mistral-v0.3-tablegpt-small 25.5 89.1 97.0 44.1 69.1 40.7 84.8
olmo-tablellama 0.6 63.8 83.8 12.6 50.3 0.0 10.7 3.1 12.5 3.7 62.2 65.5 59.7 63.0 8.8 85.4 0.1
olmo-tablellm 7.6 40.5 2.5 4.1 32.1 1.5 11.0 1.1 5.9 8.6 39.8 14.3 9.6 25.0 86.7 84.4 49.6 62.6 24.3 54.0 0.1
olmo-tablebench 5.1 31.0 2.4 5.3 27.1 2.2 10.8 1.4 14.2 12.3 75.1 87.8 60.4 66.0 15.8 80.5 0.1
olmo-tablegpt 8.5 41.8 3.1 10.9 36.6 2.1 13.7 3.4 10.9 25.5 97.7 83.4 73.0 68.9 44.7 81.9 0.1
phi-3-tablellama 40.4 64.1 86.2 50.7 55.6 54.2 29.2 65.2 22.6 99.5 92.9 63.2 67.2 22.7 34.5 88.0 0.1
phi-3-tablellm 12.3 52.8 5.6 6.6 26.1 3.4 18.1 2.3 18.4 9.8 43.1 16.1 9.7 59.1 92.5 97.5 80.6 70.5 35.1 88.3 0.1
phi-3-tablebench 11.2 51.2 4.7 9.0 21.7 11.2 16.5 2.1 18.4 15.8 67.9 93.5 69.1 45.2 13.0 64.4 0.1
phi-3-tablegpt 13.8 57.4 6.1 11.3 43.6 2.2 19.9 2.8 19.5 24.5 99.9 97.3 90.6 92.3 70.5 89.4 0.1
phi-3-mini-tablellama 28.5 61.5 81.6 51.1 52.9 16.4 28.3 46.4 18.9 90.0 57.8 96.7 43.2 56.7 23.1 81.5 0.1
phi-3-mini-tablellm 10.8 52.7 3.4 6.9 29.9 3.3 18.8 1.8 13.6 9.5 42.3 15.9 9.4 17.3 57.5 96.8 43.8 57.3 15.9 80.8 0.1
phi-3-mini-tablebench 8.3 51.0 2.2 4.3 34.0 8.5 15.6 1.5 18.1 9.0 57.9 97.0 43.4 57.0 5.8 79.7 0.1
phi-3-mini-tablegpt 11.4 54.3 14.6 16.2 47.1 48.4 16.1 3.2 16.5 26.1 85.5 97.2 86.3 91.1 63.5 89.1 0.1

310 cells loaded. Metric per benchmark is BLEU-4 for FeTaQA / ToTTo / TableLLM-FeTaQA, ROUGE-L for the rest. See the paper for details.

🎯 Result: base model dominates

Shapley R² decomposition over the 3×4 grid: across the 16 table benchmarks, choice of base model contributes substantially more variance than the choice of training corpus.

Shapley R² decomposition showing base model > training data contribution

📚 Citation

@inproceedings{deng-etal-2026-really,
    title = "What Really Matters for Table {LLM}s? A Meta-Evaluation of Model and Data Effects",
    author = "Deng, Naihao  and Zhang, Sheng  and Zhu, Henghui  and Chang, Shuaichen  and Zhang, Jiani  and Li, Alexander Hanbo  and Hang, Chung-Wei  and Kobayashi, Hideo  and Hu, Yiqun  and Ng, Patrick",
    booktitle = "Findings of the Association for Computational Linguistics: EACL 2026",
    year = "2026",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-eacl.195/",
    doi = "10.18653/v1/2026.findings-eacl.195"
}