Previously submitted to: JMIR AI (no longer under consideration since Jun 19, 2026)
Date Submitted: Jun 15, 2026
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Performance of AtlasGPT and OpenEvidence on the United States Medical Licensing Examination: A comparative study of two retrieval-augmented medical AI systems on open-ended USMLE questions
ABSTRACT
Background:
Large language models (LLMs) now routinely meet or exceed the passing threshold on the United States Medical Licensing Examination (USMLE), with systems such as OpenEvidence reported to achieve perfect scores. Despite this, most evaluations use multiple-choice formats and may not reflect performance on open-ended questions which more closely reflect real life scenarios. Specialized systems like OpenEvidence and AtlasGPT seek to improve reliability through retrieval-augmented generation and structured reasoning.
Objective:
To benchmark AtlasGPT against the widely used, high-performing OpenEvidence system by comparing their performance on identical sets of USMLE-style questions across Steps 1, 2, and 3, in an open-ended format.
Methods:
Two LLMs were tested on 366 publicly available USMLE questions. Each model was presented with only the question stem and generated a single final answer. Three physicians scored responses against official keys using a binary rubric. Inter-rater reliability (Fleiss’ κ) and head-to-head accuracy (McNemar’s test) were calculated; accuracy is reported with 95% confidence intervals and head-to-head differences with matched odds ratios.
Results:
Across 366 items, AtlasGPT achieved 71.86% accuracy (263/366; 95% CI, 0.670–0.762) versus OpenEvidence 65.30% (239/366; 95% CI, 0.603–0.700); head-to-head testing favored AtlasGPT (McNemar χ²=7.3472, p=0.0067; matched OR=2.00, 95% CI, 1.23–3.27). Effects were consistent by step—significant in Step 1 (OR=2.71, 95% CI, 1.14–6.46; p=0.031) and not significant in Step 2 (p=0.210) or Step 3 (p=0.405); no heterogeneity was detected (Q=0.843, p=0.656). Inter-rater agreement was substantial (κ: AtlasGPT 0.847, 95% CI, 0.796–0.893; OpenEvidence 0.809, 95% CI, 0.757–0.856); the difference was not significant (Δκ=0.0388, p=0.085).
Conclusions:
AtlasGPT outperformed OpenEvidence on open-ended USMLE-style questions, suggesting that meaningful performance differences exist between advanced RAG systems even when both use curated medical corpora.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.