Multi-source biomedical reasoning

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

Xingyu Tan1,2 · Shiyuan Liu3 · Xiaoyang Wang1 · Qing Liu2 · Xiwei Xu2 · Xin Yuan2 · Liming Zhu2 · Wenjie Zhang1

1University of New South Wales    2CSIRO    3University of Technology Sydney

Abstract

How do we reason over biomedical questions when the evidence is scattered across knowledge graphs, literature, and the web?

Biomedical question answering (QA) increasingly requires reasoning over interacting entities, where supporting evidence is scattered across biomedical knowledge graphs, literature documents, and web-accessible resources. However, existing biomedical QA benchmarks mainly focus on exam-style knowledge, literature comprehension, or short-range multi-hop inference, leaving source-conditioned graph reasoning and evidence topology construction underexplored.

To fill this gap, we introduce BioMedHop, a multi-source graph-grounded benchmark for evaluating biomedical reasoning over structured evidence topologies, and propose BioWeave, a source-aware reasoning framework that retrieves biomedical KG paths, gathers supporting clues from documents and web sources, assembles them into a unified evidence graph, and verifies answers through entity-level evidence support.

BioWeave Framework Overview

Figure 1: Overview of the BioWeave framework assembling multi-source evidence into a unified biomedical reasoning graph.

BioMedHop Benchmark

📊 Scale & Coverage

BioMedHop contains 10,045 instances spanning four evidence settings: KG-only, document-only, web-accessible, and hybrid. Questions cover shared-neighbor matching, intersection reasoning, path-based reasoning, and counting tasks.

🎯 Rendering Formats

Instances are presented in three formats: option-based (multiple choice), open-ended (free-form), and numeric count (quantitative), enabling evaluation across diverse answer types and reasoning challenges.

🔗 Multi-Source Evidence

Evidence is drawn from biomedical KGs (Monarch, SPOKE), PubMed literature, clinical notes, and web resources. A single question may require connecting genes, diseases, drugs, phenotypes, and clinical outcomes.

🧬 Graph-Grounded Reasoning

Unlike prior benchmarks focused on single-source retrieval, BioMedHop specifically targets source-conditioned graph reasoning and evidence topology construction — the two underexplored dimensions in biomedical QA evaluation.

BioWeave Framework

🔍

KG Path Retrieval

Retrieves typed multi-hop paths from biomedical KGs (Monarch, SPOKE), grounding reasoning in validated entity-relation structures with normalized identifiers via UMLS and HPO.

📚

Multi-Source Evidence Gathering

Gathers supporting clues from PubMed abstracts, clinical documents, and web sources, resolving entity synonyms and abbreviations for consistent cross-source entity grounding.

🕸️

Unified Evidence Graph

Assembles retrieved paths and clues into a unified evidence graph, then verifies answers through entity-level evidence support — making every reasoning step explicit and traceable.

Key Results

+10.5%

Improvement in overall average over the strong hybrid baseline ToG-2 on BioMedHop.

4B ≈ GPT-4

BioWeave enables Qwen3-4B to achieve reasoning performance comparable to GPT-4-Turbo.

10,045

BioMedHop instances spanning KG, document, web, and hybrid evidence settings.

BioWeave consistently improves different LLM backbones and achieves best overall performance among all compared methods on BioMedHop.

BibTeX

@article{tan2026bioweave,
  title={Weaving Multi-Source Evidence for Biomedical Reasoning: The BioMedHop Benchmark and BioWeave Framework},
  author={Tan, Xingyu and Liu, Shiyuan and Wang, Xiaoyang and Liu, Qing and Xu, Xiwei and Yuan, Xin and Zhu, Liming and Zhang, Wenjie},
  year={2026}
}