Enhancing domain adaptation of LLM via model composition in solving medical exam questions

he advent of large language models (LLMs) has sparked interest among educators across various fields, particularly in medical education. However, general-purpose pretrained LLMs may perform inadequately on medical examination questions without domain adaptation. Full fine-tuning of a pretrained LLM requires substantial computing resources, making it impractical for resource-constrained settings. In this work, we investigate a Composition to Augment Language Models (CALM) framework for medical domain adaptation in answering real-world medical examination questions. Specifically, we fine-tune a small pretrained language model on an open-ended medical question-answering dataset to inject medical domain knowledge, and then develop MedCALM by composing this medical-specialized augmenting model with an anchor LLM. Experiments on real-world medical exam question datasets show that MedCALM outperforms the compared baselines. The results indicate that using a multi-head cross-attention module to connect two LLMs improves performance on multiple-choice question answering. In addition, the selection of connected layers significantly affects the effectiveness of composition between the anchor and augmenting models. Our experiments suggest that cross-attention between the last few layers of the two models achieves promising performance.