Weaving Multi-Source Evidence for Biomedical Reasoning: The BioMedHop Benchmark and BioWeave Framework

The article introduces the BioMedHop benchmark and the BioWeave framework, designed to enhance biomedical question answering by integrating multi-source evidence from knowledge graphs, documents, and web resources. BioMedHop comprises 10,045 instances focusing on various reasoning tasks, while BioWeave demonstrates superior performance, outperforming the ToG-2 baseline by 10.5% and enabling smaller models like Qwen3-4B to match the reasoning capabilities of larger models such as GPT-4-Turbo. This advancement is significant for practitioners as it facilitates improved reasoning in biomedical applications, leveraging diverse evidence sources.