理解医疗大语言模型在临床推理中的变异性:一项机制可解释性研究
Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study
研究评估了GPT-5、MedGemma-27B-Text-IT和OpenBioLLM-Llama3-70B在3035个系统性扰动下的临床推理稳定性,发现模型在提示格式变化下推理能力波动显著,内部编码结构影响检索和推理干预效果。
- Article
- Open access
- Published: 31 October 2026
- Mirage Modi ORCID: orcid.org/0000-0001-5998-21071,
- Jordan E. Krull ORCID: orcid.org/0000-0001-6507-80851,2,
- Timothy D. Gauntner2,3,
- Mingjia Li3,
- Donte Johnson1,
- Xiaoying Wang1,2,
- Hao Cheng1,
- Jing Zhao ORCID: orcid.org/0000-0002-2378-63711,
- Anjun Ma ORCID: orcid.org/0000-0001-6269-398X1,2,
- Ping Zhang1,4,
- Daniel G. Stover ORCID: orcid.org/0000-0001-9003-81653,
- Zihai Li ORCID: orcid.org/0000-0003-4603-927X2,3 &
- …
- Qin Ma ORCID: orcid.org/0000-0002-3264-83921,2
npj Digital Medicine (2026) Cite this article
We’re sharing this article early to provide faster access to peer-reviewed, accepted research. It is citable and carries a permanent DOI. This version is subject to further edits and will be replaced automatically by the final Version of Record. All legal disclaimers apply.
Abstract
Medical large language models achieving high benchmark accuracy exhibit unexplained variability in clinical tasks, producing errors that clinicians cannot safeguard against. We evaluated clinical reasoning stability in GPT-5, MedGemma-27B-Text-IT, and OpenBioLLM-Llama3-70B using 3035 systematic perturbations from physician-validated oncology SOAP, brief vignette, and MIMIC cases. We trained sparse autoencoders on 1 billion tokens from 50,000 MIMIC-IV clinical notes to decompose their internal representation. We find models exhibit dramatic reasoning instability, shifting staging accuracy by over 50% based solely on prompt format, or generating definitive staging in clinically insufficient scenarios. Sparse autoencoder analysis revealed hierarchical encoding in MedGemma, where high-magnitude features encode lexical identity and low-magnitude features encode contextual meaning. OpenBioLLM distributes information uniformly. We demonstrate that these internal encoding structures differentially affect retrieval and reasoning interventions, suggesting interventions effective for one architecture may harm another. These results motivate architecture-specific safety validation as a component of pre-deployment evaluation for medical LLMs.
Acknowledgements
We thank Dr. Fei He for his feedback regarding the manuscript.
Funding
This work was funded by the National Institutes of Health (T32GM139784 and P01CA278732) and by the Pelotonia Institute for Immuno-Oncology (PIIO). The content is solely the responsibility of the authors and does not necessarily represent the official views of NIH and PIIO. The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.
Author information
Authors and Affiliations
Department of Biomedical Informatics, The Ohio State University College of Medicine, Columbus, OH, USA
Mirage Modi, Jordan E. Krull, Donte Johnson, Xiaoying Wang, Hao Cheng, Jing Zhao, Anjun Ma, Ping Zhang & Qin Ma
Pelotonia Institute for Immuno-Oncology, The Ohio State University Comprehensive Cancer Center, College of Medicine, The Ohio State University Wexner Medical Center, Columbus, OH, USA
Jordan E. Krull, Timothy D. Gauntner, Xiaoying Wang, Anjun Ma, Zihai Li & Qin Ma
Division of Medical Oncology, Department of Internal Medicine, The Ohio State University Comprehensive Cancer Center – James Cancer Hospital and Solove Research Institute, Columbus, OH, USA
Timothy D. Gauntner, Mingjia Li, Daniel G. Stover & Zihai Li
Department of Computer Science and Engineering, The Ohio State University, Columbus, OH, USA
Ping Zhang
Authors
- Mirage Modi
- Jordan E. Krull
- Timothy D. Gauntner
- Mingjia Li
- Donte Johnson
- Xiaoying Wang
- Hao Cheng
- Jing Zhao
- Anjun Ma
- Ping Zhang
- Daniel G. Stover
- Zihai Li
- Qin Ma
Corresponding authors
Correspondence to Jordan E. Krull or Qin Ma.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Modi, M., Krull, J.E., Gauntner, T.D. et al. Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study. npj Digit. Med. (2026). https://doi.org/10.1038/s41746-026-03423-6
Received: 20 February 2026
Accepted: 06 October 2026
Published: 31 October 2026
DOI: https://doi.org/10.1038/s41746-026-03423-6
来源:Nature 健康护理主题 · nature.com