qhjqhj00 medqa-usmle-eval
Domain Fine-Tuning vs. Retrieval-Augmented Generation for Medical Multiple-Choice Question Answering: A Controlled Comparison at the 4B-Parameter Scale — Avi-ad Avraam Buskila et al. (2026) (arXiv:2604.23801, 2026)
What this evaluates
Evaluates medical multiple-choice question answering capability of 4B-parameter LLMs, specifically comparing the impact of domain fine-tuning versus retrieval-augmented generation (RAG) on accuracy.
Datasets
- MedQA-USMLE — total 1273; splits: test (1273)
Metrics
Majority-vote accuracy(primary) — range: [0, 1]- Proportion of correctly answered questions out of the total test set. Answers are generated across three decoding repetitions at temperature T=0.1, and the most frequent option is selected as the final prediction.
Input / output format
Input: Medical multiple-choice questions with four options. For RAG setups, relevant medical textbook passages are injected as context.
Output: A single letter corresponding to the chosen option. The pipeline allows up to two retries if the initial output does not match the expected answer schema.
Scoring recipe
# predictions: list of majority-vote answers per question (length N)
# golds: list of correct option letters (length N)
correct = sum(1 for p, g in zip(predictions, golds) if p == g)
accuracy = correct / len(golds)
Common pitfalls
- RAG does not improve performance and may slightly decrease accuracy for domain-tuned models; the effect is not statistically significant.
- Single-pass decoding is insufficient due to stochasticity; the protocol requires three repetitions at T=0.1 to compute majority-vote accuracy.
- Outputs must strictly match the answer schema; parse failures trigger automatic retries, which affects the effective success rate.
Evidence (verbatim from paper)
Majority-vote accuracy and Wilson 95% confidence intervals for the four setups appear in Table 2 and Figure 1. The domain-tuned backbone (MedGemma 4B) leads both with and without RAG, and the best non-domain setup (Gemma 3 4B + RAG) does not catch up to the worst domain setup (MedGemma 4B + RAG).
Citation
@misc{buskila2026domain,
title={Domain Fine-Tuning vs. Retrieval-Augmented Generation for Medical Multiple-Choice Question Answering: A Controlled Comparison at the 4B-Parameter Scale},
author={Avi-ad Avraam Buskila et al. (2026)},
year={2026},
note={arXiv:2604.23801}
}
- arXiv: 2604.23801