Skip to main navigation Skip to search Skip to main content

DeepSeek R1 Distilled Fails to Perform Well Against the USMLE and Other LLMs with and Without Semantics

  • Peter L. Elkin
  • , Guresh Mehta
  • , Aaron N. Elkin
  • , Jonathan R. Nebeker
  • , Steven H. Brown
  • SUNY Buffalo
  • Knowledge Based Systems, Inc.
  • Vanderbilt University

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

DeepSeek-R1 Distilled Large Language model (LLM) was touted to be almost as good and much cheeper to generate than traditional LLMs. We compared its ability to answer medical questions from the United States Medical Licensing Examinations (USMLE) and found that the performance drop from the original model to DeepSeek was significant. DeepSeek is not yet a good alternative for medical question answering.

Original languageEnglish
Title of host publicationOpening the Personal Gate between Technology and Health Care - Proceedings of MIE 2026
EditorsMaria Hagglund, Lars Lindskold, Lenka Lhotska, Sara Marceglia, Enea Parimbelli, Lucia Sacchi, Paolo Soda, Lacramioara Stoicu-Tivadar, Pierangelo Veltri, Patrizia Vizza, Mauro Giacomini, Jaime Delgado, Theodoros N. Arvanitis, Elisavet Andrikopoulou, Arriel Benis, Gabriella Balestra, Riccardo Bellazzi, Parisis G. Gallos, Roberto Gatta, Daniele Roberto Giacobbe, Noemi Giordano
PublisherIOS Press BV
Pages510-511
Number of pages2
ISBN (Electronic)9781643686615
DOIs
StatePublished - May 21 2026
Event36th Medical Informatics Europe Conference, MIE 2026 - Genoa, Italy
Duration: May 25 2026May 28 2026

Publication series

NameStudies in Health Technology and Informatics
Volume336
ISSN (Print)0926-9630
ISSN (Electronic)1879-8365

Conference

Conference36th Medical Informatics Europe Conference, MIE 2026
Country/TerritoryItaly
CityGenoa
Period05/25/2605/28/26

Keywords

  • AI
  • ChatBots
  • Clinical Decision Support
  • Machine Learning

Fingerprint

Dive into the research topics of 'DeepSeek R1 Distilled Fails to Perform Well Against the USMLE and Other LLMs with and Without Semantics'. Together they form a unique fingerprint.

Cite this