mlsanityData to 7 Oct 2026

Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models

arXiv 2504.02821 ↗First posted 3 Apr 2025cs.CV cs.AI cs.LGcited by 70

On the map: Multimodal LLMs / vision-language models → VLM benchmarks and visual reasoning

Abstract (arXiv)

Sparse Autoencoders (SAEs) have recently gained attention as a means to improve the interpretability and steerability of Large Language Models (LLMs), both of which are essential for AI safety. In this work, we extend the application of SAEs to Vision-Language Models (VLMs), such as CLIP, and introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in visual representations. To ensure that our evaluation aligns with human perception, we propose a benchmark derived from a large-scale user study. Our experimental results reveal that SAEs trained on VLMs significantly enhance the monosemanticity of individual neurons, with sparsity and wide latents being the most influential factors. Further, we demonstrate that applying SAE interventions on CLIP's vision encoder directly steers multimodal LLM outputs (e.g., LLaVA), without any modifications to the underlying language model. These findings emphasize the practicality and efficacy of SAEs as an unsupervised tool for enhancing both interpretability and control of VLMs. Code and benchmark data are available at https://github.com/ExplainableML/sae-for-vlm.

Key facts. Cited by 70 papers in the map (data through 7 Oct 2026). Placed in VLM benchmarks and visual reasoning, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 44 of its 62 references that are in the map. First posted on arXiv 3 Apr 2025.

Builds on

44 of 62 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024COSMOS: Cross-Modality Self-Distillation for Vision Language Pre-training2 Dec

Other references · 43

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  2. 2020Denoising Diffusion Probabilistic Modelscited by 20,851
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  4. 2023Visual Instruction Tuningcited by 7,792
  5. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,764
  6. 2023Sigmoid Loss for Language Image Pre-Trainingcited by 3,048
  7. 2023InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuningcited by 2,855
  8. 2023MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGIcited by 2,044
  9. 2023A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questionscited by 1,582
  10. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
  11. 2022Toy Models of Superpositioncited by 928
  12. 2022When and why vision-language models behave like bags-of-words, and what to do about it?cited by 603
31 more
  1. 2024Scaling and evaluating sparse autoencoderscited by 530
  2. 2021SLIP: Self-supervision meets Language-Image Pre-trainingcited by 434
  3. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 283
  4. 2022Matryoshka Representation Learningcited by 268
  5. 2013k-Sparse Autoencoderscited by 209
  6. 2023RSGPT: A Remote Sensing Vision Language Model and Benchmarkcited by 204
  7. 2023BiomedGPT: A Generalist Vision-Language Foundation Model for Diverse Biomedical Taskscited by 201
  8. 2023Interpreting CLIP's Image Representation via Text-Based Decompositioncited by 189
  9. 2021Benchmarking Representation Learning for Natural World Image Collectionscited by 167
  10. 2016Linear Algebraic Structure of Word Senses, with Applications to Polysemycited by 165
  11. 2022CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networkscited by 144
  12. 2024BatchTopK Sparse Autoencoderscited by 126
  13. 2024Vision-Language Models for Medical Report Generation and Visual Question Answering: A Reviewcited by 121
  14. 2024Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)cited by 116
  15. 2024Multimodal Autoregressive Pre-training of Large Vision Encoderscited by 108
  16. 2023Vision-by-Language for Training-Free Compositional Image Retrievalcited by 102
  17. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
  18. 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 75
  19. 2024Sparse autoencoders reveal selective remapping of visual concepts during adaptationcited by 66
  20. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignmentcited by 62
  21. 2025Scaling Language-Free Visual Representation Learningcited by 62
  22. 2024FLAIR: VLM with Fine-grained Language-informed Image Representationscited by 47
  23. 2023Rosetta Neurons: Mining the Common Units in a Model Zoocited by 46
  24. 2024A Concept-Based Explainability Framework for Large Multimodal Modelscited by 39
  25. 2024Interpreting the Second-Order Effects of Neurons in CLIPcited by 36
  26. 2024One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Modelscited by 30
  27. 2024textitRevelio: Interpreting and leveraging semantic information in diffusion modelscited by 29
  28. 2025Interpretable and Testable Vision Features via Sparse Autoencoderscited by 24
  29. 2024Large Multi-modal Models Can Interpret Features in Large Multi-modal Modelscited by 23
  30. 2025Analyzing Finetuning Representation Shift for Multimodal LLMs Steeringcited by 15
  31. 2023DISCOVER: Making Vision Networks Interpretable via Competition and Dissectioncited by 8

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 10; 2026: 60.

70 papers.

Sort:

By area:

  1. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 22
  2. 2025Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEscited by 14
  3. 2026Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Modelscited by 14
  4. 2026Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Modelscited by 14
  5. 2026Cross-Modal Redundancy and the Geometry of Vision-Language Embeddingscited by 9
  6. 2026Arbitration Failure, Not Perceptual Blindness: How Vision-Language Models Resolve Visual-Linguistic Conflictscited by 9
  7. 2025Vision-and-Language Training Helps Deploy Taxonomic Knowledge but Does Not Fundamentally Alter Itcited by 6
  8. 2025Interpretable and Steerable Concept Bottleneck Sparse Autoencoderscited by 6
  9. 2026Interpretable Debiasing of Vision-Language Models for Social Fairnesscited by 6
  10. 2025From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliancecited by 4
  11. 2025Causal Interpretation of Sparse Autoencoder Features in Visioncited by 4
  12. 2026Probing Cross-modal Information Hubs in Audio-Visual LLMscited by 4
38 more
  1. 2025Sparse Autoencoders, Again?cited by 3
  2. 2025Model-Agnostic Gender Bias Control for Text-to-Image Generation via Sparse Autoencodercited by 3
  3. 2026Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformerscited by 3
  4. 2025Insights into a radiology-specialised multimodal large language model with sparse autoencoderscited by 2
  5. 2026Decomposing multimodal embedding spaces with group-sparse autoencoderscited by 2
  6. 2026Towards Interpretable Hallucination Analysis and Mitigation in LVLMs via Contrastive Neuron Steeringcited by 2
  7. 2026ELROND: Exploring and decomposing intrinsic capabilities of diffusion modelscited by 2
  8. 2026Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Imagescited by 2
  9. 2026Improving Sparse Autoencoder with Dynamic Attentioncited by 2
  10. 2026On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoderscited by 2
  11. 2026Diagnosing Visual Ignorance in Vision-Language Modelscited by 2
  12. 2026Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoderscited by 2
  13. 2026Evaluating the Interpretability of Sparse Autoencoders with Concept Annotationscited by 2
  14. 2026LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discoverycited by 1
  15. 2026MonoLoss: A Training Objective for Interpretable Monosemantic Representationscited by 1
  16. 2026Beyond Semantics: Disentangling Information Scope in Sparse Autoencoders for CLIPcited by 1
  17. 2026From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Modelscited by 1
  18. 2026GeoSAE: Geometric Prior-Guided Layer-Wise Sparse Autoencoder Annotation of Brain MRI Foundation Modelscited by 1
  19. 2026Superposition Is Not Necessary: A Mechanistic Interpretability Analysis of Transformer Representations for Time Series Forecastingcited by 1
  20. 2026Event-Grounded Sparse Autoencoders for Vision-Language-Action Policiescited by 1
  21. 2026Interpretability Transfer from Language to Vision via Sparse Autoencoderscited by 1
  22. 2026Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoderscited by 1
  23. 2026Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?cited by 1
  24. 2026What's in an Earth Embedding? An Explainability Analysis of Location Encoderscited by 1
  25. 2026DiMaS: Distribution Matching for Steering Vision-Language-Action Modelscited by 1
  26. 2026What Do They See? Interpreting Complex Road Scenarios Through the Eyes of Vision-Language-Action Models for Safe and Trustworthy Autonomous Vehicle Learningcited by 1
  27. 2026Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherencecited by 1
  28. 2025TRANSPORTER: Transferring Visual Semantics from VLM Manifoldscited by 0
  29. 2025SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networkscited by 0
  30. 2026TimeSAE: Causal Sparse Decoding for Faithful Explanations of Black-Box Time Series Modelscited by 0
  31. 2026RAIGen: Rare Attribute Identification in Text-to-Image Generative Modelscited by 0
  32. 2026Automatic Image-Level Morphological Trait Annotation for Organismal Imagescited by 0
  33. 2026Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMscited by 0
  34. 2026From Heads to Neurons: Causal Attribution and Steering in Multi-Task Vision-Language Modelscited by 0
  35. 2026LatentDiff: Scaling Semantic Dataset Comparison to Millions of Imagescited by 0
  36. 2026Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learningcited by 0
  37. 2026Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Modelscited by 0
  38. 2026Interpretability Without Tradeoffs: Disentangling Polysemanticity At Equal Predictive Performancecited by 0

50 most cited of 70 All 70 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Probing the Representational Power of Sparse Autoencoders in Vision Models
  2. 2025Evaluating Sparse Autoencoders for Monosemantic Representation
  3. 2025Explaining How Visual, Textual and Multimodal Encoders Share Concepts
  4. 2025Interpreting CLIP with Hierarchical Sparse Autoencoders
  5. 2025Interpreting the linear structure of vision-language model embedding spaces
  6. 2026Steering Vision-Language Models with Joint Sparse Autoencoders
  7. 2025Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words
  8. 2026Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models