NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning

Ruiyu Yan; Bowen Chen; Shaowen Wan; Lin Zhao

cs.AISep 29, 2026

NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning

Authors: Ruiyu Yan, Bowen Chen, Shaowen Wan, Lin Zhao

Organizations: New York University · New Jersey Institute of Technology

Abstract

Current vision-language models (VLMs) encode visual information in dense hidden states where object identity, spatial layout, and local attributes are implicitly entangled rather than explicitly disentangled, limiting their ability to isolate and modulate the specific visual evidence required by a given language query. Inspired by sparse population coding and top-down modulation in biological vision, we introduce NeuronEye, a plug-in framework that constructs a sparse, concept-level neuron vocabulary from intermediate VLM representations and selectively activates query-relevant visual concepts during inference. NeuronEye decomposes vision-token states into an overcomplete sparse basis organized by concept-level clusters, uses the language query to activate relevant clusters and localize the patches where selected concepts are expressed, and injects the focused evidence back into vision tokens. A complementary suppression mechanism attenuates dominant perceptual directions to preserve weaker but relevant cues. All operations run in a single forward pass over a frozen VLM backbone. On Qwen2.5-VL-7B, NeuronEye raises CV-Bench overall accuracy by +3.1 with gains of +9.5 on Distance, and improves BLINK Multi-view by +8.3, with similar trends on LLaVA-1.6-7B. These results suggest that sparse neuron vocabularies can serve not only as post-hoc interpretability tools but also as active interfaces for concept-level visual reasoning.

Figures & tables

Figure 1 : NeuronEye overview. Given an image and a language query, NeuronEye decomposes intermediate VLM representations into a sparse neuron vocabulary and identifies query-relevant concept clusters (e.g., Human and Apparel ). Each cluster localizes its corresponding visual evidence in the image, and the focused activations are injected back into the VLM to guide reasoning.

Figure 2 : Construction of Sparse Neuron Space. Intermediate vision-token states are encoded by a sparse autoencoder, whose activated neurons are visualized through top-activating patches and grouped into concept-level neuron clusters.

Figure 3 : NVF and PCS. NVF selects query-relevant neuron cluster indices and active patches for local cross-attention, while PCS suppresses dominant perceptual components at a later layer.

Representative VLM backbones
	CV-Bench					BLINK			Other Benchmarks
Model	Overall	Count	Depth	Dist.	Rel.	Overall	MV.	Loc.	RealWorldQA	MMStar
DeepSeek-VL1 Lu et al. (2024)	61.6	59.0	63.2	58.2	68.5	38.1	50.4	37.7	50.5	38.9
Idefics3-8B-Llama3 Laurençon et al. (2024)	67.7	60.5	72.8	67.2	73.4	42.7	45.9	50.8	62.0	49.3
Phi-4 Multimodal Abouelenin et al. (2025)	71.7	68.5	74.2	70.8	75.1	49.7	48.1	56.6	61.8	59.7
InternVL3-8B Zhu and others (2025)	81.3	70.9	84.8	83.1	89.7	51.3	51.1	58.2	68.2	59.4

Table 1 : Benchmark results across VLM backbones. The upper block lists representative VLMs for reference. The lower block shows NeuronEye applied to LLaVA-1.6-7B and Qwen2.5-VL-7B, with

\Delta

denoting the improvement over each base model.

Vision-token reduction methods
	CV-Bench					BLINK			Other Benchmarks
Model	Overall	Count	Depth	Dist.	Rel.	Overall	MV.	Loc.	RealWorldQA	MMStar
FastV Chen et al. (2024a)	75.9	64.0	83.5	74.5	85.5	48.4	55.6	49.2	68.8	55.1
SparseVLM Zhang et al. (2025b)	69.8	54.3	74.7	68.8	86.3	46.9	55.6	59.0	50.9	53.8
PruMerge Shang et al. (2025)	71.6	55.2	78.0	74.0	83.2	46.8	54.9	51.6	64.3	50.6
MustDrop Liu et al. (2024b)	73.7	58.6	81.5	74.8	83.9	47.2	55.6	52.5	65.1	53.0

Table 2 : Comparison with visual representation modulation methods on Qwen2.5-VL-7B. All methods are evaluated under the same protocol.

Panel of Figure 4 — Figure 4 : Neuron-level feature visualization. Each subfigure shows top-activating patches for one SAE neuron, with non-activating regions masked.

Panel of Figure 5 — Figure 5 : Question-guided visual focus examples. Each shows the input image, overall focus map, and individual neuron cluster activations.

	CV-Bench					BLINK
Setting	Overall	Count	Depth	Dist.	Rel.	Overall	MV.	Loc.
Baseline (Qwen2.5-VL-7B)	78.5	67.1	87.2	76.2	87.2	52.1	55.6	53.3
Baseline + FT	76.5 $\pm$ 0.3	66.6 $\pm$ 0.2	86.7 $\pm$ 0.3	69.8 $\pm$ 0.9	86.5 $\pm$ 0.6	51.1 $\pm$ 0.1	55.6 $\pm$ 1.9	55.7 $\pm$ 0.4
+ Dense Cross-Attn + PCS	79.4 $\pm$ 0.3	66.2 $\pm$ 0.6	87.0 $\pm$ 0.4	79.5 $\pm$ 1.2	88.8 $\pm$ 0.3	51.3 $\pm$ 0.3	55.6 $\pm$ 1.9	54.9 $\pm$ 1.8
+ Random SNS + NVF	78.6 $\pm$ 0.4	67.6 $\pm$ 0.5	86.5 $\pm$ 0.6	76.6 $\pm$ 1.0	86.4 $\pm$ 0.6	51.6 $\pm$ 0.3	55.6 $\pm$ 1.5	54.9 $\pm$ 1.0
+ PCS only	78.6 $\pm$ 0.4	67.3 $\pm$ 0.4	86.8 $\pm$ 0.3	76.8 $\pm$ 1.4	87.2 $\pm$ 0.6	51.3 $\pm$ 0.2	55.6 $\pm$ 1.9	54.9 $\pm$ 0.4

Table 3 : Ablation study on component contributions.

	CV-Bench					BLINK
Layer $\ell$	Overall	Count	Depth	Dist.	Rel.	Overall	MV.	Loc.
4	80.5 $\pm$ 0.1	67.3 $\pm$ 0.2	87.0 $\pm$ 0.6	83.0 $\pm$ 0.4	88.5 $\pm$ 0.3	51.6 $\pm$ 0.3	57.2 $\pm$ 1.2	56.6 $\pm$ 1.1
8	81.6 $\pm$ 0.3	67.9 $\pm$ 0.6	87.0 $\pm$ 0.4	85.7 $\pm$ 0.8	89.5 $\pm$ 0.5	52.8 $\pm$ 0.2	63.9 $\pm$ 0.4	56.6 $\pm$ 0.9
12	78.3 $\pm$ 0.4	65.2 $\pm$ 0.8	85.7 $\pm$ 0.8	79.5 $\pm$ 1.2	86.3 $\pm$ 0.3	51.3 $\pm$ 0.2	55.6 $\pm$ 0.1	54.1 $\pm$ 0.1
16	79.0 $\pm$ 0.3	67.7 $\pm$ 0.6	87.2 $\pm$ 0.2	76.9 $\pm$ 0.5	87.7 $\pm$ 0.6	51.8 $\pm$ 0.5	56.9 $\pm$ 1.9	56.3 $\pm$ 0.5
20	78.6 $\pm$ 0.4	66.8 $\pm$ 0.3	86.8 $\pm$ 0.4	77.0 $\pm$ 1.1	86.9 $\pm$ 0.6	51.7 $\pm$ 0.1	55.6 $\pm$ 1.9	55.7 $\pm$ 1.3

Table 4 : Effect of insertion layer. Layer 8 achieves the best results, suggesting a balance between low-level perceptual features and high-level semantic abstraction.

	CV-Bench					BLINK
Clusters	Overall	Count	Depth	Dist.	Rel.	Overall	MV.	Loc.
32	80.4 $\pm$ 0.2	68.4 $\pm$ 0.5	86.1 $\pm$ 0.4	82.1 $\pm$ 1.2	88.2 $\pm$ 0.5	52.0 $\pm$ 0.2	57.2 $\pm$ 1.5	53.9 $\pm$ 2.6
64	81.6 $\pm$ 0.3	67.9 $\pm$ 0.6	87.0 $\pm$ 0.4	85.7 $\pm$ 0.8	89.5 $\pm$ 0.5	52.8 $\pm$ 0.2	63.9 $\pm$ 0.4	56.6 $\pm$ 0.9
128	80.3 $\pm$ 0.1	66.4 $\pm$ 1.1	87.1 $\pm$ 0.4	82.8 $\pm$ 0.7	88.1 $\pm$ 0.4	51.7 $\pm$ 0.3	57.8 $\pm$ 1.7	58.3 $\pm$ 1.3

Table 5 : Effect of the number of neuron clusters K. Too few clusters conflate functionally distinct neurons, while too many fragment coherent functional groups and complicate routing.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6 : Offline prompts used for sparse neuron interpretation and cluster-index supervision.

Component	Hyperparameter	Value
Sparse Neuron Space	Expansion ratio $\alpha$	32
	Sparsity level $k$	32
	Sparsity coefficient $\lambda_{s}$	0.05
	Min. image threshold $M_{\min}$	3
Neuron-guided Visual Focus	Top- $n$ patches per cluster	60
	Top- $K_{\mathrm{sel}}$ clusters	5

Appendix

Table 6 : Hyperparameter settings for all experiments.

Panel of Figure 7 — Figure 7 : Additional sparse neuron visualizations. Each subfigure shows the top-activating patches for a single SNS neuron across multiple images, with non-activating regions masked.

Panel of Figure 8 — Figure 8 : Additional question-guided visual focus examples.

Evaluation Setting	Base Qwen2.5-VL	NeuronEye	Difference
Overall Accuracy	65.30%	63.10%	-2.20
Anatomy Identification	48.52%	48.10%	-0.42
Disease Diagnosis	63.74%	60.94%	-2.80
Lesion Grading	57.61%	55.43%	-2.18
Modality Recognition	97.57%	96.18%	-1.39
Other Biological Attributes	71.72%	66.21%	-5.51

Appendix

Table 7 : Performance comparison on OmniMedVQA-Mini by overall accuracy and question type.

Configuration	Latency (s)	Lat. $\Delta$ (%)	Peak Mem. (MB)	Mem. $\Delta$ (%)
Base (Qwen2.5-VL)	0.382	–	16,857	–
+ SNS + NVF	0.456	+19.4	23,933	+42.0
+ SNS + NVF + PCS	0.457	+19.7	23,933	+42.0

Appendix

Table 8: Inference overhead relative to the base Qwen2.5-VL model.

Module	Component	Params	% of Base	Stage 1	Stage 2
SNS	SAE	822.08M	10.79%	Train	Frozen
SNS	Neuron Clusters	–	–	Built	Frozen
NVF	Router	6.54M	0.09%	Train	Train
	Vision-Side Scorer	6.54M	0.09%	Train	Train
	Projection Layers	12.85M	0.17%	–	Train
	Localized Cross-Attention	51.39M	0.67%	–	Train

Appendix

Table 9 : Parameter breakdown of the added components. Percentages are relative to the 7.6B-parameter base model.

Explore similar work

Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders
Jun 19, 2026Sergio Lanza, Jae Hee Lee, Stefan WermterExtraction
Knowledge Technology, Department of Informatics, University of Hamburg, Hamburg, Germany
Mixture of Cognitive Experts in Large Vision-Language Models
Jul 12, 2026Robert Wijaya, Ngai-Man CheungRecent Vision-Language Models Multimodal Reasoning
Singapore University of Technology and Design (SUTD), Singapore
Interpretability Transfer from Language to Vision via Sparse Autoencoders
May 24, 2026Alexey Kravets, Da Li, Chuan Li +2Sparse Autoencoder Features Interpretability

Jun 19, 2026, cs.CVJ/K move · Enter open · S save

Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders

Sergio Lanza, Jae Hee Lee, Stefan Wermter

Knowledge Technology, Department of Informatics, University of Hamburg, Hamburg, Germany

Extraction

Open paper AskPDF arXiv

Vision Language Models (VLMs) have demonstrated impressive performance in tasks requiring joint understanding of images and text, such as image captioning and Visual Question Answering (VQA), but our understanding of their internal processes remains limited. Recently, Sparse Autoencoders (SAEs) have emerged as a promising tool to support the interpretation of concepts encoded in VLMs. However, most SAE-based approaches focus only on textual or visual concepts separately, ignoring multimodal concepts. This limitation hinders a comprehensive understanding of VLMs, since concepts that integrate both modalities can be misclassified. Moreover, previous visual approaches often produce low-quality visual concept descriptions that are vague or incomplete, limiting their usefulness for understanding model reasoning. We propose a framework based on SAEs to extract and analyze visual, textual, and multimodal concepts from VLMs. For each neuron, we propose a candidate human-interpretable concept and compute the alignment between the concept and the dataset samples using cosine similarity scores. Experiments on a VQA dataset (LLaVA-NeXT) demonstrate that our framework improves visual concept quality by up to 45% compared to existing SAE-based methods, while maintaining high textual concept quality and enabling systematic identification of multimodal concepts. This work contributes new insights into the conceptual space of VLMs, providing a structured approach to distinguish between visual, textual, and multimodal concepts. The code is available at https://github.com/PHDLanza/Multidata_SAE

Abstract

Explore similar work

Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders

Mixture of Cognitive Experts in Large Vision-Language Models

Interpretability Transfer from Language to Vision via Sparse Autoencoders