Back
Top 0.6%
14.3%
Top 9%
14.3%
Top 2%
11.6%
Top 0.3%
11.6%
Top 3%
6.3%
Top 3%
5.8%
Top 19%
5.8%
Top 3%
5.5%
Top 5%
3.8%
Top 2%
2.7%
Top 3%
1.9%
Top 95%
1.5%
Top 4%
1.5%
Top 2%
1.5%
Top 7%
0.9%
Top 2%
0.9%
Top 12%
0.9%
Top 13%
0.9%
Top 56%
0.7%
Top 31%
0.7%
Why Large Language Models' Clinical Reasoning Fails: Insights from Explainable Deep Learning
2026-01-27
health informatics
Title + abstract only
View on medRxiv
Show abstract
Medical large language models (LLMs) 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 355 systematic perturbations of physician-validated oncology cases and trained sparse autoencoders on 1 billion tokens from 50,000 MIMIC-IV clinical notes to decompose their internal representation. We find models...
Predicted journal destinations
1
npj Digital Medicine
85 training papers
2
Scientific Reports
701 training papers
3
Journal of the American Medical Informatics Association
53 training papers
4
Journal of Biomedical Informatics
37 training papers
5
PLOS Digital Health
88 training papers
6
JAMIA Open
35 training papers
7
Nature Communications
483 training papers
8
BMC Medical Informatics and Decision Making
36 training papers
9
Journal of Medical Internet Research
81 training papers
10
Computers in Biology and Medicine
39 training papers
11
International Journal of Medical Informatics
25 training papers
12
PLOS ONE
1737 training papers
13
JMIR Medical Informatics
16 training papers
14
Communications Medicine
63 training papers
15
BMC Medical Research Methodology
41 training papers
16
Patterns
15 training papers
17
PLOS Computational Biology
141 training papers
18
Nature Medicine
88 training papers
19
eLife
262 training papers
20
JAMA Network Open
125 training papers