acceptodds
Withdrawn from ICLR 2027

How Does Text Help in Tables? A Systematic Evaluation of Text-Rich Tabular Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world tabular datasets often contain free-form text alongside numerical and categorical features, yet existing methods incorporate such textual information through substantially different interfaces and are typically evaluated under separate protocols. We present a systematic evaluation of text-rich tabular learning that compares four strategies under a unified framework: tabular-only prediction, embedding-augmented learning with frozen text representations, native text-aware modeling, and large language model (LLM) direct prediction across 44 datasets spanning binary classification, multiclass classification, and regression. Through controlled experiments and analyses, we establish three key findings: first, incorporating even basic text representations consistently enhances supervised tabular learners; second, the strongest embedding-augmented model outperforms nearly all native text-aware models; and third, LLM direct prediction underperforms both supervised paradigms due to numerical insensitivity and output compression. Together, these results show that the effectiveness of text-rich tabular prediction depends critically on how textual evidence is represented and integrated with structured supervision. Code and evaluation pipelines are available at https://anonymous.4open.science/r/TextTableBench.

Primary area:
datasets and benchmarks
Keywords:
Text-Rich Tabular Learning, Datasets and Benchmarks, Tabular foundation models, Large language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2NbpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

-This paper presents an empirical study that systematically compares four strategies for incorporating textual information into predictive modeling on text-rich tabular data under a unified evaluation framework. The strategies considered are (i) tabular-only learning that excludes text fields, (ii) embedding-augmented learning that combines frozen representations from pretrained text encoders with structured features, (iii) native text-aware models that directly process textual and structured features, and (iv) LLM direct prediction using prompts constructed from individual tabular records.

-The authors evaluate these strategies on 44 datasets selected from the existing TabSTAR benchmark, comprising 4 binary classification, 11 multiclass classification, and 29 regression tasks. The main findings are as follows.

-First, adding frozen text embeddings to the same tabular learner improves predictive performance on most datasets (Table 1).

-Second, TabICLv2 augmented with E5/PCA50 achieves performance comparable to, or better than, several native text-aware models (Table 2).

-Third, under the direct prediction protocol adopted in this study, LLMs underperform supervised learning approaches, particularly in regression tasks, where prediction variance is reduced, and performance degrades with additional few-shot demonstrations (Table 6, Appendices B.7-8).

-This paper's primary contribution is an empirical comparison of text-integration strategies and diagnostic analyses of prediction behavior, not a new learning algorithm.

Critical strengths

1) The authors conduct paired comparisons with and without textual information across five learners on the same datasets, controlling for differences in the underlying predictive models when evaluating text representations. Across 44 datasets, adding text improves performance on 36--39 datasets, and the positive findings generally hold when classification and regression tasks are analyzed separately (Tables1, 11-12). These results show that the practical benefits of incorporating textual information are not limited to specific models or a small subset of datasets.

2) The comparison between TF-IDF/SVD and pretrained embeddings (Appendix B.5), text shuffling and removal experiments (Section 5.2), and analyses of LLM regression predictions using Spearman rank correlation, prediction variance, oracle recalibration, and tied classification scores (Section 5.3 and Appendices B.7-8) provide useful insights for interpreting the experimental results.

Critical weaknesses
  1. The paper's independent research contributions beyond existing benchmarks are unclear.
  • Mraz et al. (2025) already evaluated strategies for integrating text representations into tabular data and the performance of foundation models. STRABLE (Blayer et al., 2026) also analyzed the relative performance of modular text encoding and end-to-end approaches, and the competitiveness of simple text representations, using 108 datasets and multiple pipelines. The paper also selected the 44 datasets from the existing TabSTAR benchmark rather than collecting new ones. Although the focus on genuinely free-form text and the inclusion of LLM direct prediction may differentiate the paper, several main findings overlap with prior studies.
  1. The experimental design does not sufficiently isolate the effects of text integration strategies.
  • The main comparison in Section 4.3 is conducted between TabICLv2 augmented with E5/PCA50 and native text-aware models that differ in their underlying predictive architectures, pretraining procedures, and learning mechanisms. Consequently, it is difficult to attribute the observed performance differences specifically to the choice between frozen embeddings and native text integration. Moreover, the median performance difference relative to TabSTAR is small, while the difference relative to TabPFN-3-Plus is not statistically significant (Table 2).
Questions for authors
  1. Excluding the benefits of adding text representations and the comparisons between modular and native approaches already reported by STRABLE (Blayer et al., 2026) and Mraz et al.(2025), what new conclusions does this paper establish?

  2. Have the authors verified whether the performance degradation from 0-shot to 32-shot is robust to changes in demonstration sampling strategies, input context length, text truncation, and output format?

  3. Section 3.2 requires text fields to have an average character length of at least 50. However, Table 7 reports an average text length of 12.9 for “nba-draft-basketball” and 29.5 for “top-ramen-ratings-2022”, each containing only one text field. What units and aggregation procedures were used to calculate the average text lengths reported in Table 7, and how can these values be reconciled with the stated text-field selection criteria?

Reviewer_Mp1YOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a systematic evaluation of text-rich tabular learning, comparing four text-integration strategies. Through experiments and analyses, the study shows: incorporating even basic text representations consistently boosts supervised tabular performance; the strongest embedding-augmented model baseline outperforms nearly all native text-aware models, and LLM direct prediction underperforms supervised paradigms due to numerical insensitivity and output compression, with in-context demonstrations further degrading performance.

Critical strengths

In general, the paper is organized and easy to follow. The analysis shown provide some learning to readers.

Critical weaknesses
  • My main concern for the paper is that there is not too much different finding from previous studies. Most of the claims are well-established from previous studies (CARTE, TabSTAR, TabuLa, STRABLE benchmark papers).
  • Baselines must be enlarged and improved. It is missing ContextTab, and information on hyperparameter optimization, search spaces is missing, which is quite critical in these analytical studies.
  • From the list of datasets, most of them corresponds to datasets from CARTE, with additional ones from TabStar. As noted from the recent works, these datasets are highly curated, which can cause challenges of its own.
Questions for authors
  • What are the grounds for determining the textual columns?
  • Also, please refer the weaknesses.
Reviewer_r1umOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper provides a relatively comprehensive study of text-rich tabular learning. Under a unified data processing and evaluation framework, the authors compare four common text-integration strategies: tabular-only prediction, embedding-augmented learning with frozen text representations, native text-aware modeling, and LLM direct prediction. The experiments cover most of the datasets currently used in tabular learning. The results show that textual information consistently improves prediction performance on the vast majority of datasets, and that simple frozen text embedding methods are already highly competitive. In comparison, LLM direct prediction still performs much worse than supervised tabular models. Overall, this paper provides a detailed analysis of the general value of textual features for tabular learning and offers several useful conclusions.

Critical strengths

(1) The paper builds a unified evaluation framework for text-rich tabular learning. It compares four main text-integration strategies across many commonly used datasets and task types. The experiments also include representative recent leading tabular learning methods, such as TabPFN-3-Plus. Therefore, the complete study provides a useful view of the current performance of text-rich tabular learning methods.

(2) The paper goes beyond comparing the performance of different methods. Through a series of interpretable ablation study, it further studies whether and how textual information is used by the tabular models. These analyses are useful.

(3) The paper provides an in-depth study of the direct prediction behavior of LLMs on text-rich tabular tasks. In addition to standard performance experiments, it further analyzes properties such as output scale. This analysis is helpful for practical need to use LLMs for prediction on tabular data with text.

Critical weaknesses

(1) The main research questions and key empirical findings in the supervised learning part of the paper overlap with the cited TabSTAR and STRABLE studies. As a result, the scope of the paper’s novelty is not fully clear.

(2) In Sections 3.2 and 3.3, the authors explain that the datasets are selected from the TabSTAR benchmark using filtering rules based on text length. However, Table 7 still contains more than ten datasets with clearly short average text lengths. This raises an additional question: do short and long texts require different text encoding and integration methods?

(3) The paper lacks a text-only baseline that uses only textual features. The current comparisons mainly include a structured-only setting and a setting that uses both structured features and text.

Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.