Learning or Cheating? Assessing Data Contamination in Large Vision-Language Models

Le résumé fourni par la source

Large Vision-Language Models (VLMs) have demonstrated remarkable capabilities in tasks that involve both visual and textual understanding, including chart interpretation, document comprehension, and geometric reasoning. However, concerns remain about whether these models truly generalize or if their performance on popular benchmarks is influenced by overfitting and data contamination. To investigate these concerns, we propose a systematic evaluation framework for assessing data contamination of both closed-source and open-source VLMs using multiple visual question answering benchmarks, including ChartQA, DocVQA, InfoVQA, and MathVista. Our approach applies structured perturbations such as image replacement and question completion to four visual question answering benchmarks (ChartQA, DocVQA, InfoVQA, MathVista). By comparing original and perturbed accuracies, we compute a contamination score that measures how much each model relies on memorized examples versus genuine multimodal reasoning. Our results reveal significant performance shifts and conclude that both closed-source and open-source VLMs suffer from data contamination. These findings underscore the necessity of rigorous data filtering and independent evaluations to ensure robust generalization before deploying VLMs in real-world applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.