Exploring the Robustness of Language Models for Tabular Question Answering via Attention Analysis

TL;DR

This paper investigates how large language models handle tabular question answering, analyzing attention shifts and robustness across domains, and emphasizes the need for more interpretable and structure-aware models for reliable table comprehension.

Contribution

It provides a detailed analysis of attention behavior and robustness in LLMs for TQA, highlighting the impact of model scale, instruction tuning, and domain bias, and proposes directions for improving interpretability and reliability.

Findings

Larger, instruction-tuned LLMs show improved robustness.

Attention shifts correlate with performance drops under perturbations.

Middle layers of models are most sensitive to perturbations.

Abstract

Large Language Models (LLMs), already shown to ace various unstructured text comprehension tasks, have also remarkably been shown to tackle table (structured) comprehension tasks without specific training. Building on earlier studies of LLMs for tabular tasks, we probe how in-context learning (ICL), model scale, instruction tuning, and domain bias affect Tabular QA (TQA) robustness by testing LLMs, under diverse augmentations and perturbations, on diverse domains: Wikipedia-based $\textbf{WTQ}$, financial $\textbf{TAT-QA}$, and scientific $\textbf{SCITAB}$. Although instruction tuning and larger, newer LLMs deliver stronger, more robust TQA performance, data contamination and reliability issues, especially on $\textbf{WTQ}$, remain unresolved. Through an in-depth attention analysis, we reveal a strong correlation between perturbation-induced shifts in attention dispersion and the drops in performance, with sensitivity peaking in the model's middle layers. We highlight the need for improved interpretable methodologies to develop more reliable LLMs for table comprehension. Through an in-depth attention analysis, we reveal a strong correlation between perturbation-induced shifts in attention dispersion and performance drops, with sensitivity peaking in the model's middle layers. Based on these findings, we argue for the development of structure-aware self-attention mechanisms and domain-adaptive processing techniques to improve the transparency, generalization, and real-world reliability of LLMs on tabular data.