KB2Bench: Toward a Benchmark Framework for Large Language Models on Medical Knowledge

While Large Language Models (LLMs) have trans-formed question answering tasks, their propensity for hallucinations continues to drive an area of active research. Efforts toward creating benchmarks to test LLMs' performance on queries, in particular for the field of medicine, have led to a few reputable benchmarks, but these are limited in scope because of the amount of human annotation required. Our framework addresses this issue by leveraging existing, large knowledge bases for medicine to generate vast query and answer sets dynamically, which are less likely to be memorized by LLMs. The framework rests on designing a few key knowledge patterns, which can then generate millions (potentially billions) of queries. This offers a more efficient, cost-effective, and scalable alternative to human-curated annotations used in medical question-and-answer benchmarks. Applying our framework to a small sample of five drug related ontologies, we are already capable of more than 100,000 unique drug related queries, which is 10 to 1000 times larger than existing various human annotation efforts. This paper introduces the KB2Bench framework.