Structural Graph Probing of Vision–Language Models

Paper | PDF

Haoyu He, Yue Zhuo, Yu Zheng, Qi R. Wang IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026

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.

Key Findings

• Topology encodes behavior. Neuron-neuron correlation graphs carry signal predictive of VQA correctness — GCN classifiers trained on these graphs outperform MLP baselines using last-token activations alone.
• Cross-modal consolidation. Visual and textual representations progressively align with depth; visual–textual (VT) correlation strengthens in later layers across all tested VLMs.
• Hub neurons are causally influential. A small set of high-degree neurons recur consistently across samples and layers; ablating them substantially degrades model accuracy, confirming their causal role.
• Topology is architecture-agnostic. Findings hold consistently across LLaVA, Qwen2.5-VL, InternVL3, and Gemma3 model families.

Repository Layout

Core modules:

• model.py — model loading, hidden-state extraction, correlation graph construction, GCNPredictor / MLPPredictor
• dataset.py — dataset loading and VQA prompt construction
• utils.py — shared helpers, constants, and text sanitization utilities

Experiment scripts (under probing/):

Script	Description
extract_graphs.py	Extract hidden states, build sparse correlation graphs, save per-layer artifacts
predict_gcn.py	Train GCNPredictor on extracted graphs to probe behavioral signal in topology
degree_analysis.py	Layer-wise degree and last-token activation analysis across all layers
hub_neurons.py	Identify hub neurons by degree / activation frequency across samples
intervene_neuron.py	Neuron-level ablation / scaling intervention experiments
intervene_edge.py	Edge-level (neuron-pair) intervention experiments
modality_corr.py	Cross-modality (visual-text) token correlation analysis

Runner scripts: scripts/ — ready-to-run shell scripts for each pipeline step (start here).

Environment Setup

conda create -n vlm-graph-probing python=3.10
conda activate vlm-graph-probing
pip install -r requirements.txt

Notes:

• A working PyTorch installation compatible with your CUDA setup is required for GPU experiments.
• Some Hugging Face model checkpoints may require authentication or acceptance of model terms.

Supported Models

Model	Hugging Face checkpoint
LLaVA-1.5-7B	llava-hf/llava-1.5-7b-hf
LLaVA-1.5-13B	llava-hf/llava-1.5-13b-hf
LLaVA-v1.6-Mistral	llava-hf/llava-v1.6-mistral-7b-hf
Qwen2.5-VL-3B	Qwen/Qwen2.5-VL-3B-Instruct
Qwen2.5-VL-7B	Qwen/Qwen2.5-VL-7B-Instruct
InternVL3-1B	OpenGVLab/InternVL3-1B-hf
InternVL3-2B	OpenGVLab/InternVL3-2B-hf
InternVL3-8B	OpenGVLab/InternVL3-8B-hf

See utils.py for the complete list of supported checkpoints.

Data Setup

Dataset	Source	Setup
CLEVR	HuggingFace	Loaded automatically via datasets (laion/clevr-webdataset). No manual download needed.
COCO	HuggingFace	Loaded automatically via datasets (lmms-lab/COCO-Caption2017). No manual download needed.
TDIUC	Manual download	Download and place under data/TDIUC/. Also requires COCO val2014 images under data/val2014/.

For TDIUC, the expected directory structure is:

data/
├── TDIUC/
│   ├── Questions/
│   │   └── OpenEnded_mscoco_val2014_questions.json
│   └── Annotations/
│       └── mscoco_val2014_annotations.json
└── val2014/
    └── COCO_val2014_*.jpg

You can override dataset locations via CLI flags or environment variables:

export VLM_GRAPH_DATA_ROOT=/path/to/data
export VLM_GRAPH_TDIUC_ROOT=/path/to/data/TDIUC
export VLM_GRAPH_COCO_VAL_ROOT=/path/to/data/val2014

Quickstart

Ready-to-run scripts live under scripts/. Each script has configurable variables (model, dataset, device, etc.) at the top. Run them in order:

# Step 1: Extract hidden states → per-layer correlation graphs + last-token activations
bash scripts/01_extract_graphs.sh

# Step 2: Train GCNPredictor → probes whether graph topology predicts VQA correctness
bash scripts/02_predict_gcn.sh

# Step 3: Layer-wise degree & activation statistics → avg_node_degree / avg_last_token .npy
bash scripts/03_degree_analysis.sh

# Step 4: Hub neuron identification → hub_neurons_<model>_<dataset>_<category>_top<k>.json
bash scripts/04_hub_neurons.sh

# Step 5: Neuron-level intervention → accuracy delta vs. baseline (requires step 4 output)
bash scripts/05_intervene_neuron.sh

# Step 6: Edge-level intervention → top-k edge ablation accuracy results
bash scripts/06_intervene_edge.sh

# Step 7: Cross-modality correlation → per-layer VV / TT / VT correlation .npz files
bash scripts/07_modality_corr.sh

All scripts default to InternVL3-1B on CLEVR color with cuda:0. Edit the variables at the top of each script to change the model, dataset, category, or GPU device.

Citation

If you use this code in your research, please cite our paper:

@inproceedings{he2026structural,
  title={Structural Graph Probing of Vision-Language Models},
  author={He, Haoyu and Zhuo, Yue and Zheng, Yu and Wang, Qi R.},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}

License

This project is licensed under the MIT License. See LICENSE for details.