Large Language Models have demonstrated remarkable capabilities in natural language understanding and generation. However, their application in high-stakes domains such as medicine and science is significantly constrained by the tendency to generate f...
Large Language Models have demonstrated remarkable capabilities in natural language understanding and generation. However, their application in high-stakes domains such as medicine and science is significantly constrained by the tendency to generate factually incorrect information, a phenomenon known as hallucination. Retrieval-Augmented Generation (RAG) has emerged as a promising framework to mitigate this issue by grounding model responses in external evidence. Despite its potential, the efficacy of RAG relies heavily on the precision of the retrieved documents, which is often suboptimal when using general-purpose retrieval systems in specialized domains. This study proposes a reranker-driven RAG tailored for medical and scientific question answering. We conduct a rigorous comparative analysis between traditional discriminative cross-encoders and emerging generative rerankers. Our investigation reveals that deploying pre-trained retrieval models without adaptation frequently leads to retrieval-induced degradation, where noisy context deteriorates performance compared to direct inference. To address this, we implement supervised fine-tuning on a composite biomedical corpus. The experimental results across five medical and scientific question answering benchmarks demonstrate that the domain-adapted generative reranker significantly outperforms both the discriminative cross-encoder and direct inference baselines. Notably, the fine-tuned generative model achieves significant improvements on BioASQ and MedMCQA, and competitive results on PubMedQA and MMLU, surpassing recent advanced frameworks. Furthermore, an ablation study indicates that while generative architectures require more data to converge, they possess a higher asymptotic performance ceiling than discriminative models. These findings suggest that leveraging the reasoning capabilities of generative models for relevance ranking is a viable strategy for building trustworthy and accurate medical artificial intelligence systems.