- 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
- 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.
- 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
-
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?
-
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?
-
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?