Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pretraining on publicly available biomedical corpora. These works typically claim that such domain-adaptive pretraining (DAPT) improves performance on downstream medical tasks, such as answering medical licensing exam questions. In this paper, we compare seven public “medical” LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting regime for medical question-answering (QA) tasks. For instance, across the tasks and model pairs we consider in the 3-shot setting, medical LLMs only outperform their base models in 12.1% of cases, reach a (statistical) tie in 49.8% of cases, and are significantly worse than their base models in the remaining 38.2% of cases. Our conclusions are based on (i) comparing each medical model head-to-head, directly against the corresponding base model; (ii) optimizing the prompts for each model separately; and (iii) accounting for statistical uncertainty in comparisons. While these basic practices are not consistently adopted in the literature, our ablations show that they substantially impact conclusions. Our findings suggest that state-of-the-art general-domain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.

Summarizing documents with Large Language Models (LLMs) warrants a rigorous inspection of the resulting outputs by humans. However, unaided verification of generated outputs is time-intensive and intractable at scale. For high-stakes applications like healthcare where verification is necessary, expediting this step can unlock massive gains in productivity. In this paper, we focus on the task of evidence extraction for abstractive summarization: for each summary line, extract the corresponding evidence spans from a source document. Viewing this evidence extraction problem through the lens of extractive question answering, we train a set of fast and scalable hierarchical architectures: EarlyFusion, MidFusion, and LateFusion. Our experiments show that (i) our method outperforms the state-of-the-art by 1.4% relative F1-Score; (ii) our model architecture reduces latency by 4x over a RoBERTa-Large baseline; and (iii) pretraining on an extractive QA corpus confers positive transfer to evidence extraction, especially in low-resource regimes.