Do VLMs Have Bad Eyes? Diagnosing Compositional Failures via Mechanistic Interpretability

TL;DR

This paper investigates why vision-language models like CLIP struggle with compositional generalization, revealing that neuron superposition in the vision encoder impairs their ability to represent and reason about complex object combinations.

Contribution

The study uncovers mechanistic reasons behind compositional failures in VLMs, highlighting neuron superposition as a key factor affecting their reasoning capabilities.

Findings

Neuron superposition in CLIP's vision encoder hampers compositional representation

Mechanistic interpretability reveals neurons encode multiple features

Addressing superposition could improve VLMs' compositional reasoning

Abstract

Vision-Language Models (VLMs) have shown remarkable performance in integrating visual and textual information for tasks such as image captioning and visual question answering. However, these models struggle with compositional generalization and object binding, which limit their ability to handle novel combinations of objects and their attributes. Our work explores the root causes of these failures using mechanistic interpretability techniques. We show evidence that individual neurons in the MLP layers of CLIP's vision encoder represent multiple features, and this "superposition" directly hinders its compositional feature representation which consequently affects compositional reasoning and object binding capabilities. We hope this study will serve as an initial step toward uncovering the mechanistic roots of compositional failures in VLMs. The code and supporting results can be found https://github.com/Mystic-Slice/Do-VLMs-Have-Bad-Eyes.