aiqbench analysis
This arXiv study tests whether LLMs' strong scores on medical exam benchmarks reflect clinical competence or format effects and memorization. Accuracy drops sharply when answer choices are withheld (82.3% to 50.9%), and the strongest models can reconstruct about one in four questions from the choices alone. For readers, the key takeaway is that headline medical benchmark results need scrutiny for contamination and format sensitivity, and smaller locally deployable models still lag their larger counterparts.
From Small to Large: Assessing Medical Reasoning, Benchmark Contamination, and Model Size in Language Models
arXiv:2611.00460v1 Announce Type: new Abstract: Large language models (LLMs) now answer medical exam questions at or above expert level, but it is unclear whether such scores reflect clinical competence, or are instead affected by the format of the questions being asked (e.g., multiple-choice versus open-ended), prior exposure to the benchmarks, and the difficulty of the questions. These benchmarks are also public and widely used online, which may be resulting in LLM memorization. Models may perform better on questions that they have seen before. It is also not evident whether smaller models, which are more locally deployable in HIPAA-compliant hospital settings, have achieved human performance. In this study, we evaluate fourteen current models from six leading providers, one smaller and one larger model from each provider with two additional open-weight models, on 3,076 questions from four medical benchmarks (MedQA-USMLE, MedMCQA, PubMedQA, and MedBullets) under standard multiple choice, the same questions with the answer choices withheld, and reconstruction of a hidden question from its choices alone. In the standard multiple-choice setting, the strongest models exceed the human-expert baseline on MedQA-USMLE and PubMedQA, but the smaller models trail their larger models by 8.7 percentage points and only two of the six pass the human-expert bar on the licensing examination. Withholding the answer choices, however, makes every model struggle: mean accuracy falls from 82.3% to 50.9%, and the two largest drops belong to the medically specialised models. On memorization, the strongest models reconstruct one question in four from the answer choices alone, yet accuracy holds on the questions they cannot reconstruct. We further weigh accuracy against inference cost and test whether ensembling several models improves medical question answering.
Abstract:Large language models (LLMs) now answer medical exam questions at or above expert level, but it is unclear whether such scores reflect clinical competence, or are instead affected by the format of the questions being asked (e.g., multiple-choice versus open-ended), prior exposure to the benchmarks, and the difficulty of the questions. These benchmarks are also public and widely used online, which may be resulting in LLM memorization. Models may perform better on questions that they have seen before. It is also not evident whether smaller models, which are more locally deployable in HIPAA-compliant hospital settings, have achieved human performance. In this study, we evaluate fourteen current models from six leading providers, one smaller and one larger model from each provider with two additional open-weight models, on 3,076 questions from four medical benchmarks (MedQA-USMLE, MedMCQA, PubMedQA, and MedBullets) under standard multiple choice, the same questions with the answer choices withheld, and reconstruction of a hidden question from its choices alone. In the standard multiple-choice setting, the strongest models exceed the human-expert baseline on MedQA-USMLE and PubMedQA, but the smaller models trail their larger models by 8.7 percentage points and only two of the six pass the human-expert bar on the licensing examination. Withholding the answer choices, however, makes every model struggle: mean accuracy falls from 82.3% to 50.9%, and the two largest drops belong to the medically specialised models. On memorization, the strongest models reconstruct one question in four from the answer choices alone, yet accuracy holds on the questions they cannot reconstruct. We further weigh accuracy against inference cost and test whether ensembling several models improves medical question answering.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2611.00460 [cs.AI]
(or arXiv:2611.00460v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2611.00460 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 367 words
Related coverage
Quantifying the Effect of Test Set Contamination on Generative Evaluations
arXiv · Sep 30, 2026
When Large Language Models Know the Table: A Framework for Assessing Data Contamination in Tabular Datasets
arXiv · Aug 6, 2026
Robustness of Anomaly Detection Models for Industrial Control Systems under Training-Time Data Contamination
arXiv · Aug 25, 2026
Prism-SQA: An Interpretable and Adaptable Neural Framework for Surface Electromyography Quality Assessment
arXiv · Sep 14, 2026