Team XSZ at BioLaySumm2025: Section-Wise Summarization, Retrieval-Augmented LLM, and Reinforcement Learning Fine-Tuning for Lay Summaries

We present a multi-stage pipeline for BioLay-Summ 2025 Subtask 1.1 that improves readability, relevance, and factuality.First, we select the top-5 relevant sections and generate summaries with BioBART.Next, we retrieve a Kshot demonstration using BGE embeddings to prompt Llama 3 8B and fine-tune it with LoRA.We then merge section summaries via a second BioBART pass.Finally, we apply reinforcement learning (PPO and GRPO) with a composite reward combining factuality (AlignScore, SummaC), relevance (ROUGE-L, BERTScore), and readability (LENS, FKGL, DCRS, CLI).On PLOS and eLife validation sets, our pipeline reduces DCRS from 9.23 to 8.56 and CLI from 12.98 to 12.65, and boosts AlignScore from 0.722 to 0.862, demonstrating balanced gains in lay-summary quality.