Structural Graph Probing of Vision-Language Models

1. Problem Statement

Vision-Language Models (VLMs) have achieved remarkable performance in multimodal tasks, yet the internal organization of their computation remains poorly understood. Existing interpretability methods primarily focus on local signals (e.g., attention maps, saliency, patch attribution) or isolated units (individual neurons/tokens). These approaches fail to capture the large-scale, population-level organization through which multimodal reasoning is realized. The authors argue that complex computation in transformers is distributed across interacting populations rather than concentrated in isolated pathways, necessitating a shift from local attribution to analyzing the structural topology of neuron interactions.

2. Methodology: Neural Topology

The authors propose a framework called Neural Topology, which treats each transformer layer as a weighted correlation graph derived from neuron co-activations during inference.

A. Graph Construction

• Input: For a given image-text pair, hidden activations $H^{(\ell)}$ are recorded at each layer $\ell$.
• Nodes: Each neuron in the hidden dimension $d$ corresponds to a node in the graph ($|V| = d$).
• Edges: The edge weight $W_{ij}^{(\ell)}$ between neurons $i$ and $j$ is defined as the Pearson correlation of their activation profiles across all tokens in the multimodal context.
  $$W_{ij}^{(\ell)} = \text{corr}(H_{i,:}^{(\ell)}, H_{j,:}^{(\ell)})$$
• Modality-Specific Graphs: To analyze cross-modal dynamics, the authors construct three types of graphs per layer:
  1. Multimodal ($G^{(\ell)}$): Based on the full hidden state.
  2. Vision ($G_{vis}^{(\ell)}$): Based only on visual token activations.
  3. Text ($G_{text}^{(\ell)}$): Based only on textual token activations.

B. Graph Representation & Probing

• Node Embedding: To avoid exposing raw activation magnitudes or token semantics, neurons are assigned unique learnable one-hot identity embeddings.
• Graph Neural Network (GCN): A GCN processes the correlation graph to produce node representations $Z^{(\ell)}$.
• Layer Signature: Global pooling (Mean and Max) aggregates node representations into a fixed-dimensional structural signature $h^{(\ell)}$ for each layer.
• Probing Tasks: These signatures are used to train linear and GCN probes to predict:
  • Task outputs (e.g., object counting, color recognition).
  • Hallucination status (grounded vs. hallucinated responses).

C. Causal Intervention

To test causality, the authors perform targeted perturbations:

• Neuron Ablation: Zeroing out the top 1% of neurons selected by graph degree vs. activation magnitude.
• Edge Intervention: Replacing the activation of one endpoint of a high-degree edge with its partner's activation (Identical), negated partner (Opposite), or random noise.
• Hub Scaling: Directly scaling the activation of identified "hub" neurons to observe performance degradation.

3. Key Contributions

1. Neural Topology as an Interpretability Lens: Introduces a novel intermediate scale of analysis (between local attribution and full circuit recovery) that treats VLM layers as correlation graphs.
2. Behavioral Predictability: Demonstrates that graph-based probes significantly outperform linear baselines in predicting model behavior, proving that correlation topology carries recoverable behavioral signals.
3. Cross-Modal Structural Dynamics: Reveals that multimodal integration is not uniform; it progressively consolidates in deeper layers around a compact set of recurrent "hub" neurons.
4. Causal Identification: Identifies structurally central neurons (hubs) whose targeted perturbation causes substantial performance degradation, validating that topology identifies causally influential components.

4. Key Results

A. Predictability

• Superior Performance: Graph-based probes (GCN) consistently outperformed linear baselines across multiple VLMs (InternVL3, Qwen2.5-VL, LLaVA) and datasets (CLEVR, TDIUC, MMMU).
  • Example: On CLEVR counting, GCN improved accuracy by 7.7% over linear baselines for LLaVA-1.5-7B.
• Regression: Graph probes also improved regression metrics (MSE, $R^2$) for object counting, indicating they capture functional dependencies beyond discrete classification.
• Hallucination Detection: Graph probes achieved significantly higher accuracy (e.g., 91.0% for Qwen2.5-VL) in detecting hallucinations compared to text-only baselines (~65%), proving that hallucination is linked to structural correlation patterns.

B. Structural Dynamics

• Depth Evolution: Cross-modal (Vision-Text) correlations increase with depth, while Vision-Vision correlations remain flat. This suggests later layers are where multimodal integration solidifies.
• Hub Stability: Neurons identified as "hubs" (high degree in the correlation graph) show high recurrence across different samples and inputs, particularly in intermediate layers. This indicates that multimodal processing is organized around persistent topological loci rather than being uniformly distributed.
• Graph Alignment: Multimodal graphs (Image+Text) align better with each other than unimodal graphs (Text vs. Image) or text graphs from unimodal backbones (LLaMA), suggesting multimodal fine-tuning creates a distinct, partially aligned structural space.

C. Causal Intervention

• Ablation: Removing neurons selected by graph degree caused significantly larger performance drops than removing neurons selected by activation magnitude.
• Edge Perturbation: Replacing an edge endpoint with the "Opposite" (negated) partner activation caused the most severe performance degradation, indicating that the sign and alignment of coordinated activity are critical.
• Hub Sensitivity: Small perturbations (scaling) of identified hub neurons resulted in symmetric performance degradation (both amplification and suppression hurt performance), confirming these hubs operate within a narrow, critical functional range.

5. Significance and Implications

• New Interpretability Paradigm: The paper argues that VLM interpretability should move beyond "which token is important" to "how populations of neurons are organized."
• Efficiency: The approach is computationally tractable (using sparse graphs) and does not require full circuit recovery, making it applicable to large-scale models.
• Mechanistic Insight: It provides evidence that multimodal reasoning is an emergent property of organized neural populations centered around stable hubs, rather than a simple sum of local effects.
• Practical Utility: The ability to identify causally influential hubs via topology offers a new method for model debugging, safety interventions, and understanding failure modes (like hallucinations).

In summary, this work establishes neural topology as a powerful, behaviorally meaningful, and causally relevant framework for understanding the internal mechanics of Vision-Language Models.