Structural Graph Probing of Vision-Language Models
Provides a new intermediate-scale interpretability method for vision-language models, bridging local attribution and full circuit recovery.
The authors propose using within-layer correlation graphs from neuron co-activations to study population-level structure in vision-language models, finding that cross-modal structure consolidates around hub neurons whose perturbation alters model output, offering a tractable interpretability approach.
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.