Structural Graph Probing of Vision-Language Models

TL;DR

This paper introduces a neural topology approach to analyze vision-language models, revealing how population structure relates to behavior and identifying influential neurons through correlation graphs.

Contribution

It presents a novel method of representing VLMs as correlation graphs to interpret internal organization and causal influence of neurons.

Findings

Correlation topology encodes behavioral signals.

Cross-modal structure consolidates with depth around hub neurons.

Targeted perturbation of hub neurons significantly impacts model output.

Abstract

Vision-language models (VLMs) achieve strong multimodal performance, yet how computation is organized across populations of neurons remains poorly understood. In this work, we study VLMs through the lens of neural topology, representing each layer as a within-layer correlation graph derived from neuron-neuron co-activations. This view allows us to ask whether population-level structure is behaviorally meaningful, how it changes across modalities and depth, and whether it identifies causally influential internal components under intervention. We show that correlation topology carries recoverable behavioral signal; moreover, cross-modal structure progressively consolidates with depth around a compact set of recurrent hub neurons, whose targeted perturbation substantially alters model output. Neural topology thus emerges as a meaningful intermediate scale for VLM interpretability: richer than local attribution, more tractable than full circuit recovery, and empirically tied to multimodal behavior. Code is publicly available at https://github.com/he-h/vlm-graph-probing.