EUVIMEDSkopri EUVIMED
Open access

Token Probabilities to Mitigate Large Language Models Overconfidence in Answering Medical Questions: Quantitative Study

Raphaël Bentegeac iD, Bastien Le Guellec iD, Grégory Kuchcinski iD, Philippe Amouyel iD, Aghiles Hamroun iD

Journal of Medical Internet Research · 2025-08-29 · 27 · e64348-e64348

DOI: 10.2196/64348 ↗

At a glance

Journal article in Journal of Medical Internet Research, published online on 2025-08-29. Written by Raphaël Bentegeac and 4 co-authors. Institution: Department of Public Health, Lille University, Lille University Hospital Center, avenue du Professeur Emile Laine, Lille, 59037, France. Major MEDLINE topics: Language, Educational Measurement. Cited 22 times so far (Crossref); 38 references. The full text is freely available in PubMed Central.

AstrattSubjectsDettalji biblijografiċiCiteRelated studies

Abstract (original language)

Background

Chatbots have demonstrated promising capabilities in medicine, scoring passing grades for board examinations across various specialties. However, their tendency to express high levels of confidence in their responses, even when incorrect, poses a limitation to their utility in clinical settings.

Objective

The aim of the study is to examine whether token probabilities outperform chatbots’ expressed confidence levels in predicting the accuracy of their responses to medical questions.

Methods

In total, 9 large language models, comprising both commercial (GPT-3.5, GPT-4, and GPT-4o) and open source (Llama 3.1-8b, Llama 3.1-70b, Phi-3-Mini, Phi-3-Medium, Gemma 2-9b, and Gemma 2-27b), were prompted to respond to a set of 2522 questions from the United States Medical Licensing Examination (MedQA database). Additionally, the models rated their confidence from 0 to 100, and the token probability of each response was extracted. The models’ success rates were measured, and the predictive performances of both expressed confidence and response token probability in predicting response accuracy were evaluated using area under the receiver operating characteristic curves (AUROCs), adapted calibration error, and Brier score. Sensitivity analyses were conducted using additional questions sourced from other databases in English (MedMCQA: n=2797), Chinese (MedQA Mainland China: n=3413 and Taiwan: n=2808), and French (FrMedMCQA: n=1079), different prompting strategies, and temperature settings.

Results

Overall, mean accuracy ranged from 56.5% (95% CI 54.6‐58.5) for Phi-3-Mini to 89% (95% CI 87.7‐90.2) for GPT-4o. Across the United States Medical Licensing Examination questions, all chatbots consistently expressed high levels of confidence in their responses (ranging from 90, 95% CI 90-90 for Llama 3.1-70b to 100, 95% CI 100-100 for GPT-3.5). However, expressed confidence failed to predict response accuracy (AUROC ranging from 0.52, 95% CI 0.50‐0.53 for Phi-3-Mini to 0.68, 95% CI 0.65‐0.71 for GPT-4o). In contrast, the response token probability consistently outperformed expressed confidence for predicting response accuracy (AUROCs ranging from 0.71, 95% CI 0.69‐0.73 for Phi-3-Mini to 0.87, 95% CI 0.85‐0.89 for GPT-4o; all P

Subjects

MeSH terms (MEDLINE)

LanguageEducational MeasurementHumansProbabilityUnited StatesLarge Language Models

Author keywords

ChatGPTMedQANLPaccuracyartificial intelligencechatbotconfidencelanguage modellarge language modelmachine learningmedical questionmedicinenatural language processingprobabilityquestionnaire

Authors and institutions

  • Department of Public Health, Lille University, Lille University Hospital Center, avenue du Professeur Emile Laine, Lille, 59037, France.
  • Univ. Lille, Inserm, Centre Hosp. Univ Lille, Institut Pasteur de Lille, UMR1167-Labex DISTALZ-RID-AGE-Risk Factors and Molecular Determinants of Aging-Related Diseases, Lille, France.
  • Department of Neuroradiology, Lille University, Lille University Hospital Center, Lille, France.
  • Univ. Lille, Inserm, CHU Lille, U1172-LilNCog-Lille Neuroscience & Cognition, Lille, France.
  • Univ. Lille, CNRS, Inserm, CHU Lille, Institut Pasteur de Lille, US 41 - UAR 2014 - PLBS, Lille, France.

ORCID profiles: Raphaël Bentegeac, Bastien Le Guellec, Grégory Kuchcinski, Philippe Amouyel, Aghiles Hamroun

Cite

APA

Bentegeac, R., Le Guellec, B., Kuchcinski, G., Amouyel, P., & Hamroun, A. (2025). Token Probabilities to Mitigate Large Language Models Overconfidence in Answering Medical Questions: Quantitative Study. Journal of Medical Internet Research, 27, e64348-e64348. https://doi.org/10.2196/64348

Vancouver

Bentegeac R, Le Guellec B, Kuchcinski G, Amouyel P, Hamroun A. Token Probabilities to Mitigate Large Language Models Overconfidence in Answering Medical Questions: Quantitative Study. Journal of Medical Internet Research. 2025;27:e64348-e64348. doi:10.2196/64348

BibTeX
@article{Bentegeac2025,
  author = {Bentegeac, Raphaël and Le Guellec, Bastien and Kuchcinski, Grégory and Amouyel, Philippe and Hamroun, Aghiles},
  title = {Token Probabilities to Mitigate Large Language Models Overconfidence in Answering Medical Questions: Quantitative Study},
  journal = {Journal of Medical Internet Research},
  year = {2025},
  volume = {27},
  pages = {e64348--e64348},
  doi = {10.2196/64348}
}
Start for free →