Abstract:
Biomedical summarization aims to transform complex scientific literature into concise, accessible summaries tailored to diverse audiences. This paper introduces a novel approach to controlling the ”layness” of biomedical summaries by varying technical depth and complexity for three audience categories: pre-med students, researchers, and domain experts. Using large language models (LLMs) like GPT-4o and Gemini, we evaluate one-shot and few-shot prompting strategies, developing metrics S1 and S2 to quantify the balance between accessibility and technical precision. The methodology highlights the challenges of reproducibility, dependency on prompts, and the limitations of automated evaluation for nuanced aspects like readability and factuality. The findings reveal that one- shot prompting consistently outperforms alternative strategies in generating tailored summaries. However, scalability and subjective human evaluations remain significant barriers.
