mlsanityData to 14 Oct 2026

Steering CLIP's vision transformer with sparse autoencoders

arXiv 2504.08729 ↗First posted 11 Apr 2025cs.CV cs.AI cs.LGcited by 38

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

Abstract (arXiv)

While vision models are highly capable, their internal mechanisms remain poorly understood -- a challenge which sparse autoencoders (SAEs) have helped address in language, but which remains underexplored in vision. We address this gap by training SAEs on CLIP's vision transformer and uncover key differences between vision and language processing, including distinct sparsity patterns for SAEs trained across layers and token types. We then provide the first systematic analysis on the steerability of CLIP's vision transformer by introducing metrics to quantify how precisely SAE features can be steered to affect the model's output. We find that 10-15% of neurons and features are steerable, with SAEs providing thousands more steerable features than the base model. Through targeted suppression of SAE features, we then demonstrate improved performance on three vision disentanglement tasks (CelebA, Waterbirds, and typographic attacks), finding optimal disentanglement in middle model layers, and achieving state-of-the-art performance on defense against typographic attacks.

Key facts. Cited by 38 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 31 of its 47 references that are in the map. First posted on arXiv 11 Apr 2025.

Builds on

31 of 47 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024Sparse autoencoders reveal selective remapping of visual concepts during adaptation6 Dec

Other references · 30

  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. 2024The Llama 3 Herd of Modelscited by 14,715
  4. 2023Visual Instruction Tuningcited by 7,869
  5. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,968
  6. 2023Improved Baselines with Visual Instruction Tuningcited by 4,418
  7. 2022DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generationcited by 3,300
  8. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  9. 2022Magic3D: High-Resolution Text-to-3D Content Creationcited by 1,290
  10. 2019Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalizationcited by 1,247
  11. 2022Multi-Concept Customization of Text-to-Image Diffusioncited by 1,106
  12. 2020Concept Bottleneck Modelscited by 1,032
18 more
  1. 2023Large Language Modelscited by 780
  2. 2023Label-Free Concept Bottleneck Modelscited by 285
  3. 2023Interpreting CLIP's Image Representation via Text-Based Decompositioncited by 196
  4. 2022Patching open-vocabulary models by interpolating weightscited by 190
  5. 2021Simple data balancing achieves competitive worst-group-accuracycited by 178
  6. 2024AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformerscited by 120
  7. 2024Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)cited by 119
  8. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
  9. 2022Spread Spurious Attribute: Improving Worst-group Accuracy with Spurious Attribute Estimationcited by 88
  10. 2022Disentangling visual and written concepts in CLIPcited by 58
  11. 2024The Missing Curve Detectors of InceptionV1: Applying Sparse Autoencoders to InceptionV1 Early Visioncited by 37
  12. 2024Interpreting the Second-Order Effects of Neurons in CLIPcited by 36
  13. 2022Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergencecited by 25
  14. 2024An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generationcited by 25
  15. 2023Defense-Prefix for Preventing Typographic Attacks on CLIPcited by 24
  16. 2024Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformerscited by 11
  17. 2023Discriminative Class Tokens for Text-to-Image Diffusion Modelscited by 9
  18. 2021The Effect of Model Size on Worst-Group Generalizationcited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 11; 2026: 27.

38 papers.

Sort:

By area:

  1. 2025Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Videocited by 21
  2. 2026Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Modelscited by 15
  3. 2025Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrievalcited by 7
  4. 2025From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliancecited by 6
  5. 2025From Noise to Narrative: Tracing the Origins of Hallucinations in Transformerscited by 4
  6. 2026Suppression or Deletion: A Restoration-Based Representation-Level Analysis of Machine Unlearningcited by 4
  7. 2025Symbolic Rule Extraction from Attention-Guided Sparse Representations in Vision Transformerscited by 3
  8. 2025A Geometric Unification of Concept Learning with Concept Conescited by 3
  9. 2026Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speechcited by 3
  10. 2026Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformerscited by 3
  11. 2025Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steeringcited by 2
  12. 2026Sparse CLIP: Co-Optimizing Interpretability and Performance in Contrastive Learningcited by 2
26 more
  1. 2026Where Vision Becomes Text: Locating the OCR Routing Bottleneck in Vision-Language Modelscited by 2
  2. 2026Language Models Can Explain Visual Features via Steeringcited by 2
  3. 2026Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP modelscited by 2
  4. 2026On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoderscited by 2
  5. 2026Evaluating the Interpretability of Sparse Autoencoders with Concept Annotationscited by 2
  6. 2025Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoderscited by 1
  7. 2025SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networkscited by 1
  8. 2025LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfercited by 1
  9. 2026SCoCCA: Multi-modal Sparse Concept Decomposition via Canonical Correlation Analysiscited by 1
  10. 2026Interpretability Transfer from Language to Vision via Sparse Autoencoderscited by 1
  11. 2026Representation Alignment Rests on Linear Structurecited by 1
  12. 2025MEIcoder: Decoding Visual Stimuli from Neural Activity by Leveraging Most Exciting Inputscited by 0
  13. 2026Interpreting Video Representations with Spatio-Temporal Sparse Autoencoderscited by 0
  14. 2026Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMscited by 0
  15. 2026From Attribution to Action: A Human-Centered Application of Activation Steeringcited by 0
  16. 2026Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Visioncited by 0
  17. 2026LatentDiff: Scaling Semantic Dataset Comparison to Millions of Imagescited by 0
  18. 2026Vision-Language Asymmetry in Bistable Image Captioningcited by 0
  19. 2026The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensionscited by 0
  20. 2026At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalizationcited by 0
  21. 2026IRIS: A Visual Cortex-Inspired Framework for Analyzing Orientation Selectivity in Vision Transformerscited by 0
  22. 2026Multimodal Model Diffing for Feature Discovery and Controlcited by 0
  23. 2026EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoderscited by 0
  24. 2026Detect and Suppress: A Mechanistic Defense against Adversarial Patches in VLA Modelscited by 0
  25. 2026Metonymic Circuits for Abstract Concept Grounding in Vision Transformerscited by 0
  26. 2026Forensic Reserve: Eliciting Latent Knowledge for Image Forgery Detectioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models
  2. 2025Probing the Representational Power of Sparse Autoencoders in Vision Models
  3. 2025Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alone
  4. 2024Analyzing (In)Abilities of SAEs via Formal Languages
  5. 2025Interpretable and Testable Vision Features via Sparse Autoencoders
  6. 2025Interpreting CLIP with Hierarchical Sparse Autoencoders
  7. 2026Steering Vision-Language Models with Joint Sparse Autoencoders
  8. 2025Interpreting the linear structure of vision-language model embedding spaces