mlsanityData to 14 Oct 2026

Interpreting CLIP with Hierarchical Sparse Autoencoders

arXiv 2502.20578 ↗First posted 27 Feb 2025cs.CV cs.AI cs.LGcited by 60

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

Abstract (arXiv)

Sparse autoencoders (SAEs) are useful for detecting and steering interpretable features in neural networks, with particular potential for understanding complex multimodal representations. Given their ability to uncover interpretable features, SAEs are particularly valuable for analyzing large-scale vision-language models (e.g., CLIP and SigLIP), which are fundamental building blocks in modern systems yet remain challenging to interpret and control. However, current SAE methods are limited by optimizing both reconstruction quality and sparsity simultaneously, as they rely on either activation suppression or rigid sparsity constraints. To this end, we introduce Matryoshka SAE (MSAE), a new architecture that learns hierarchical representations at multiple granularities simultaneously, enabling a direct optimization of both metrics without compromise. MSAE establishes a new state-of-the-art Pareto frontier between reconstruction quality and sparsity for CLIP, achieving 0.99 cosine similarity and less than 0.1 fraction of variance unexplained while maintaining ~80% sparsity. Finally, we demonstrate the utility of MSAE as a tool for interpreting and controlling CLIP by extracting over 120 semantic concepts from its representation to perform concept-based similarity search and bias analysis in downstream tasks like CelebA. We make the codebase available at https://github.com/WolodjaZ/MSAE.

Key facts. Cited by 60 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 57 of its 74 references that are in the map. First posted on arXiv 27 Feb 2025.

Builds on

57 of 74 references.

References · 57

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2014ImageNet Large Scale Visual Recognition Challengecited by 8,498
  3. 2023Visual Instruction Tuningcited by 7,869
  4. 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
  5. 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
  6. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,070
  7. 2023SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesiscited by 4,356
  8. 2014Deep Learning Face Attributes in the Wildcited by 3,693
  9. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  10. 2023Sigmoid Loss for Language Image Pre-Trainingcited by 3,109
  11. 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,371
  12. 2013Visualizing and Understanding Convolutional Networkscited by 2,290
38 more
  1. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  2. 2019Similarity of Neural Network Representations Revisitedcited by 1,755
  3. 2021LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairscited by 1,555
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  5. 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,319
  6. 2022Reproducible scaling laws for contrastive language-image learningcited by 1,173
  7. 2018Sanity Checks for Saliency Mapscited by 1,156
  8. 2020Concept Bottleneck Modelscited by 1,032
  9. 2020Quantifying Attention Flow in Transformerscited by 775
  10. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  11. 2023CogVLM: Visual Expert for Pretrained Language Modelscited by 708
  12. 2024Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Thinkcited by 703
  13. 2024Scaling and evaluating sparse autoencoderscited by 553
  14. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  15. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  16. 2024The Platonic Representation Hypothesiscited by 382
  17. 2021How Much Can CLIP Benefit Vision-and-Language Tasks?cited by 321
  18. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 300
  19. 2016Not Just a Black Box: Learning Important Features Through Propagating Activation Differencescited by 297
  20. 2023Label-Free Concept Bottleneck Modelscited by 285
  21. 2022Matryoshka Representation Learningcited by 282
  22. 2023Interpreting CLIP's Image Representation via Text-Based Decompositioncited by 196
  23. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 179
  24. 2020Part-aware Prototype Network for Few-shot Semantic Segmentationcited by 162
  25. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 160
  26. 2022CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networkscited by 147
  27. 2022Natural Language Descriptions of Deep Visual Featurescited by 137
  28. 2024BatchTopK Sparse Autoencoderscited by 130
  29. 2024Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)cited by 119
  30. 2025Sparse Autoencoders Trained on the Same Data Learn Different Featurescited by 93
  31. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
  32. 2022Distilling Model Failures as Directions in Latent Spacecited by 82
  33. 2022Overlooked factors in concept-based explanations: Dataset choice, concept learnability, and human capabilitycited by 46
  34. 2020On Interpretability of Deep Learning based Skin Lesion Classifiers using Concept Activation Vectorscited by 39
  35. 2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Modelscited by 33
  36. 2023Labeling Neural Representations with Inverse Recognitioncited by 32
  37. 2022VAEL: Bridging Variational Autoencoders and Probabilistic Logic Programmingcited by 31
  38. 2022Exploring Visual Interpretability for Contrastive Language-Image Pre-trainingcited by 30

50 most cited of 57 All 57 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 13; 2026: 47.

60 papers.

Sort:

By area:

  1. 2025From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuitcited by 49
  2. 2026Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Modelscited by 15
  3. 2026DNA: Uncovering Universal Latent Forgery Knowledgecited by 14
  4. 2025Measuring and Guiding Monosemanticitycited by 10
  5. 2025Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactionscited by 6
  6. 2025Emergent World Representations in OpenVLAcited by 6
  7. 2025Interpretable and Steerable Concept Bottleneck Sparse Autoencoderscited by 6
  8. 2025Causal Interpretation of Sparse Autoencoder Features in Visioncited by 5
  9. 2025MoME: Mixture of Matryoshka Experts for Audio-Visual Speech Recognitioncited by 4
  10. 2025Model Science: getting serious about verification, explanation and control of AI systemscited by 3
  11. 2026Representations of Text and Images Align From Layer Onecited by 3
  12. 2026Latent Anomaly Knowledge Excavation: Unveiling Sparse Sensitive Neurons in Vision-Language Modelscited by 3
38 more
  1. 2026Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformerscited by 3
  2. 2025Semantic Regexes: Auto-Interpreting LLM Features with a Structured Languagecited by 2
  3. 2025MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoderscited by 2
  4. 2026Decomposing multimodal embedding spaces with group-sparse autoencoderscited by 2
  5. 2026Sparse CLIP: Co-Optimizing Interpretability and Performance in Contrastive Learningcited by 2
  6. 2026Evaluating the Interpretability of Sparse Autoencoders with Concept Annotationscited by 2
  7. 2025FaCT: Faithful Concept Traces for Explaining Neural Network Decisionscited by 1
  8. 2025Extracting Interaction-Aware Monosemantic Concepts in Recommender Systemscited by 1
  9. 2026Inspiration Seeds: Learning Non-Literal Visual Combinations for Generative Explorationcited by 1
  10. 2026Hierarchical Concept Embedding & Pursuit for Interpretable Image Classificationcited by 1
  11. 2026SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Modelscited by 1
  12. 2026Beyond Semantics: Disentangling Information Scope in Sparse Autoencoders for CLIPcited by 1
  13. 2026From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Modelscited by 1
  14. 2026SoftSAE: Dynamic Top-K Selection for Adaptive Sparse Autoencoderscited by 1
  15. 2026TEVI: Text-Conditioned Editing of Visual Representations via Sparse Autoencoders for Improved Vision-Language Alignmentcited by 1
  16. 2026Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMscited by 1
  17. 2026Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?cited by 1
  18. 2026When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalitiescited by 1
  19. 2026Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherencecited by 1
  20. 2025The Deleuzian Representation Hypothesiscited by 0
  21. 2026RAIGen: Rare Attribute Identification in Text-to-Image Generative Modelscited by 0
  22. 2026Med-SegLens: Latent-Level Model Diffing for Interpretable Medical Image Segmentationcited by 0
  23. 2026Sparse Visual Thought Circuits in Vision-Language Modelscited by 0
  24. 2026LatentDiff: Scaling Semantic Dataset Comparison to Millions of Imagescited by 0
  25. 2026Birds of a Feather Flock Together: Background-Invariant Representations via Linear Structure in VLMscited by 0
  26. 2026Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Modelscited by 0
  27. 2026Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Modelscited by 0
  28. 2026Interpretability Without Tradeoffs: Disentangling Polysemanticity At Equal Predictive Performancecited by 0
  29. 2026Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoderscited by 0
  30. 2026Aligning Sentence Embeddings to Human Concepts via Sparse Autoencoderscited by 0
  31. 2026Towards Robustness against Typographic Attack with Training-free Concept Localizationcited by 0
  32. 2026The Hyperspherical Geometry of CLIP Latent Space: A Semantic Mixture Modelcited by 0
  33. 2026SADe: Sparse-Atom Support Decontamination for Few-Shot Segmentation with Weak Support Annotationscited by 0
  34. 2026V-FIND: Revealing the Intrinsic Forgery Knowledge Encoded in Video Forgery Detectorscited by 0
  35. 2026IRIS: A Visual Cortex-Inspired Framework for Analyzing Orientation Selectivity in Vision Transformerscited by 0
  36. 2026SCOUT: Semantic Concept Discovery for Open-Vocabulary Editing of face Recognition Templatescited by 0
  37. 2026A survey detection channel overrides the pixels in an astronomical foundation model, and biases tomographic mean redshiftscited by 0
  38. 2026FailSAE: Towards Interpretable Failure Prediction for Vision-Language Models via Sparse Autoencoderscited by 0

50 most cited of 60 All 60 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models
  2. 2026Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models
  3. 2025Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy
  4. 2025Probing the Representational Power of Sparse Autoencoders in Vision Models
  5. 2026Verifying the Linear Representation Hypothesis: How Interpretable Are Vision SAEs?
  6. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoders
  7. 2026Are Sparse Autoencoder Benchmarks Reliable?
  8. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment