Pith. sign in

REVIEW 3 major objections 6 minor 26 references

Towards Benchmarking Foundation Models for Tabular Data With Text

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Text columns reliably boost tabular model accuracy when embedded and downsampled.

desk verdict Useful benchmark corpus with real free-text columns, but the 11/13 'text helps' headline is conditional on an ambiguous SHAP downsampling protocol and best-embedding selection. read the letter →

arxiv 2507.07829 v1 pith:L6M2TAWU submitted 2025-07-10 cs.LG

Martin Mráz , Breenda Das , Anshul Gupta , Lennart Purucker , Frank Hutter This is my paper
classification cs.LG
keywords tabulardatawithtextembeddingsfoundationmodelsbenchmarkcurationfeaturedownsamplingTabPFNv2AutoGluonfast
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that tabular foundation models should be benchmarked on data that actually contains free text, and that current practice mostly avoids this setting. It curates 13 real-world datasets with genuinely semantic text fields alongside structured columns, and shows that replacing those text columns with numeric embeddings usually improves accuracy across TabPFNv2, XGBoost, and two AutoML frameworks. The central empirical claim is that text embeddings beat dropping the text in 11 of 13 datasets.

What carries the argument

The central mechanism is the per-column text-embedding pipeline: embed each free-text field with fastText, Skrub's GapEncoder, or AutoGluon's text encoder, then downsample the resulting high-dimensional vectors to 300 features (often via SHAP importance) before feeding them to a tabular model.

What would settle it

Recompute Table 2 with a single pre-registered embedding method (e.g., fastText only) and a single downsampler across all 13 datasets; if the with-text accuracy does not beat no-text in at least 9 of 13 datasets, the headline claim fails.

Watch

Extended reading notes

Core claim

On 11 of 13 curated datasets, text embeddings improve over no-text baselines for all models and frameworks tested; no single embedding method dominates, and SHAP-based feature downsampling is most frequently the best selector.

Load-bearing premise

The 11/13 accuracy claim assumes it is fair to report the best embedding method for each dataset after seeing the test results, rather than a fixed pipeline chosen in advance.

Editorial extensions

If this is right

  • Practitioners with free-text columns can expect accuracy gains from embedding them, though the best embedding method varies by dataset.
  • Downsampling text embeddings is essential; SHAP-based selection is a strong default but not universal.
  • Tabular foundation model evaluations that omit text miss a realistic source of signal.
  • Current embedding methods remain brittle, motivating better text encoders for tabular tasks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 11/13 result likely overstates what a practitioner would see with a fixed embedding method, since Table 2 reports the best method per dataset after seeing results.
  • A testable extension would fix one embedding method and one downsampler and measure the average gain across all 13 datasets.
  • The qualitative synthetic failures (synonym OOD, noise, semantic ambiguity) suggest the true bottleneck is not the tabular model but the text encoder's robustness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes TextTabBench, a manually curated set of 13 real-world tabular datasets containing free-text columns, and evaluates four text-embedding routes (FastText, Skrub's TableVectorizer/GapEncoder, AutoGluon's feature generator, and a BERT variant) combined with TabPFNv2, XGBoost, the TabPFNv2 API, and AutoGluon's Tabular Predictor. It reports that adding text embeddings improves accuracy in 11/13 datasets under a protocol that selects the best embedding per dataset after SHAP-based feature downsampling, and it presents a synthetic qualitative study of embedding failure modes plus a detailed critique of the CARTE benchmark. The paper also contributes acquisition scripts, preprocessing notebooks, and a discussion of benchmark-design principles for tabular data with text.

Significance. If the empirical protocol is valid, the paper is a useful step: the curated corpus fills a real gap, the preprocessing is documented in unusual detail, and the qualitative synthetic experiments in Section 3 clearly illustrate distinct embedding failure modes. The claim that text helps tabular models is actionable, and the benchmark itself is likely to be reused by the community. However, the headline 11/13 result is only as strong as the evaluation protocol in Table 2, and the current protocol has two features—post-hoc selection of the best embedding per dataset and a SHAP downsampling step whose nesting inside cross-validation is not specified—that can inflate the with-text advantage. The CARTE critique in Appendix G is thoughtful and useful, but the LLM-based grading of dataset quality has no human agreement check, which weakens the benchmark-construction argument.

major comments (3)
  1. [Section 5, Table 2 caption; Appendix B] The central claim that textual features boost accuracy in 11/13 datasets depends on SHAP-based downsampling not leaking validation target information. Appendix B states that SHAP 'computes the average absolute SHAP value for each feature across the dataset and selects the top-k most important ones,' but the paper never states that this supervised selection is nested inside the 5-fold cross-validation loop described in Section 5. If SHAP is fit on the full 3000-row subsample before splitting, the selector sees validation-fold labels, while the no-text baseline undergoes no such supervised selection; this biases the comparison in favor of the with-text condition. The same ambiguity applies to the other supervised selectors in Appendix F (t-test, ANOVA, correlation, L1). Please state explicitly whether feature selection is performed per training fold only, and if it is not, rerun the experiments with properly nested selection and report whether the 11/13 result survives.
  2. [Table 2 and Section 5, result (i)] The headline comparison reports, for each dataset and model, the best performance among all embedding techniques downsampled by SHAP. Because the best embedding is chosen after seeing the cross-validated scores, the 11/13 win rate is an upper bound over methods rather than the performance of a fixed pipeline a practitioner would deploy; for example, on airbnb only AGPIPE reaches 0.692 while FastText under SHAP reaches 0.682, and on osha only AGPIPE gives the reported gain. The paper should report the fixed-embedding comparison (e.g., FastText with SHAP for every dataset) or a nested model-selection procedure, and should accompany the aggregate claim with a paired significance test, since several improvements (e.g., airbnb 0.692±0.048 vs 0.679±0.045 for TabPFNv2) are within one standard deviation.
  3. [Appendix G.1, G.4, G.10] The benchmark-quality assessment of CARTE and of the authors' own datasets relies on GPT-4o judgments (domain grouping in G.4 and Green/Yellow/Red classification in G.10) without any human inter-annotator agreement check or error analysis. Since Appendix G is used to motivate the new corpus and to criticize an existing benchmark, the LLM judgments should be validated on a sample by human annotators, and the prompts and outputs should be released; otherwise the critique is self-referential and cannot be independently evaluated.
minor comments (6)
  1. [Section 2 and Section 5] There are several typos: 'pre-column embeddings' should be 'per-column embeddings', 'dowsampling' should be 'downsampling', and Section 3's contribution list says 'show casing' instead of 'showcasing'.
  2. [Section 5, Experimental Setup] The paper states that rows are restricted to 3000 for all evaluations, but Table 1 lists datasets with far more rows (e.g., mercari 100000, complaints 96935); please clarify whether the 3000-row cap is applied before or after the 100,000-row crop and whether Table 1 reports raw or evaluation sizes.
  3. [Appendix G.4, Figures 6 and 7] The figure captions say that binary-thresholded coverage at or above 0.5 is highlighted in blue, while the surrounding text says green cells indicate strong similarity; please harmonize the color descriptions.
  4. [Appendix B, Table 3] The Correlation row is marked as both Zero-Shot and Supervised; since correlation with the target uses target information, the Zero-Shot check appears to be an error and should be corrected.
  5. [Section 4 and Appendix A] The code repository is mentioned by name ('TextTabBench repository' and 'Towards TextTabBench repository') but no URL is provided in the manuscript; please include a working link for reproducibility.
  6. [Appendix E and Section 3] The main text says the qualitative study uses five OpenML datasets, while Table 4's caption in Appendix E says four; please correct the inconsistency.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: the central 'text helps' claim rests on external models and real datasets, but one self-referential Appendix G validation loops back to the authors' own criteria.

  1. other [Appendix G.1, G.4, G.10 (supporting Section 2's critique of CARTE); Listing 8 prompt]
    "To approximate this process in a transparent and relatively independent way, we incorporate outputs from a highly capable LLM (ChatGPT-4o via API), along with the corresponding prompts. ... Based on GPT's evaluations, the majority of CARTE's datasets were categorized as either Red or Yellow ... In contrast, our benchmark datasets were predominantly classified as Green."

    The 'relatively independent' LLM check is not independent of the paper's own definitions: the prompt in Listing 8 states the authors' five fitness conditions (semantic richness, dual signal, native target, enough signal), and G.1 says these rules were derived from examining CARTE's limitations. GPT-4o's verdict that the authors' datasets are 'predominantly Green' while CARTE's are 'Red/Yellow' is therefore the authors' suitability definition re-applied to both pools by a black-box judge, with no human-agreement check reported. As evidence that the new benchmark is better suited than CARTE, the support reduces to the paper's own stated criteria rather than an external standard.

full rationale

This paper contains no mathematical derivation chain; its claims are empirical benchmark results. The central claim, that text embeddings improve accuracy in 11/13 datasets (Section 5.2, Table 2), is supported by 5-fold CV comparisons of TabPFNv2, XGBoost, TabPFNv2 API, and AutoGluon on 13 real datasets against a no-text baseline, using standard statistics of these external models. This comparison is self-contained and does not reduce to the paper's inputs: the win rate is not guaranteed by construction (permits and houses lose under some frameworks), and the TabPFNv2 citation is to an externally published Nature model co-authored by two of the present authors, which is ordinary co-authorship, not load-bearing circularity. Suspected issues do not rise to pith circularity. First, Table 2's protocol, 'we report best performance amongst all embedding techniques downsampled by SHAP,' gives the with-text condition a best-of-n selection advantage over a single no-text baseline; this can inflate the reported gain but does not force it, and fixed embeddings (e.g., FastText) also beat no-text in most datasets, so this is a claim-strength and statistical-design caveat, not a reduction by construction. Second, Appendix B describes SHAP as computing average absolute SHAP values 'across the dataset' without stating whether selection is nested inside the 5-fold CV; if not nested, validation targets could leak into with-text feature selection but not into the no-text baseline. This is a potential correctness risk that cannot be adjudicated from the text and does not match any circularity pattern. Third, the Section 3 synthetic study builds texts to trigger specific embedding failures; the observed failures follow from the construction, but that is a designed controlled experiment, and the paper acknowledges the tests 'may seem artificial.' The one genuinely self-referential element is Appendix G: GPT-4o certifies the authors' own benchmark as Green and CARTE's as largely Red/Yellow using a prompt that encodes the authors' own selection criteria, making the 'relatively independent' validation an echo of the paper's own definitions. This is minor and not load-bearing for the central empirical claim, so the score sits at the top of the no-significant-circularity band.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The paper contributes an empirical benchmark and introduces no equations or fitted constants. The free parameters are experimental caps and curation thresholds that influence the headline 11/13 result. The axioms are mostly assumptions about benchmark validity and evaluation protocol, several of which are acknowledged or left unquantified in the paper.

free parameters (4)
  • max_features_cap = 300
    All text embeddings are downsampled to at most 300 features to fit TabPFNv2 memory; this cap affects which downsampling technique performs best.
  • max_rows_cap = 3000
    Rows are restricted to 3000 for all evaluations, which may change the text signal available in larger datasets.
  • categorical_threshold = 50 unique values or 5% of rows, manually adjusted
    Appendix A uses this threshold to decide whether a string column is categorical or textual, with manual adjustments in a few cases; this shapes which columns count as text.
  • skrub_gapencoder_components = 30 (default)
    Skrub's TableVectorizer uses GapEncoder with 30 components by default; the paper's ablation shows that higher dimensions improve performance but not enough to change the main conclusions.
assumptions (5)
  • domain assumption Each selected dataset has both a meaningful structured signal and a semantic free-text signal (dual-signal criterion).
    Section 4 rule (ii) requires this, but Appendix G.8 explicitly says the dual-signal criterion was not quantified and served only as a guiding principle.
  • ad hoc to paper The best embedding reported per dataset is a legitimate target-independent choice.
    Table 2 reports the best performance over embeddings after seeing results; no fixed protocol or nested model selection is described, so the 11/13 aggregate presumes oracle selection is valid.
  • ad hoc to paper SHAP-based feature selection does not leak target information into the 5-fold cross-validation.
    Appendix B describes SHAP as supervised and model-based, but Section 5 does not state that feature selection is nested inside each training fold.
  • domain assumption GPT-4o assessments are reliable for classifying dataset quality and domain redundancy.
    Appendix G.4 and G.10 use GPT-4o to group CARTE domains and grade dataset fit, but no human validation or agreement measure is reported.
  • domain assumption The synthetic dilution tasks in Section 3 are representative of real-world text variability.
    Section 3 says the tests 'may seem artificial' but assumes synonym variation, random noise, and semantic ambiguity are common in real long texts; no real-text validation is provided.
invented entities (1)
  • TextTabBench dataset corpus independent evidence
    purpose: A collection of 13 real-world tabular datasets with free-text columns for benchmarking tabular foundation models with text.
    The datasets are real and are sourced from Kaggle and similar portals; acquisition scripts and preprocessing notebooks are promised in a repository. This is a curated resource, not a hypothetical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Benchmarking Foundation Models for Tabular Data With Text." pith.science (2026). https://pith.science/paper/L6M2TAWU

@misc{pith2026250707829,
  author       = {Pith},
  title        = {Pith review of: Towards Benchmarking Foundation Models for Tabular Data With Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L6M2TAWU}},
  note         = {Machine review of arXiv:2507.07829}
}
read the original abstract

Foundation models for tabular data are rapidly evolving, with increasing interest in extending them to support additional modalities such as free-text features. However, existing benchmarks for tabular data rarely include textual columns, and identifying real-world tabular datasets with semantically rich text features is non-trivial. We propose a series of simple yet effective ablation-style strategies for incorporating text into conventional tabular pipelines. Moreover, we benchmark how state-of-the-art tabular foundation models can handle textual data by manually curating a collection of real-world tabular datasets with meaningful textual features. Our study is an important step towards improving benchmarking of foundation models for tabular data with text.

Figures

Figures reproduced from arXiv: 2507.07829 by the authors.

Figure 1
Figure 1. Qualitative investigation of textual embeddings for tabular prediction tasks: TabPFN v2 accuracy across five datasets. Baselines “No-Text” uses original input features, “Complete Leak” has target leakage, therefore 100% accuracy for all. Following tests embed targets into textual modality. “N-Gram Break” shows TF-IDF breaking under unseen synonyms, “Simple NLP break” shows FastText underperforming under noise, “LLM … view at source ↗
Figure 2
Figure 2. Performance improvement with FastText over mean accuracy across all ablations on TabPFN v2: FastText generally gives the best scores, with a few exceptions.These exceptions generally have longer length of texts than average. We present our complete study of each embedding’s performance under every downsampling technique. Tables 6, 7 and 8 compare performance of TabPFN and XGBoost under different downsampling techniq… view at source ↗
Figure 3
Figure 3. Ablation over larger embedding sizes for Skrub GapEncoder over five datasets: beer, calif houses, laptops, sf permits, wine. Bars show percentage difference between best score and best skrub embedding score (across all downsampling techniques), averaged across all chosen datasets. Skrub’s performance improves with larger embedding dimensions, though the improvement slows down, and any further increase in dimensional… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: SHAP performance improvement over mean accuracy across all ablations on FastText embeddings: SHAP generally coincides with the best score using FastText for most datasets, on few exceptions, the performance is slightly below the best score [PITH_FULL_IMAGE:figures/ful…
Figure 5
Figure 5. Figure 5: Performance improvement with BERT variant over best accuracy across all ablations on TabPFN v2: BERT is sub-optimal compared to best scores over other embeddings: FastText, Skrub and AGPIPE per dataset. This can be possibly due to higher embedding dimensions than the r…
Figure 6
Figure 6. Figure 6: Directional feature coverage between used vehicle listing datasets. Left: proportion of semantically similar features from dataset A to B. Right: binary thresholded version where coverage ≥ 0.5 is highlighted in blue. Grey cells indicate missing comparisons [PITH_FULL…
Figure 7
Figure 7. Figure 7: Directional feature coverage between cross-domain datasets with low expected similarity. The left heatmap shows continuous coverage from dataset A to B; the right applies a binary threshold of ≥ 0.5. Grey cells indicate unpaired or unmatched datasets. Directional Table…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 4 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    G., van Rijn, J

    Bischl, B., Casalicchio, G., Feurer, M., Hutter, F., Lang, M., Mantovani, R. G., van Rijn, J. N., and Vanschoren, J. Openml benchmarking suites. arXiv:1708.03731v2 [stat.ML], 2019

    arXiv 2019

  3. [3]

    Enriching word vectors with subword information, 2017

    Bojanowski, P., Grave, E., Joulin, A., and Mikolov, T. Enriching word vectors with subword information, 2017. URL https://arxiv.org/abs/1607.04606

    arXiv 2017

  4. [4]

    V., Na, L., Ma, Y., Boussioux, L., Zeng, C., Soenksen, L

    Carballo, K. V., Na, L., Ma, Y., Boussioux, L., Zeng, C., Soenksen, L. R., and Bertsimas, D. Tabtext: A flexible and contextual approach to tabular data representation, 2023. URL https://arxiv.org/abs/2206.10381

    arXiv 2023

  5. [5]

    and Guestrin, C

    Chen, T. and Guestrin, C. XGBoost : A scalable tree boosting system. In Krishnapuram, B., Shah, M., Smola, A., Aggarwal, C., Shen, D., and Rastogi, R. (eds.), Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining ( KDD '16) , pp.\ 785--794, 2016

    work page 2016

  6. [7]

    Autogluon-tabular: Robust and accurate automl for structured data, 2020 b

    Erickson, N., Mueller, J., Shirkov, A., Zhang, H., Larroy, P., Li, M., and Smola, A. Autogluon-tabular: Robust and accurate automl for structured data, 2020 b . URL https://arxiv.org/abs/2003.06505

    arXiv 2020

  7. [8]

    Gijsbers, P., Bueno, M. L. P., Coors, S., LeDell, E., Poirier, S., Thomas, J., Bischl, B., and Vanschoren, J. Amlb: an automl benchmark, 2022. URL https://arxiv.org/abs/2207.12560

    arXiv 2022

  8. [9]

    L., Amaral, L

    Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff, J. M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody, G. B., Peng, C.-K., and Stanley, H. E. Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals. Circulation, 101 0 (23): 0 e215--e220, 2000

    work page 2000

Show all 26 references
  1. [10]

    J., and Varoquaux, G

    Grinsztajn, L., Oyallon, E., Kim, M. J., and Varoquaux, G. Vectorizing string entries for data processing on tables: when are larger language models better?, 2023. URL https://arxiv.org/abs/2312.09634

    2023 arXiv

  2. [11]

    Tabllm: Few-shot classification of tabular data with large language models, 2022

    Hegselmann, S., Buendia, A., Lang, H., Agrawal, M., Jiang, X., and Sontag, D. Tabllm: Few-shot classification of tabular data with large language models, 2022. URL https://arxiv.org/abs/2210.10723

    2022 arXiv

  3. [12]

    N., Chang, S., Hartvigsen, T., and Singh, H

    Hegselmann, S., Parziale, A., Shanmugam, D., Tang, S., Asiedu, M. N., Chang, S., Hartvigsen, T., and Singh, H. Machine learning for health symposium 2023 -- findings track, 2023. URL https://arxiv.org/abs/2312.00655

    2023 arXiv

  4. [13]

    u ller, S., Purucker, L., Krishnakumar, A., K \

    Hollmann, N., M \"u ller, S., Purucker, L., Krishnakumar, A., K \"o rfer, M., Hoo, S. B., Schirrmeister, R. T., and Hutter, F. Accurate predictions on small data with a tabular foundation model. Nature, 01 2025. doi:10.1038/s41586-024-08328-6. URL https://www.nature.com/articl...

    2025 doi

  5. [14]

    J., Grinsztajn, L., and Varoquaux, G

    Kim, M. J., Grinsztajn, L., and Varoquaux, G. Carte: Pretraining and transfer for tabular learning, 2024. URL https://arxiv.org/abs/2402.16785

    2024 arXiv

  6. [15]

    Llm embeddings for deep learning on tabular data, 2025

    Koloski, B., Margeloiu, A., Jiang, X., Škrlj, B., Simidjievski, N., and Jamnik, M. Llm embeddings for deep learning on tabular data, 2025. URL https://arxiv.org/abs/2502.11596

    2025 arXiv

  7. [16]

    Talent: A tabular analytics and learning toolbox

    Liu, S.-Y., Cai, H.-R., Zhou, Q.-L., and Ye, H.-J. Talent: A tabular analytics and learning toolbox. arXiv preprint arXiv:2407.04057, 2024

    2024 arXiv

  8. [17]

    Mug: A multimodal classification benchmark on game data with tabular, textual, and visual fields

    Lu, J., Qian, Y., Zhao, S., Xi, Y., and Yang, C. Mug: A multimodal classification benchmark on game data with tabular, textual, and visual fields. In Findings of the Association for Computational Linguistics: EMNLP 2023, pp.\ 5332–5346. Association for Computational Linguistic...

    2023 doi

  9. [18]

    Lundberg, S. M. and Lee, S.-I. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017

    2017

  10. [19]

    C., Golestan, K., Yu, G., Volkovs, M., and Caterini, A

    Ma, J., Thomas, V., Hosseinzadeh, R., Kamkari, H., Labach, A., Cresswell, J. C., Golestan, K., Yu, G., Volkovs, M., and Caterini, A. L. Tabdpt: Scaling tabular foundation models, 2024. URL https://arxiv.org/abs/2410.18164

    2024

  11. [20]

    and Ratajczak, W

    Maćkiewicz, A. and Ratajczak, W. Principal components analysis (pca). Computers & Geosciences, 19 0 (3): 0 303--342, 1993. ISSN 0098-3004. doi:https://doi.org/10.1016/0098-3004(93)90090-R. URL https://www.sciencedirect.com/science/article/pii/009830049390090R

    1993

  12. [21]

    McElfresh, D., Khandagale, S., Valverde, J., Prasad C , V., Ramakrishnan, G., Goldblum, M., and White, C. When do neural nets outperform boosted trees on tabular data? In Proceedings of the 37th International Conference on Advances in Neural Information Processing Systems ( N ...

    2023

  13. [22]

    P., Feuer, B., Hegde, C., Ramakrishnan, G., Goldblum, M., and White, C

    McElfresh, D., Khandagale, S., Valverde, J., C, V. P., Feuer, B., Hegde, C., Ramakrishnan, G., Goldblum, M., and White, C. When do neural nets outperform boosted trees on tabular data?, 2024. URL https://arxiv.org/abs/2305.02997

    2024 arXiv

  14. [23]

    Shi, X., Mueller, J., Erickson, N., Li, M., and Smola, A. J. Benchmarking multimodal automl for tabular data with text fields, 2021. URL https://arxiv.org/abs/2111.02705

    2021 arXiv

  15. [24]

    A., Gonzalez, J., Duvenaud, D., and Turner, R

    Shysheya, A., Bronskill, J., Requeima, J., Siddiqui, S. A., Gonzalez, J., Duvenaud, D., and Turner, R. E. Jolt: Joint probabilistic predictions on tabular data using llms, 2025. URL https://arxiv.org/abs/2502.11877

    2025 arXiv

  16. [25]

    Autogluon-multimodal (automm): Supercharging multimodal automl with foundation models, 2024

    Tang, Z., Fang, H., Zhou, S., Yang, T., Zhong, Z., Hu, T., Kirchhoff, K., and Karypis, G. Autogluon-multimodal (automm): Supercharging multimodal automl with foundation models, 2024. URL https://arxiv.org/abs/2404.16233

    2024 arXiv

  17. [26]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, 2020

    Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., and Zhou, M. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, 2020. URL https://arxiv.org/abs/2002.10957

    2020 arXiv

  18. [27]

    Tablellm: Enabling tabular data manipulation by llms in real office usage scenarios, 2024

    Zhang, X., Luo, S., Zhang, B., Ma, Z., Zhang, J., Li, Y., Li, G., Yao, Z., Xu, K., Zhou, J., Zhang-Li, D., Yu, J., Zhao, S., Li, J., and Tang, J. Tablellm: Enabling tabular data manipulation by llms in real office usage scenarios, 2024. URL https://arxiv.org/abs/2403.19318

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.