Evaluating Latent Knowledge of Public Tabular Datasets in Large Language Models

2025-10-23

AI Summary

This study reveals that large language models’ (LLMs) strong performance on public tabular datasets often stems from data contamination—i.e., memorization of semantic cues (e.g., column names, value distributions) encountered during training—rather than genuine reasoning. To isolate memory from reasoning, we design controlled probing experiments employing semantic denoising and column-name randomization. Our results show that removing such semantic cues collapses model accuracy to chance level, exposing severe overestimation of generalization capability. We introduce the novel concept of *semantic contamination bias* and propose a principled evaluation paradigm that explicitly disentangles memorization from reasoning. This framework establishes a methodological foundation and empirical basis for rigorous LLM assessment on structured data. (126 words)

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large Language Models (LLMs) are increasingly evaluated on their ability to reason over structured data, yet such assessments often overlook a crucial confound: dataset contamination. In this work, we investigate whether LLMs exhibit prior knowledge of widely used tabular benchmarks such as Adult Income, Titanic, and others. Through a series of controlled probing experiments, we reveal that contamination effects emerge exclusively for datasets containing strong semantic cues-for instance, meaningful column names or interpretable value categories. In contrast, when such cues are removed or randomized, performance sharply declines to near-random levels. These findings suggest that LLMs' apparent competence on tabular reasoning tasks may, in part, reflect memorization of publicly available datasets rather than genuine generalization. We discuss implications for evaluation protocols and propose strategies to disentangle semantic leakage from authentic reasoning ability in future LLM assessments.

Problem

• Assessing LLMs' latent knowledge of public tabular datasets
• Investigating dataset contamination effects in tabular benchmarks
• Disentangling memorization from genuine reasoning in evaluations

Innovation

• Probing LLMs for dataset contamination via controlled experiments
• Revealing performance decline with randomized semantic cues
• Proposing strategies to separate memorization from true reasoning