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.
Figures & tables
Figure 1: Overview of neural topology construction.
InternVL3-1B
Qwen2.5-VL-3B
LLaVA-1.5-7B
Dataset
Linear
GCN
Linear
GCN
Linear
GCN
Acc
F1
Acc
F1
Acc
F1
Acc
F1
Acc
F1
Acc
F1
TDIUC
0.884
0.846
0.965
0.959
0.943
0.930
0.976
0.970
0.971
0.966
0.954
0.947
CLEVR
0.980
0.980
0.993
0.993
0.920
0.919
0.963
0.963
0.602
0.563
0.679
0.676
MMMU
0.293
0.253
0.321
0.253
0.293
0.211
0.336
0.335
0.314
0.300
0.279
0.162
MMMU-Pro
0.309
0.226
0.359
0.260
0.286
0.164
0.318
0.290
0.291
0.171
0.314
0.305
Table 1: Graph probing across multimodal benchmarks. Accuracy and F1 of linear vs. graph-based probes trained on layer-wise neuron-correlation topology for TDIUC, CLEVR, MMMU, MMMU-Pro, BLINK, and EMMA. Best result in each model–dataset pair is in bold . Graph-based probes outperform linear baselines in most settings, indicating that population-level connectivity carries task-relevant information. See Table 5 for Precision/Recall.
Model
Probe
MSE ↓
R2↑
Pearson ↑
InternVL3-1B
Linear
0.020
0.996
0.998
GCN
0.007
0.999
0.999
Qwen2.5-VL-3B
Linear
0.081
0.985
0.992
GCN
0.038
0.993
0.996
LLaVA-1.5-7B
Linear
0.605
0.884
0.949
GCN
0.379
0.928
0.963
Table 2: Regression on object counting from neural topology. Linear and graph-based probes on CLEVR counting. Graph-based probes improve regression performance across all three VLMs.
Figure 2: Sparsity robustness and depth dependence of graph probing. Left: probing accuracy as a function of graph sparsity (top 1%–20% of neuron correlations retained). Right: probing accuracy across normalized layer depth. Accuracy is stable across sparsity levels, while depth-wise peak predictiveness differs by architecture.
Method
InternVL3-1B
Qwen2.5-VL-3B
LLaVA-1.5-7B
Emb. Avg.
0.664
0.654
0.649
Length
0.500
0.633
0.642
GCN
0.789
0.910
0.908
Table 3: Hallucination detection from neural topology on MHaluBench. Accuracy of a graph-based probe versus two text-only baselines (mean word2vec embedding and token length) for binary hallucination classification. Graph-based probes outperform both baselines across all models.
Figure 3: Token-level cross-modal correlation dynamics across depth. Layer-wise token–token correlations for Vision–Vision, Vision–Text, and Text–Text pairs on TDIUC (mean ± std) across multiple VLM families and scales. Vision–Text correlations increase with depth, consistent with progressively stronger multimodal integration in later layers.
Figure 4: Cross-sample stability of hub definitions. Recurrence of top 1% hubs across samples on TDIUC for graph-wide, modality-specific, and activation-based hub definitions. Graph-derived hubs are the most stable, indicating that structurally central neurons occupy more persistent roles than alternative hub candidates.
Figure 5: Layer-wise stability of graph hubs. Recurrence of top 1% graph-derived hubs across samples at different depths on TDIUC. Intermediate layers show the strongest hub stability, suggesting the most persistent population-level organization emerges in mid-depth representations.
Ω Modality
Γ Modality
GAUC
Image+Text
Image+Text
0.960
Text
Image
0.819
LLaMA (Text)
LLaVA (Text)
0.680
Table 4: Cross-modal graph alignment. GAUC at layer 6 of LLaVA-1.5-7B for graph embeddings matched across modalities. Higher GAUC indicates stronger structural correspondence.
Figure 6: Ablation by zeroing selected neurons. Accuracy on TDIUC and CLEVR after zeroing the top 1% of neurons selected per sample by random choice, activation magnitude, or graph degree for InternVL3-1B (a, layer 11) and Qwen2.5-VL-3B (b, layer 0). Zeroing graph-selected neurons yields the largest drop.
Figure 7: Edge-level intervention on top-ranked neuron pairs. For the strongest dataset-level graph-defined neuron pair, one endpoint is replaced by its partner’s activation (Identical), the negated partner activation (Opposite), or a random matched-shape vector (Random), and downstream accuracy is measured on color and counting tasks for InternVL3-1B and Qwen2.5-VL-3B. Opposite intervention causes the largest performance drop.
Figure 8: Direct perturbation of topology-defined hub neurons. Selected high-degree hub neurons in InternVL3-1B (neuron 62, layer 11) and Qwen2.5-VL-3B (neurons 71, 318, 294, 528, 583, layer 0) are scaled and evaluated on color and counting tasks. Performance degrades under both positive and negative perturbation.
InternVL3-1B
Qwen2.5-VL-3B
LLaVA-1.5-7B
Dataset
Linear
GCN
Linear
GCN
Linear
GCN
Prec
Rec
Prec
Rec
Prec
Rec
Prec
Rec
Prec
Rec
Prec
Rec
TDIUC
0.879
0.839
0.961
0.957
0.932
0.929
0.970
0.970
0.964
0.967
0.945
0.949
CLEVR
0.980
0.980
0.993
0.993
0.919
0.919
0.962
0.963
0.619
0.594
0.712
0.684
MMMU
0.270
0.288
0.231
0.300
0.226
0.279
0.337
0.337
0.320
0.324
0.389
0.261
MMMU-Pro
0.368
0.280
0.272
0.323
0.157
0.256
0.299
0.300
0.224
0.260
0.306
0.306
Table 5: Precision and recall for graph probing across multimodal benchmarks. Companion to Table 1 , reporting Precision and Recall for linear and graph-based probes on TDIUC, CLEVR, MMMU, MMMU-Pro, BLINK, and EMMA. Best result in each model–dataset pair is in bold .
Vision-language models (VLMs) are powerful but remain opaque black boxes. We introduce the first framework for transparent circuit tracing in VLMs to systematically analyze multimodal reasoning. By utilizing transcoders, attribution graphs, and attention-based methods, we uncover how VLMs hierarchically integrate visual and semantic concepts. We reveal that distinct visual feature circuits can handle mathematical reasoning and support cross-modal associations. Validated through feature steering and circuit patching, our framework proves these circuits are causal and controllable, laying the groundwork for more explainable and reliable VLMs.
Jingcheng Yang, Tianhu Xiong, Shengyi Qian +2
University of Illinois Urbana-Champaign · Independent Researcher
Recent work has increasingly explored neuron-level interpretation in vision-language models (VLMs) to identify neurons critical to final predictions. However, existing neuron analyses generally focus on single tasks, limiting the comparability of neuron importance across tasks. Moreover, ranking strategies tend to score neurons in isolation, overlooking how task-dependent information pathways shape the write-in effects of feed-forward network (FFN) neurons. This oversight can exacerbate neuron polysemanticity in multi-task settings, introducing noise into the identification and intervention of task-critical neurons. In this study, we propose HONES (Head-Oriented Neuron Explanation & Steering), a gradient-free framework for task-aware neuron attribution and steering in multi-task VLMs. HONES ranks FFN neurons by their causal write-in contributions conditioned on task-relevant attention heads, and further modulates salient neurons via lightweight scaling. Experiments on four diverse multimodal tasks and two popular VLMs show that HONES outperforms existing methods in identifying task-critical neurons and improves model performance after steering. Our source code is released at: https://github.com/petergit1/HONES.
Qidong Wang, Junjie Hu, Ming Jiang
Tongji University · University of Wisconsin-Madison
Vision-language models can describe an image with remarkable accuracy, yet a more fundamental question remains unanswered: what visual information actually drives their answers? In this work, we investigate this question through causal tracing, and we observe that highly causal vision tokens often lie outside the target region. Extending the analysis to larger vision-language models reveals a similar pattern across models and corruption settings, suggesting that strong multimodal performance does not necessarily imply spatially localized causal representations. We further investigate: can these models preserve visual structure when appearance cues are removed? and find that visual cues are exploited to understand visual structures. Together, our experiments expose a gap between seeing, using, and reasoning over visual structure, and provide a causal framework for studying how visual information is transformed, preserved, and ultimately used by modern vision-language models.
Naren Kumar S, Tirth Bhatt, Mayank Singh
LINGO Research Group, Indian Institute of Technology Gandhinagar, India