mlsanityData to 14 Oct 2026

Interpreting vision and multimodal models

105 papers0 this weekHigh confidence in this grouping: the papers form one clear topicHigh confidence in this grouping: the papers form one clear topic · About

Keywords: visual, vision, multimodal, vision-language, sparse

Key facts. Interpreting vision and multimodal models: 105 papers on arXiv since 2021 (data through 14 Oct 2026). 30 of its 105 mapped papers (29%) cite “Sparse autoencoders reveal selective remapping of visual concepts during adaptation” (2024, arXiv 2412.05276). 4 new papers 24 Sep–14 Oct.

What’s new

24 Sep – 14 Oct 2026

  1. Earlier · 4
  2. Decoding the Functional Roles of Register and High-Norm Patch Tokens in Vision Transformers
  3. Corrupted but Correct: Why Vision-Language Models Lie to Themselves Internally
  4. From Patching to Pruning Visual Computation in Vision Language Models
  5. Augmenting Visual Anomaly Detection with Automated Interpretability

Foundations

Most cited by papers here

  1. Sparse autoencoders reveal selective remapping of visual concepts during adaptation2024cited by 30 here
  2. Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models2025cited by 20 here
  3. Interpreting CLIP with Hierarchical Sparse Autoencoders2025cited by 16 here
  4. Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment2025cited by 14 here
  5. Steering CLIP's vision transformer with sparse autoencoders2025cited by 13 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Toy Models of Superpositioncited by 30 here
  2. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 14 here
  3. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 52 here
  4. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 11 here
  5. 2024Scaling and evaluating sparse autoencoderscited by 36 here
  6. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 21 here
  7. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 10 here
  8. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 7 here
  9. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignmentcited by 14 here
  10. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Modelscited by 20 here
14 more it builds on, not in this chain
  1. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 18 here
  2. 2023Towards Vision-Language Mechanistic Interpretability: A Causal Tracing Tool for BLIPcited by 8 here
  3. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 8 here
  4. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 9 here
  5. 2023Steering Llama 2 via Contrastive Activation Additioncited by 7 here
  6. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 9 here
  7. 2024Transcoders Find Interpretable LLM Feature Circuitscited by 7 here
  8. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 22 here
  9. 2024Interpreting and Editing Vision-Language Representations to Mitigate Hallucinationscited by 7 here
  10. 2024Sparse autoencoders reveal selective remapping of visual concepts during adaptationcited by 30 here
  11. 2024BatchTopK Sparse Autoencoderscited by 13 here
  12. 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 8 here
  13. 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 16 here
  14. 2025Steering CLIP's vision transformer with sparse autoencoderscited by 13 here

The 24 most cited of 156.

Browse

30 of 105, ranked 6–35

  1. 6Towards Vision-Language Mechanistic Interpretability: A Causal Tracing Tool for BLIP20238
  2. 7Large Multi-modal Models Can Interpret Features in Large Multi-modal Models20247
  3. 8Interpretable and Testable Vision Features via Sparse Autoencoders20256
  4. 9SAE-V: Interpreting Multimodal Models for Enhanced Alignment20255
  5. 10Interpreting the linear structure of vision-language model embedding spaces20255
  6. 11A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models20254
  7. 12How Visual Representations Map to Language Feature Space in Multimodal LLMs20254
  8. 13VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set20254
  9. 14Causal Interpretation of Sparse Autoencoder Features in Vision20253
  10. 15LVLM-Interpret: An Interpretability Tool for Large Vision-Language Models20243
All 30
  1. 16What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Gaussian-Noise-free Text-Image Corruption and Evaluation20243
  2. 17Decomposing multimodal embedding spaces with group-sparse autoencoders20262
  3. 18Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video20252
  4. 19Probing the Representational Power of Sparse Autoencoders in Vision Models20252
  5. 20Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation20252
  6. 21Understanding Video Transformers via Universal Concept Discovery20242
  7. 22Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering20242
  8. 23Circuit Tracing in Vision-Language Models: Understanding the Internal Mechanisms of Multimodal Thinking20261
  9. 24Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers20261
  10. 25SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoders20251
  11. 26M2IV: Towards Efficient and Fine-grained Multimodal In-Context Learning via Representation Engineering20251
  12. 27Analyzing Hierarchical Structure in Vision Models with Sparse Autoencoders20251
  13. 28From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliance20251
  14. 29Line of Sight: On Linear Representations in VLLMs20251
  15. 30Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders20251
  16. 31SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability20251
  17. 32Explaining How Visual, Textual and Multimodal Encoders Share Concepts20251
  18. 33Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alone20251
  19. 34V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models20251
  20. 35LP++: A Surprisingly Strong Linear Probe for Few-Shot CLIP20241