mlsanityData to 14 Oct 2026

Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers

arXiv 2606.06664 ↗First posted 4 Jun 2026cs.CV cs.AI cs.LGcited by 3

On the map: Mechanistic interpretability → Interpreting vision and multimodal models

Abstract (arXiv)

Despite high accuracy, Vision Transformer (ViT) predictions can be driven by spurious cues, raising the need to understand their inner workings before safe deployment. Sparse autoencoders (SAEs) provide a promising lens for decomposing model representations into human-interpretable concepts, yet adapting SAE-based interpretation to ViTs remains challenging due to limited control over concept coverage and subjective, non-scalable feature interpretation. To fill the gaps, motivated by neuroscience-inspired principles, we propose ViSAE, a mechanistic interpretability toolbox for understanding ViT inner workings through concept circuits. ViSAE consists of three components: (1) A probing suite with 64K images and a 16K visually grounded concept vocabulary, improving concept coverage efficiency by 20x over ImageNet and interpretation accuracy by 28.7% over existing concept sets. (2) Top-down concept reading and Bottom-up circuit tracing algorithms that automatically recover ViT inner workings via concept circuits. (3) Applications for auditing and steering ViT behavior. Through concept editing, ViSAE improves the worst-group accuracy on WaterBirds by 48.2%, outperforming existing methods by 23.8%. Our data and code: https://github.com/deep-real/ViSAE.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Interpreting vision and multimodal models, one of 13 directions in Mechanistic interpretability. Builds on 54 of its 71 references that are in the map. First posted on arXiv 4 Jun 2026.

Builds on

54 of 71 references.

Closest ancestors · 6

Same direction, newest first

  1. 2025Steering CLIP's vision transformer with sparse autoencoders11 Apr
  2. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models3 Apr
  3. 2025Interpreting CLIP with Hierarchical Sparse Autoencoders27 Feb
  4. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment6 Feb
  5. 2024Sparse autoencoders reveal selective remapping of visual concepts during adaptation6 Dec
  6. 2024Large Multi-modal Models Can Interpret Features in Large Multi-modal Models22 Nov

Other references · 48

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  3. 2023GPT-4 Technical Reportcited by 18,396
  4. 2014Microsoft COCO: Common Objects in Contextcited by 17,608
  5. 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
  6. 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
  7. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,070
  8. 2023Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detectioncited by 3,332
  9. 2015ShapeNet: An Information-Rich 3D Model Repositorycited by 2,897
  10. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  11. 2016Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotationscited by 2,656
  12. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
36 more
  1. 2013Describing Textures in the Wildcited by 2,125
  2. 2019Invariant Risk Minimizationcited by 1,854
  3. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  4. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  5. 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,319
  6. 2018Sanity Checks for Saliency Mapscited by 1,156
  7. 2020Concept Bottleneck Modelscited by 1,032
  8. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  9. 2018This Looks Like That: Deep Learning for Interpretable Image Recognitioncited by 737
  10. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  11. 2017Network Dissection: Quantifying Interpretability of Deep Visual Representationscited by 708
  12. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  13. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  14. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  15. 2017The (Un)reliability of saliency methodscited by 333
  16. 2021Generic Attention-model Explainability for Interpreting Bi-Modal and Encoder-Decoder Transformerscited by 320
  17. 2022Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classificationcited by 312
  18. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 300
  19. 2017Right for the Right Reasons: Training Differentiable Models by Constraining their Explanationscited by 294
  20. 2022Post-hoc Concept Bottleneck Modelscited by 293
  21. 2013Direct and Indirect Effectscited by 280
  22. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 179
  23. 2022Quantus: An Explainable AI Toolkit for Responsible Evaluation of Neural Network Explanations and Beyondcited by 171
  24. 2022CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networkscited by 147
  25. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
  26. 2024BatchTopK Sparse Autoencoderscited by 130
  27. 2024Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)cited by 119
  28. 2021Do Concept Bottleneck Models Learn as Intended?cited by 116
  29. 2020Neuron Shapley: Discovering the Responsible Neuronscited by 98
  30. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
  31. 2024Decomposing and Editing Predictions by Modeling Model Computationcited by 19
  32. 2023Are Data-driven Explanations Robust against Out-of-distribution Data?cited by 14
  33. 2024DEAL: Disentangle and Localize Concept-level Explanations for VLMscited by 14
  34. 2025Interpretable Failure Detection with Human-Level Conceptscited by 7
  35. 2024Beyond Accuracy: Ensuring Correct Predictions With Correct Rationalescited by 6
  36. 2024Beyond Accuracy: On the Effects of Fine-tuning Towards Vision-Language Model's Prediction Rationalitycited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Medical AI Encodes a "Feeling of Error": Verifying Cancer Segmentation via Internal Conceptscited by 0
  2. 2026"World Knowledge" in the Weights: Reading Concept Circuits of Vision Transformerscited by 0
  3. 2026The Platonic brain bridge hypothesis: human brain networks as an architectural prior for multimodal large language modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Interpretable and Testable Vision Features via Sparse Autoencoders
  2. 2026ViT-Explainer: An Interactive Walkthrough of the Vision Transformer Pipeline
  3. 2026SpIn-ViT: Designing a Sparsity-Induced Vision Transformer That Is Mechanistically Interpretable
  4. 2025Interpreting vision transformers via residual replacement model
  5. 2026Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision
  6. 2023Interpretability-Aware Vision Transformer
  7. 2025Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video
  8. 2026ViT-5: Vision Transformers for The Mid-2020s