Both Text and Images Leaked! A Systematic Analysis of Data Contamination in Multimodal LLM

TL;DR

This paper systematically analyzes data contamination in multimodal large language models, revealing significant contamination issues that impact fair evaluation and model reliability, with contamination often originating during unimodal pre-training.

Contribution

The paper introduces MM-Detect, an analytical framework for quantifying multimodal data contamination, and provides comprehensive analysis across multiple models and benchmarks.

Findings

Significant data contamination detected in proprietary and older models.

Contamination often originates during unimodal pre-training, not just multimodal fine-tuning.

Contamination impacts evaluation fairness and model reliability.

Abstract

The rapid advancement of multimodal large language models (MLLMs) has significantly enhanced performance across benchmarks. However, data contamination-unintentional memorization of benchmark data during model training-poses critical challenges for fair evaluation. Existing detection methods for unimodal large language models (LLMs) are inadequate for MLLMs due to multimodal data complexity and multi-phase training. We systematically analyze multimodal data contamination using our analytical framework, MM-Detect, which defines two contamination categories-unimodal and cross-modal-and effectively quantifies contamination severity across multiple-choice and caption-based Visual Question Answering tasks. Evaluations on twelve MLLMs and five benchmarks reveal significant contamination, particularly in proprietary models and older benchmarks. Crucially, contamination sometimes originates during unimodal pre-training rather than solely from multimodal fine-tuning. Our insights refine contamination understanding, guiding evaluation practices and improving multimodal model reliability.