Insights into a radiology-specialised multimodal large language model with sparse autoencoders
On the map: Multimodal LLMs / vision-language models → Medical vision-language models
Abstract (arXiv)
Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.
Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Medical vision-language models, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 35 of its 54 references that are in the map. First posted on arXiv 17 Jul 2025.
Builds on
35 of 54 references.
Closest ancestors · 2
Same direction, newest first
- 2024Worse than Random? An Embarrassingly Simple Probing Evaluation of Large Multimodal Models in Medical VQA30 May
- 2023MAIRA-1: A specialised large multimodal model for radiology report generation22 Nov
Other references · 33
- 2023Visual Instruction Tuningcited by 7,869
- 2019CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparisoncited by 1,423
- 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
- 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
- 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
- 2023Steering Language Models With Activation Engineeringcited by 922
- 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
- 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
- 2023Towards Generalist Biomedical AIcited by 289
- 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 160
- 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 146
- 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
21 more
- 2024Exploring scalable medical image encoders beyond text supervisioncited by 132
- 2024BatchTopK Sparse Autoencoderscited by 130
- 2023Dynamic Graph Enhanced Contrastive Learning for Chest X-ray Report Generationcited by 118
- 2023METransformer: Radiology Report Generation by Transformer with Multiple Learnable Expert Tokenscited by 101
- 2024Steering Language Model Refusal with Sparse Autoencoderscited by 95
- 2024Advancing Multimodal Medical Capabilities of Geminicited by 90
- 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Modelscited by 71
- 2024InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoderscited by 66
- 2024A Multimodal Automated Interpretability Agentcited by 63
- 2024Multimodal Healthcare AI: Identifying and Designing Clinically Relevant Vision-Language Applications for Radiologycited by 45
- 2025Enhancing Automated Interpretability with Output-Centric Feature Descriptionscited by 41
- 2024A Concept-Based Explainability Framework for Large Multimodal Modelscited by 39
- 2024Efficient Dictionary Learning with Switch Sparse Autoencoderscited by 39
- 2024An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generationcited by 25
- 2025Interpretable and Testable Vision Features via Sparse Autoencoderscited by 25
- 2024Relational Composition in Neural Networks: A Survey and Call to Actioncited by 23
- 2024Large Multi-modal Models Can Interpret Features in Large Multi-modal Modelscited by 23
- 2025SAE-V: Interpreting Multimodal Models for Enhanced Alignmentcited by 23
- 2020Trust and Medical AI: The challenges we face and the expertise needed to overcome themcited by 21
- 2024Learning biologically relevant features in a pathology foundation model using sparse autoencoderscited by 14
- 2024GPT-4V Cannot Generate Radiology Reports Yetcited by 7
Built on by
Citing papers per year: 2025: 1; 2026: 1.
2 papers.
Similar, not cited
Close in wording, no citation either way
- 2025MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders
- 2026Volumetric Radiology AI in the Era of Multimodal Large Language Models
- 2025Resurrecting the Salmon: Rethinking Mechanistic Interpretability with Domain-Specific Sparse Autoencoders
- 2026RadLite: Multi-Task LoRA Fine-Tuning of Small Language Models for CPU-Deployable Radiology AI
- 2025mAIstro: an open-source multi-agentic system for automated end-to-end development of radiomics and deep learning models for medical imaging
- 2026RadSight: Towards Perceptually Reliable Multimodal Radiology Image Understanding
- 2026Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models
- 2025Vision-Language Models for Automated Chest X-ray Interpretation: Leveraging ViT and GPT-2