# Steering CLIP's vision transformer with sparse autoencoders

arXiv: https://arxiv.org/abs/2504.08729 · HTML: https://mlsanity.com/paper/2504.08729/

## 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.

## Placed in

- Mechanistic interpretability / [Interpreting vision and multimodal models](https://mlsanity.com/mechinterp/interpreting-vision-and-multimodal-models/)

## Builds on (most cited first, 30 of 31)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [Emerging Properties in Self-Supervised Vision Transformers](https://mlsanity.com/paper/2104.14294/) (2021)
- [Improved Baselines with Visual Instruction Tuning](https://mlsanity.com/paper/2310.03744/) (2023)
- [DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation](https://mlsanity.com/paper/2208.12242/) (2022)
- [Sparse Autoencoders Find Highly Interpretable Features in Language Models](https://mlsanity.com/paper/2309.08600/) (2023)
- [Magic3D: High-Resolution Text-to-3D Content Creation](https://mlsanity.com/paper/2211.10440/) (2022)
- [Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization](https://mlsanity.com/paper/1911.08731/) (2019)
- [Multi-Concept Customization of Text-to-Image Diffusion](https://mlsanity.com/paper/2212.04488/) (2022)
- [Concept Bottleneck Models](https://mlsanity.com/paper/2007.04612/) (2020)
- [Large Language Models](https://mlsanity.com/paper/2307.05782/) (2023)
- [Label-Free Concept Bottleneck Models](https://mlsanity.com/paper/2304.06129/) (2023)
- [Interpreting CLIP's Image Representation via Text-Based Decomposition](https://mlsanity.com/paper/2310.05916/) (2023)
- [Patching open-vocabulary models by interpolating weights](https://mlsanity.com/paper/2208.05592/) (2022)
- [Simple data balancing achieves competitive worst-group-accuracy](https://mlsanity.com/paper/2110.14503/) (2021)
- [AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers](https://mlsanity.com/paper/2402.05602/) (2024)
- [Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)](https://mlsanity.com/paper/2402.10376/) (2024)
- [Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discovery](https://mlsanity.com/paper/2407.14499/) (2024)
- [Spread Spurious Attribute: Improving Worst-group Accuracy with Spurious Attribute Estimation](https://mlsanity.com/paper/2204.02070/) (2022)
- [Sparse autoencoders reveal selective remapping of visual concepts during adaptation](https://mlsanity.com/paper/2412.05276/) (2024)
- [Disentangling visual and written concepts in CLIP](https://mlsanity.com/paper/2206.07835/) (2022)
- [The Missing Curve Detectors of InceptionV1: Applying Sparse Autoencoders to InceptionV1 Early Vision](https://mlsanity.com/paper/2406.03662/) (2024)
- [Interpreting the Second-Order Effects of Neurons in CLIP](https://mlsanity.com/paper/2406.04341/) (2024)
- [Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence](https://mlsanity.com/paper/2202.03482/) (2022)
- [An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generation](https://mlsanity.com/paper/2410.03334/) (2024)
- [Defense-Prefix for Preventing Typographic Attacks on CLIP](https://mlsanity.com/paper/2304.04512/) (2023)
- [Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers](https://mlsanity.com/paper/2412.06639/) (2024)
- [Discriminative Class Tokens for Text-to-Image Diffusion Models](https://mlsanity.com/paper/2303.17155/) (2023)

## Built on by (most cited first, 30 of 38)

- [Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video](https://mlsanity.com/paper/2504.19475/) (2025)
- [Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models](https://mlsanity.com/paper/2603.19233/) (2026)
- [Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrieval](https://mlsanity.com/paper/2512.03276/) (2025)
- [From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliance](https://mlsanity.com/paper/2505.20229/) (2025)
- [From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers](https://mlsanity.com/paper/2509.06938/) (2025)
- [Suppression or Deletion: A Restoration-Based Representation-Level Analysis of Machine Unlearning](https://mlsanity.com/paper/2602.18505/) (2026)
- [Symbolic Rule Extraction from Attention-Guided Sparse Representations in Vision Transformers](https://mlsanity.com/paper/2505.06745/) (2025)
- [A Geometric Unification of Concept Learning with Concept Cones](https://mlsanity.com/paper/2512.07355/) (2025)
- [Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speech](https://mlsanity.com/paper/2606.01479/) (2026)
- [Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers](https://mlsanity.com/paper/2606.06664/) (2026)
- [Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering](https://mlsanity.com/paper/2506.01247/) (2025)
- [Sparse CLIP: Co-Optimizing Interpretability and Performance in Contrastive Learning](https://mlsanity.com/paper/2601.20075/) (2026)
- [Where Vision Becomes Text: Locating the OCR Routing Bottleneck in Vision-Language Models](https://mlsanity.com/paper/2602.22918/) (2026)
- [Language Models Can Explain Visual Features via Steering](https://mlsanity.com/paper/2603.22593/) (2026)
- [Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models](https://mlsanity.com/paper/2605.15961/) (2026)
- [On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders](https://mlsanity.com/paper/2605.31518/) (2026)
- [Evaluating the Interpretability of Sparse Autoencoders with Concept Annotations](https://mlsanity.com/paper/2606.24716/) (2026)
- [Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoders](https://mlsanity.com/paper/2511.17735/) (2025)
- [SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networks](https://mlsanity.com/paper/2512.15938/) (2025)
- [LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer](https://mlsanity.com/paper/2512.18930/) (2025)
- [SCoCCA: Multi-modal Sparse Concept Decomposition via Canonical Correlation Analysis](https://mlsanity.com/paper/2603.13884/) (2026)
- [Interpretability Transfer from Language to Vision via Sparse Autoencoders](https://mlsanity.com/paper/2605.24946/) (2026)
- [Representation Alignment Rests on Linear Structure](https://mlsanity.com/paper/2605.28870/) (2026)
- [MEIcoder: Decoding Visual Stimuli from Neural Activity by Leveraging Most Exciting Inputs](https://mlsanity.com/paper/2510.20762/) (2025)
- [Interpreting Video Representations with Spatio-Temporal Sparse Autoencoders](https://mlsanity.com/paper/2604.03919/) (2026)
- [Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs](https://mlsanity.com/paper/2604.08846/) (2026)
- [From Attribution to Action: A Human-Centered Application of Activation Steering](https://mlsanity.com/paper/2604.11467/) (2026)
- [Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision](https://mlsanity.com/paper/2604.13304/) (2026)
- [LatentDiff: Scaling Semantic Dataset Comparison to Millions of Images](https://mlsanity.com/paper/2605.00899/) (2026)
- [Vision-Language Asymmetry in Bistable Image Captioning](https://mlsanity.com/paper/2606.08031/) (2026)

## 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.