What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Gaussian-Noise-free Text-Image Corruption and Evaluation

TL;DR

This paper introduces NOTICE, a novel interpretability pipeline for Vision-Language Models that uses noise-free corruption techniques to analyze internal decision processes and identify key attention mechanisms across tasks.

Contribution

It presents the first noise-free corruption and evaluation pipeline for mechanistic interpretability in VLMs, enabling causal analysis of multimodal integration.

Findings

Identification of crucial middle-layer cross-attention heads.

Discovery of universal cross-attention heads with distinct functions.

Insights into VLM decision-making across multiple datasets.

Abstract

Vision-Language Models (VLMs) have gained community-spanning prominence due to their ability to integrate visual and textual inputs to perform complex tasks. Despite their success, the internal decision-making processes of these models remain opaque, posing challenges in high-stakes applications. To address this, we introduce NOTICE, the first Noise-free Text-Image Corruption and Evaluation pipeline for mechanistic interpretability in VLMs. NOTICE incorporates a Semantic Minimal Pairs (SMP) framework for image corruption and Symmetric Token Replacement (STR) for text. This approach enables semantically meaningful causal mediation analysis for both modalities, providing a robust method for analyzing multimodal integration within models like BLIP. Our experiments on the SVO-Probes, MIT-States, and Facial Expression Recognition datasets reveal crucial insights into VLM decision-making, identifying the significant role of middle-layer cross-attention heads. Further, we uncover a set of ``universal cross-attention heads'' that consistently contribute across tasks and modalities, each performing distinct functions such as implicit image segmentation, object inhibition, and outlier inhibition. This work paves the way for more transparent and interpretable multimodal systems.