Year: 2025
Type: conference-paper
Abstract: The growing need to make biomedical research accessible to non-expert audiences has motivated the development of effective lay summarization systems.While large language models (LLMs) have set recent benchmarks, their computational demands limit widespread adoption.This paper explores the use of smallscale, state-of-the-art LLMs (4B-7B parameters) for biomedical lay summarization in the BioLaySumm 2025 shared task.Leveraging dynamic 4-bit quantization, extractive preprocessing, prompt engineering, data augmentation, and Direct Preference Optimization, our system, based on Gemma3 4B, Qwen3 4B, and GPT-4.1-mini,ranked second in its category, showing that compact models can deliver highquality, factually accurate summaries.
Authors: Ilyass Ramzi, Isabel Bedmar
Institution: Universidad Carlos III de Madrid
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4412889116