Due to necessary scheduled maintenance, the JMIR Publications website will be unavailable from Wednesday, July 01, 2020 at 8:00 PM to 10:00 PM EST. We apologize in advance for any inconvenience this may cause you.
Who will be affected?
Readers: No access to all 28 journals. We recommend accessing our articles via PubMed Central
Authors: No access to the submission form or your user account.
Reviewers: No access to your user account. Please download manuscripts you are reviewing for offline reading before Wednesday, July 01, 2020 at 7:00 PM.
Editors: No access to your user account to assign reviewers or make decisions.
Copyeditors: No access to user account. Please download manuscripts you are copyediting before Wednesday, July 01, 2020 at 7:00 PM.
Application of Sparse Autoencoders to Enhance Mechanistic Interpretability of Large Language Models in Medicine
Andre Metzger;
Shiv Patil;
Lauren Sugarmann;
Mert Karabacak;
Konstantinos Margetis
ABSTRACT
Large language models (LLMs) are being increasingly incorporated into clinical workflows due to their ability to synthesize medical knowledge and support diagnosis and treatment planning. However, their opaque internal decision-making processes limit trust, reliability, and safe clinical adoption. Mechanistic interpretability seeks to address this challenge by revealing how LLMs transform inputs into outputs. This paper explores the use of sparse autoencoders (SAEs) as a promising approach to improving mechanistic interpretability of LLMs in medicine. We discuss how SAE-based analyses can illuminate model reasoning, detect potential failure modes, and complement existing interpretability frameworks. Improving mechanistic interpretability through SAEs may be essential for safely deploying LLMs as trustworthy cognitive aids in clinical medicine.
Citation
Please cite as:
Metzger A, Patil S, Sugarmann L, Karabacak M, Margetis K
Application of Sparse Autoencoders to Enhance Mechanistic Interpretability of Large Language Models in Medicine