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

• The framework draws on sparse concept-level neural coding and top-down modulation of task-relevant responses in biological vision.Source: §1 · p. 2
• NeuronEye decomposes vision-token states into sparse concept clusters, activates query-relevant concepts in localized patches, and suppresses dominant directions in one pass over a frozen backbone.Source: §1 · p. 2
• On Qwen2.5-VL-7B, CV-Bench overall accuracy rises from 78.5 to 81.6 (+3.1), with +9.5 on Distance and +8.3 on BLINK Multi-view.Source: §1 · p. 2
• NeuronEye uses sparse autoencoder features as an active vocabulary for concept-level visual reasoning rather than only as post-hoc interpretability tools.Source: §1 · p. 2
• Evaluation covers four vision-centric benchmarks and two VLM backbones, with improvements on spatial and localization-sensitive tasks.Source: §1 · p. 2

2.1 Visual Representation Modulation

• Token-selection methods retain dense visual content within each kept patch, while representation-level steering adjusts dense states uniformly rather than selecting query-specific concepts.Source: §2.1 · p. 3
• Unlike token selection or uniform dense-state steering, NeuronEye selects sparse concept subsets within localized patches for query-specific modulation.Source: §2.1 · p. 3, §2.2 · p. 3

2.2 Sparse Autoencoders in Multimodal Models

• Prior multimodal SAE interventions suppress fixed hallucination-related features without query conditioning or structured routing; NeuronEye organizes features into query-selected concept clusters localized to image patches.Source: §2.2 · p. 3

3.1 Sparse Neuron Space (SNS)

• SNS encodes patch tokens in an overcomplete sparse basis, retains k active neurons per patch, and groups reliable neurons into concept-level clusters.The clusters form the sparse neuron vocabulary used by NVF.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 Neuron-guided Visual Focus

• NVF scores query-relevant clusters, localizes patches expressing selected concepts, refines their sparse codes, and injects the decoded evidence into selected vision tokens.Only selected tokens are updated; the remaining vision tokens stay unchanged.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.3 Perceptual Concept Suppression

• PCS estimates dominant perceptual directions from vision-token geometry at a later layer and attenuates them to preserve representational diversity and weaker visual cues.Its learnable suppression strength is non-negative and initialized near zero.Source: §3.3 · pp. 5–6

4.1 Dataset; 4.2 Implementation Details

• Training uses VQAv2 data disjoint from evaluation benchmarks, and evaluation spans spatial understanding, counting, relations, localization, and general visual question answering.The evaluation uses Qwen2.5-VL-7B and LLaVA-1.6-7B as frozen backbones; SNS is constructed at layer 8, NVF applied there, and PCS at layer 20.Source: §4.1 · p. 6, §4.2 · p. 6

4.3 Benchmark Results across VLM Backbones

• CV-Bench overall rises +3.1 and BLINK overall +0.7 on Qwen2.5-VL-7B; gains peak on Distance (+9.5), Multi-view (+8.3), and Localization (+3.3).On LLaVA-1.6-7B, gains include Distance (+6.9), Relation (+12.1), Multi-view (+7.7), and Localization (+8.1), while Count (-8.3) and Depth (-3.3) decrease.Source: §4.3 · p. 6, Table 1 · p. 7

4.4 Comparison with Visual Representation Modulation Methods

• NeuronEye achieves the best overall performance on CV-Bench and BLINK against visual-modulation methods, with especially large gains on Distance (+9.5) and Multi-view (+8.3).Token reduction methods degrade on structure-sensitive subtasks, while representation-level methods offer limited improvement.Source: §4.4 · pp. 6–7, Table 2 · p. 7

4.5 Visualization of Sparse Neurons and Visual Focus

• Fig. 4 shows individual SNS neurons activating on semantically coherent regions such as faces, towers, and flags.Each subfigure presents top-activating patches for one SAE neuron with non-activating regions masked.Source: §4.5 · p. 7, Figure 4 · p. 7
• Fig. 5 shows NVF selecting query-relevant clusters whose patch activations focus on localized regions rather than globally reweighting the image.Examples focus on shelves and packages, clothing and people, or two elephants and their contact region, depending on the question.Source: Figure 5 · p. 8, §4.5 · p. 8

4.6 Ablation Studies

• Fine-tuning lightweight modules without sparse intervention does not improve performance, random cluster assignments yield near-baseline results, and PCS alone changes performance only marginally.SNS + NVF improves CV-Bench Overall and Distance; the full configuration further improves Distance and BLINK Multi-view.Source: §4.6 · p. 8, Table 3 · p. 8
• Layer 8 gives the best NVF insertion results, especially on CV-Bench Distance and BLINK Multi-view; earlier layer 4 is slightly weaker, and layers 12–20 degrade toward baseline.The table caption describes layer 8 as balancing low-level perceptual features and high-level semantic abstraction.Source: §4.6 · p. 8, Table 4 · p. 9
• K = 64 gives the best overall trade-off, with the strongest results on CV-Bench Distance and BLINK Multi-view.Fewer clusters can conflate distinct concepts, while more clusters can split related neurons and destabilize routing.Source: Table 5 · p. 9, §4.6 · p. 9

Limitations

• The frozen vocabulary is trained on VQAv2 and may lack domain-specific concepts; on OmniMedVQA-Mini, overall accuracy falls from 65.30% to 63.10%.Drops are larger for disease diagnosis (-2.80), lesion grading (-2.18), and other biological attributes (-5.51).Source: §5 · p. 9
• Only approximately 20% of LLM-generated cluster-index labels were manually verified, so query-to-cluster misalignment may persist in the remainder and potentially degrade performance.Source: §5 · p. 9
• Inference adds approximately 20% latency and 42% peak memory relative to the base model, mainly from frozen SAE encoding and localized cross-attention.Source: §5 · p. 9

Future directions

• Future work could expand training corpora, build domain-specific vocabularies, extend sparse routing across layers, or learn clusters directly from data to reduce reliance on offline labels.Source: §5 · p. 9

6 Conclusion

• NeuronEye decomposes dense VLM representations into a sparse, concept-level vocabulary and selectively activates query-relevant visual concepts without modifying or retraining the backbone.The conclusion presents sparse autoencoder features as active interfaces for modulating concept-level visual reasoning.Source: §6 · p. 9
• Experiments across four benchmarks and two backbones demonstrate consistent improvements on spatial, relational, and multi-view reasoning.Source: §6 · p. 9

A Offline Interpretation Prompts

• Offline prompts assign concept labels to sparse neurons and map training questions to cluster indices; these annotations are not used by inference-time components.Figure 6 shows the prompts used for sparse-neuron interpretation and cluster-index supervision.Source: A Offline Interpretation Prompts · p. 13, Figure 6 · p. 13

B Training, Inference, and Hyperparameter Details

• Training first stabilizes query-to-cluster mapping with frozen hidden states, then enables the full NVF pipeline and PCS suppression while optimizing lightweight components.Stage 1 uses cluster-index prediction and query–vision alignment losses before feature injection; Stage 2 adds sparse feature localization, refinement, and localized injection.Source: B Training, Inference, and… · p. 14
• The VLM backbone, pretrained SAE, and neuron-cluster assignments remain fixed throughout training; only lightweight NVF modules and the PCS scalar are updated.Source: B Training, Inference, and… · p. 14
• At inference, NVF injects localized updates at layer ℓ and PCS suppresses dominant directions at layer ℓ′ in a single modified forward pass.Inference uses no external annotator, retrieval module, parameter update, or multi-pass decoding; Table 6 lists hyperparameter settings for all experiments.Source: B Training, Inference, and… · p. 14, Table 6 · p. 14

C Additional Sparse Neuron Visualizations

• Many retained neurons correspond to localized, semantically coherent visual patterns, supporting their organization into concept-level clusters.Source: C Additional Sparse Neuron… · p. 15
• Each visualization shows the top-activating patches for one SNS neuron across multiple images, with non-activating regions masked.Source: Figure 7 · p. 15

D Additional Question-Guided Focus Examples

• NVF selects a small set of query-relevant neuron clusters and increases their neuron activations, yielding cluster-specific maps of localized image regions relevant to answering the question.Source: D Additional Question-Guided Focus… · p. 16

E Additional Domain-Shift Evaluation

• Overall accuracy on OmniMedVQA-Mini falls from 65.30% to 63.10% with NeuronEye.The result suggests that a sparse neuron vocabulary built from general-domain VQAv2 activations may not fully cover specialized medical visual concepts.Source: Table 7 · p. 17, E Additional Domain-Shift Evaluation · p. 17

F Efficiency and Overhead

• ≈20% latency increase over the base model comes mainly from frozen SAE encoding and localized cross-attention; PCS adds negligible latency and no peak memory in this setting.Table 8 reports inference overhead relative to the base Qwen2.5-VL model.Source: Table 8 · p. 17, F Efficiency and Overhead · p. 17
• Memory overhead mainly comes from intermediate activations, routing buffers, and the frozen SAE, rather than additional trainable components.The passage does not quantify the memory increase.Source: F Efficiency and Overhead · p. 17
• 1.41% parameter overhead refers only to trainable modules; the SAE is trained once to construct SNS and then kept frozen during lightweight module training and inference.Trainable components include the router, vision-side scorer, projection layers, localized cross-attention, refinement module, and PCS scalar; the frozen SAE contributes non-trainable parameters and memory overhead.Source: Table 9 · p. 17, F Efficiency and Overhead · p. 17