VT-Bench: A Unified Benchmark for Visual-Tabular Multi-Modal Learning

Referee: [§3 and §4] §3 (Dataset Aggregation) and §4 (Task Definitions): the manuscript must explicitly document the selection criteria for the 14 datasets, including any exclusion rules, domain balance metrics, and preprocessing pipelines. Without these, it is impossible to assess whether the benchmark fairly captures core visual-tabular challenges or introduces selection bias toward medical data.

Authors: We agree that explicit documentation of benchmark construction is essential for transparency and to allow assessment of potential biases. Section 3 of the original manuscript provides an overview of the 14 datasets with Table 1 summarizing domains, sample counts, and sources, while §4 defines the tasks. However, we acknowledge the need for more detail on selection. In the revised manuscript, we will add a new subsection 'Dataset Selection and Preprocessing' in §3 that explicitly states: (1) Selection criteria included public availability of paired visual-tabular data, relevance to discriminative prediction or generative reasoning, minimum sample size (>1,000 for statistical reliability), and coverage of high-stakes domains; (2) Exclusion rules: datasets were excluded if they lacked one modality, contained only synthetic data, had restricted access, or were too small; (3) Domain balance: we will report metrics such as the proportion of samples per domain (medical: ~65%, pets: ~15%, media: ~10%, transportation: ~10%) and note that medical dominance reflects real-world prevalence of visual-tabular data (e.g., imaging + EHR) rather than arbitrary choice, while non-medical domains were deliberately included for diversity; (4) Preprocessing pipelines: uniform steps including image resizing to 224x224, tabular feature standardization, missing value imputation via mean/mode, and consistent train/validation/test splits (70/15/15). These additions will directly address concerns about fairness and selection bias without changing the benchmark composition. revision: yes

Referee: [§5] §5 (Model Evaluation): the reported results for the 23 models lack statistical controls such as multiple random seeds, confidence intervals, or significance tests for the claimed 'substantial challenges.' This weakens the ability to draw reliable conclusions about relative model performance across discriminative and generative tasks.

Authors: We appreciate the emphasis on statistical rigor for drawing reliable conclusions about model performance gaps. The original §5 reports single-run results across the 23 models to highlight clear trends (e.g., unimodal models struggling with cross-modal integration and VLMs showing limited tabular reasoning). To strengthen this, the revised manuscript will include: (1) averages and standard deviations over 3 random seeds for all models with stochastic components (e.g., fine-tuning or generation sampling); (2) 95% confidence intervals for primary metrics such as accuracy, F1-score (discriminative tasks), and BLEU/ROUGE (generative tasks); (3) paired statistical tests (e.g., t-tests) between model categories to quantify significance of the observed challenges. Due to substantial computational costs for re-evaluating all 23 models (particularly large VLMs and tool-augmented systems) across 14 datasets, we will apply full multi-seed analysis to representative subsets (baselines, top performers, and one from each category) and note this as a limitation for the remainder. These updates will better substantiate the claims of substantial challenges while remaining feasible for minor revision. revision: partial