Journal article · 2026 · Open access

Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study

Mirage Modi, Jordan Edward Krull, Timothy D. Gauntner, Mingjia Li, Donte Johnson, Xiaoying Wang, Hao Cheng, Jing Zhao, Qin Ma, Ping Zhang, Daniel G. Stover, Zihai Li and 1 more

npj Digital Medicine · Published 31 October 2026 · DOI 10.1038/s41746-026-03423-6

Open PDFView at the publisherCite this paper

In brief

“Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study” is a 2026 journal article by Mirage Modi and 12 co-authors, published in npj Digital Medicine. OpenAlex counts 4 citations as of 17 November 2026, which puts it in the top 3% of works published in 2026 by citations. It is open access under a CC BY-NC-ND licence: the full text is free to read at npj Digital Medicine.

Citations
4
Most cited in
2026 (4)
Field-weighted impact
4.81
References
35
Authors
13

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.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Modi, M., Krull, J. E., Gauntner, T. D., Li, M., Johnson, D., Wang, X., Cheng, H., Zhao, J., Ma, Q., Zhang, P., Stover, D. G., Li, Z., & Ma, Q. (2026). Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study. npj Digital Medicine. https://doi.org/10.1038/s41746-026-03423-6

MLA 9

Modi, Mirage, et al. “Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study.” npj Digital Medicine, 2026. https://doi.org/10.1038/s41746-026-03423-6.

Chicago

Modi, Mirage, Jordan Edward Krull, Timothy D. Gauntner, Mingjia Li, Donte Johnson, Xiaoying Wang, Hao Cheng, et al. 2026. “Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study.” npj Digital Medicine. https://doi.org/10.1038/s41746-026-03423-6.

Harvard

Modi, M. et al. (2026) ‘Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study’, npj Digital Medicine. doi:10.1038/s41746-026-03423-6.

IEEE

M. Modi et al., “Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study,” npj Digital Medicine, 2026, doi: 10.1038/s41746-026-03423-6.

BibTeX

@article{modi2026understanding,
  title     = {{Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study}},
  author    = {Modi, Mirage and Krull, Jordan Edward and Gauntner, Timothy D. and Li, Mingjia and Johnson, Donte and Wang, Xiaoying and Cheng, Hao and Zhao, Jing and Ma, Qin and Zhang, Ping and Stover, Daniel G. and Li, Zihai and Ma, Qin},
  journal   = {npj Digital Medicine},
  year      = {2026},
  doi       = {10.1038/s41746-026-03423-6},
  url       = {https://doi.org/10.1038/s41746-026-03423-6}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education

  • medical large language models
  • mechanistic interpretability
  • sparse autoencoders
  • tumor staging
  • MedGemma
  • GPT-5
  • Llama 3

Work with this paper

Most-cited papers in Medicine

All highly cited Medicine papers

Questions about this paper

How do I cite this paper in APA style?

Modi, M., Krull, J. E., Gauntner, T. D., Li, M., Johnson, D., Wang, X., Cheng, H., Zhao, J., Ma, Q., Zhang, P., Stover, D. G., Li, Z., & Ma, Q. (2026). Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study. npj Digital Medicine. https://doi.org/10.1038/s41746-026-03423-6

How many times has this paper been cited?

4 times, according to OpenAlex on 17 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at npj Digital Medicine: sign in and open the PDF from this page.

Who wrote this paper?

Mirage Modi, Jordan Edward Krull, Timothy D. Gauntner, Mingjia Li, Donte Johnson, Xiaoying Wang, Hao Cheng, Jing Zhao, Qin Ma, Ping Zhang, Daniel G. Stover, Zihai Li and 1 more.

What is the DOI of this paper?

10.1038/s41746-026-03423-6, permanently at https://doi.org/10.1038/s41746-026-03423-6. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 17 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.