Structural Graph Probing of Vision-Language Models
On the map: Multimodal LLMs / vision-language models → VLM benchmarks and visual reasoning
Abstract (arXiv)
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.
Key facts. No citing papers in the map yet (data through 13 Oct 2026). Placed in VLM benchmarks and visual reasoning, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 42 of its 51 references that are in the map. First posted on arXiv 28 Mar 2026.
Builds on
42 of 51 references.
Closest ancestors · 3
Same direction, newest first
- 2024Towards Interpreting Visual Information Processing in Vision-Language Models9 Oct
- 2024LVLM-Interpret: An Interpretability Tool for Large Vision-Language Models3 Apr
- 2022Flamingo: a Visual Language Model for Few-Shot Learning29 Apr
Other references · 39
- 2014Adam: A Method for Stochastic Optimizationcited by 37,093
- 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,588
- 2023Visual Instruction Tuningcited by 7,827
- 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,713
- 2023BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Modelscited by 6,103
- 2025Qwen2.5-VL Technical Reportcited by 5,499
- 2013Efficient Estimation of Word Representations in Vector Spacecited by 4,360
- 2022Locating and Editing Factual Associations in GPTcited by 2,650
- 2015VQA: Visual Question Answeringcited by 2,495
- 2023MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGIcited by 2,057
- 2025InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Modelscited by 1,798
- 2019Similarity of Neural Network Representations Revisitedcited by 1,719
27 more
- 2019LXMERT: Learning Cross-Modality Encoder Representations from Transformerscited by 1,551
- 2016CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoningcited by 1,450
- 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,391
- 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,280
- 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 850
- 2019Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Prunedcited by 849
- 2023EVA-CLIP: Improved Training Techniques for CLIP at Scalecited by 792
- 2020Quantifying Attention Flow in Transformerscited by 767
- 2024BLINK: Multimodal Large Language Models Can See but Not Perceivecited by 534
- 2024MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmarkcited by 502
- 2019Graph Matching Networks for Learning the Similarity of Graph Structured Objectscited by 205
- 2025Can MLLMs Reason in Multimodality? EMMA: An Enhanced MultiModal ReAsoning Benchmarkcited by 132
- 2024A Practical Review of Mechanistic Interpretability for Transformer-Based Language Modelscited by 116
- 2021Inducing Causal Structure for Interpretable Neural Networkscited by 108
- 2017An Analysis of Visual Question Answering Algorithmscited by 90
- 2024Unified Hallucination Detection for Multimodal Large Language Modelscited by 75
- 2022VL-InterpreT: An Interactive Visualization Tool for Interpreting Vision-Language Transformerscited by 53
- 2024Explainable and Interpretable Multimodal Large Language Models: A Comprehensive Surveycited by 52
- 2024A Survey on Evaluation of Multimodal Large Language Modelscited by 46
- 2023Emergent Modularity in Pre-trained Transformerscited by 37
- 2024The LLM Language Network: A Neuroscientific Approach for Identifying Causally Task-Relevant Unitscited by 34
- 2025A Survey on Mechanistic Interpretability for Multi-Modal Foundation Modelscited by 34
- 2025Representation Engineering for Large-Language Models: Survey and Research Challengescited by 19
- 2016Small-World Brain Networks Revisitedcited by 16
- 2024How Interpretable Are Interpretable Graph Neural Networks?cited by 13
- 2025Brain-Inspired Exploration of Functional Networks and Key Neurons in Large Language Modelscited by 4
- 2025Probing Neural Topology of Large Language Modelscited by 4
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2025Uncovering Brain-Like Hierarchical Patterns in Vision-Language Models through fMRI-Based Neural Encoding
- 2026Nodes Are Early, Edges Are Late: Probing Diagram Representations in Large Vision-Language Models
- 2025Deciphering Functions of Neurons in Vision-Language Models
- 2026Do LLMs and VLMs Share Neurons for Inference? Evidence and Mechanisms of Cross-Modal Transfer
- 2026Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits
- 2025Sparse Neuron Ablation Triggers Catastrophic Collapse of the Language Core in Large Vision-Language Models
- 2026Belief or Circuitry? Causal Evidence for In-Context Graph Learning
- 2025Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs