mlsanityData to 14 Oct 2026

MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders

arXiv 2510.26411 ↗First posted 30 Oct 2025cs.AIcited by 2

On the map: Mechanistic interpretability → Interpreting biomedical and science models

Abstract (arXiv)

Artificial intelligence in healthcare requires models that are accurate and interpretable. We advance mechanistic interpretability in medical vision by applying Medical Sparse Autoencoders (MedSAEs) to the latent space of MedCLIP, a vision-language model trained on chest radiographs and reports. To quantify interpretability, we propose an evaluation framework that combines correlation metrics, entropy analyses, and automated neuron naming via the MedGemma foundation model. Experiments on the CheXpert dataset show that MedSAE neurons achieve higher monosemanticity and interpretability than raw MedCLIP features. Our findings bridge high-performing medical AI and transparency, offering a scalable step toward clinically reliable representations. The source code supporting the findings of this study is available at https://github.com/EIDOSLAB/MedSAE.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Interpreting biomedical and science models, one of 13 directions in Mechanistic interpretability. Builds on 16 of its 23 references that are in the map. First posted on arXiv 30 Oct 2025.

Builds on

16 of 23 references.

Closest ancestors · 2

Same direction, newest first

  1. 2024An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generation4 Oct
  2. 2024Learning biologically relevant features in a pathology foundation model using sparse autoencoders15 Jul

Other references · 14

  1. 2019CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparisoncited by 1,423
  2. 2017A Survey on Deep Learning in Medical Image Analysiscited by 1,186
  3. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  4. 2022MedCLIP: Contrastive Learning from Unpaired Medical Images and Textcited by 570
  5. 2024Scaling and evaluating sparse autoencoderscited by 553
  6. 2025MedGemma Technical Reportcited by 436
  7. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  8. 2023Multimodal Large Language Models: A Surveycited by 195
  9. 2024Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)cited by 119
  10. 2024Automatically Interpreting Millions of Features in Large Language Modelscited by 113
  11. 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
  12. 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 60
2 more
  1. 2025Mechanistic understanding and validation of large AI models with SemanticLenscited by 29
  2. 2024Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoderscited by 18

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoderscited by 5
  2. 2026Medical AI Encodes a "Feeling of Error": Verifying Cancer Segmentation via Internal Conceptscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Sparse Autoencoders for Interpretable Medical Image Representation Learning
  2. 2025Insights into a radiology-specialised multimodal large language model with sparse autoencoders
  3. 2024Envisioning MedCLIP: A Deep Dive into Explainability for Medical Vision-Language Models
  4. 2026MedConcept: Unsupervised Concept Discovery for Interpretability in Medical VLMs
  5. 2025Resurrecting the Salmon: Rethinking Mechanistic Interpretability with Domain-Specific Sparse Autoencoders
  6. 2025Mammo-SAE: Interpreting Breast Cancer Concept Learning with Sparse Autoencoders
  7. 2025An Explainable Biomedical Foundation Model via Large-Scale Concept-Enhanced Vision-Language Pre-training
  8. 2026Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models