Structural Graph Probing of Vision-Language Models

🕒 Published (v1): 2026-03-28 01:14 UTC · Source: Arxiv · Venue: CVPR · link

Why this paper was selected

CVPR; structural graph probing reveals neural topology organization in VLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper introduces neural topology probing for VLMs, representing each transformer layer as a weighted neuron–neuron Pearson correlation graph and using GCN-based probes to analyze population-level structure. Graph-based probes outperform linear baselines in predicting model behavior and hallucination, cross-modal correlation strengthens with depth, and topology-defined hub neurons are more causally influential than activation-magnitude-ranked neurons.

Problem

Existing VLM interpretability methods focus on local attribution signals (attention patterns, saliency maps, patch attribution), which are poorly suited to characterizing large-scale computational organization across distributed neuron populations. How multimodal computation is structured at the population level — how visual and linguistic representations couple, stabilize, and govern model output — remains largely uncharacterized.

Method

For a frozen VLM given an image–question pair, hidden activations \(H^{(\ell)} \in \mathbb{R}^{d \times N}\) are recorded at each layer \(\ell\). A weighted within-layer correlation graph \(G^{(\ell)} = (V, E, W^{(\ell)})\) is constructed, where edge weight \(W^{(\ell)}_{ij} = \text{corr}(H^{(\ell)}_{i,:}, H^{(\ell)}_{j,:})\) is the Pearson correlation between neuron \(i\) and neuron \(j\) across all tokens. Three variants are constructed per layer (full multimodal, vision-only, text-only) by partitioning tokens by positional indices.

Graphs are sparsified by retaining only the top-\(k\%\) of edge weights. Neurons receive learnable one-hot identity embeddings (not raw activations), and a GCN propagates structure: $\(Z^{(\ell)} = \sigma\!\left(D^{-\frac{1}{2}} W^{(\ell)} D^{-\frac{1}{2}} X W_g\right)\)$ A layer-level structural signature is \(h^{(\ell)} = \text{Concat}(\text{Mean}(Z^{(\ell)}), \text{Max}(Z^{(\ell)}))\). Linear and GCN probes are trained on these signatures for behavioral prediction. Hub neurons are defined by weighted degree \(d^{(\ell)}_i = \sum_j |W^{(\ell)}_{ij}|\); cross-modal graph alignment is trained via a symmetric InfoNCE objective over graph-level GCN embeddings with cosine similarity.

Key Contributions

• A neural topology framework for VLM interpretability: layer-wise neuron correlation graphs as the primary analytical object.
• Empirical demonstration that GCN probes on correlation topology outperform linear probes for behavioral prediction (grounded reasoning, semantic recognition, object counting regression) and hallucination detection.
• Characterization of cross-modal correlation dynamics: vision–text token correlations increase monotonically with depth while vision–vision correlations remain flat.
• Evidence that topology-defined hub neurons (by degree) recur more stably across inputs than activation-magnitude-ranked or modality-specific hubs, particularly in middle layers.
• Causal ablation showing that zeroing degree-ranked top-1% neurons degrades accuracy more than zeroing activation-ranked neurons.
• Edge-level intervention protocol: replacing one endpoint of the highest-degree edge with the negated partner activation (OPPOSITE) causes the largest accuracy drop, indicating graph-defined edges encode functionally meaningful pairwise relations.
• Cross-modal alignment analysis: multimodal–multimodal GAUC = 0.9598; text–image within LLaVA = 0.8188; LLaVA text vs. LLaMA text = 0.6803, showing multimodal training partially but not fully aligns visual and textual topology.

Results

• Graph probing accuracy (TDIUC, CLEVR): GCN consistently outperforms linear baselines; on CLEVR counting, GCN improves over linear by 7.7% (LLaVA-1.5-7B), 4.3% (Qwen2.5-VL-3B), 1.3% (InternVL3-1B).
• CLEVR regression: GCN reduces MSE from 0.020→0.007 (\(R^2\): 0.996→0.999) for InternVL3-1B; 0.081→0.038 (\(R^2\): 0.985→0.993) for Qwen2.5-VL-3B; 0.605→0.379 (\(R^2\): 0.884→0.928) for LLaVA-1.5-7B.
• Hallucination detection (MHaluBench): GCN achieves 0.789/0.910/0.908 accuracy on InternVL3/Qwen2.5/LLaVA vs. word2vec baseline 0.664/0.654/0.649.
• Cross-modal alignment (GAUC at LLaVA layer 6): 0.9598 (multimodal–multimodal), 0.8188 (text–image), 0.6803 (LLaVA-text vs. LLaMA-text).
• Hub neuron ablation: Zeroing graph-degree-ranked top-1% neurons causes up to −85.7% accuracy drop on CLEVR Color (Qwen2.5-VL-3B), exceeding activation-ranked ablation.
• Sparsity robustness: Probe accuracy is stable from 1% to 20% edge density, indicating task-relevant signal concentrates in the strongest correlations.

Limitations

• Correlation graphs reflect co-activation, not causal wiring; the method cannot distinguish mechanism from co-occurrence.
• Graph-level probes require per-dataset training; out-of-distribution generalization is not systematically evaluated.
• Layerwise patterns (e.g., mid-layer hub stability, depth-wise predictability peaks) are described as descriptive rather than mechanistic — causal localization of multimodal fusion is not claimed.
• Hallucination detection results are framed as evidence of informativeness, not a competitive system.
• Computational scale: each layer of a large VLM yields \(O(d^2)\) neuron pairs; sparse approximation is practical but may discard structurally relevant weak correlations.
• Analysis covers three relatively small VLMs (1B–7B); behavior at larger scales is untested.

Relevance to Vision-Language Models

This work opens a new interpretability axis for VLMs by moving from token-level attribution to population-level interaction structure, filling a gap between local saliency methods and full circuit recovery. The finding that topology encodes hallucination-predictive signal directly informs efforts to diagnose and mitigate grounding failures in VLMs without accessing output probabilities. The depth-wise cross-modal coupling analysis provides a concrete empirical account of how visual and linguistic representations become integrated across transformer layers, which is foundational for architectural design and alignment research. The hub-neuron intervention results suggest that structural centrality — not activation magnitude — is the better criterion for identifying causally influential components, with implications for pruning, editing, and steering VLMs.