Beyond Reachability: A Relation-Semantic Reasoning Framework for Biomedical Multi-hop QA

Ziyang Ling; Ziyi Chen; Ronald X. Xu; Mingzhai Sun

cs.CLOct 14, 2026

Beyond Reachability: A Relation-Semantic Reasoning Framework for Biomedical Multi-hop QA

Authors: Ziyang Ling, Ziyi Chen, Ronald X. Xu, Mingzhai Sun

Organizations: University of Science and Technology of China, Hefei, China

Abstract

Biomedical multi-hop reasoning over knowledge graphs requires identifying a semantically valid relation sequence that supports the answer. Existing methods can retrieve correct answers, yet often rank them poorly within a large pool of seemingly relevant candidates. Even when a correct answer is retrieved, its associated evidence and reasoning path may fail to support the relation composition required by the question. We term this disconnect between answer reachability and evidential validity the reachable-but-invalid problem. We present ReSemReason, which shares a question-induced distribution over typed, directed relation schemas between candidate discovery and evidence selection. Schema compatibility guides graph search; ReSemAttn then uses relation-conditioned key/value projections and an uncertainty-weighted path prior to attend to candidate-specific evidence. A separate absolute-support score complements normalized attention when all retrieved paths are weak. The reported experiments on BioHopR, PrimeKGQA, and MedReason show improvements in answer ranking and answer-set quality. Path-discrimination analyses examine the distinction between endpoint correctness and evidential support. We also specify matched controls for search, attention, and schema uncertainty.

Figures & tables

Figure 1. Two paths reach correct answer chronic myeloid leukemia (CML), but only the ABL1 path supports the question. Bottom: reported BioHopR Coverage, H@1, and same-answer PairAcc. A single-column motivation figure contrasts a target-protein path through ABL1 with a contraindication path through cardiovascular disorder; both reach CML. The baseline favors invalid support, whereas ReSemReason favors the ABL1 path. Three connected blocks show soft schemas, guided search, and ReSemAttn with absolute support. Bar charts compare Coverage 86.1 versus 94.8 percent, H at 1 of 64.1 versus 77.2 percent, and same-answer PairAcc of 49.3 versus 80.6 percent.

Figure 2. Overview of ReSemReason pipeline. Left: the two-layer ReSemAttn computation. Middle: input, schema-guided discovery, attention updates, and the CML answer with its ABL1 support. Right: shared schemas, relation-conditioned attention, and absolute support. A detailed ReSemAttn block on the left shows relation-conditioned key and value projections, token queries, a shared schema bias, softmax over paths, value aggregation, and gated residual fusion. Three inclined case planes in the center progress from the imatinib question and candidate paths to the first and second attention layers, ending with chronic myeloid leukemia and the ABL1 path. Right-side explanations connect each stage to the method's innovations. The matrices show symbolic token-path weights and do not encode measured attention magnitudes.

Panel of Figure 3 — Figure 3. Per-metric ablation results on BioHopR. V1–V11 denote the eleven ablated variants, and Full denotes the complete ReSemReason model. V1 = w/o Grounding Distribution V2 = w/o Relation-Guided Search V3 = w/ Hard Question Schema V4 = w/o Schema Entropy Gate V5 = w/o Entity Anchoring V6 = w/o Relation Compatibility V7 = w/o Path Consistency V8 = w/o Topology Penalty V9 = w/o Attention Prior V10 = w/o Relation-Conditioned KV V11 = w/o Path Aggregation. Higher is better for all metrics.

Integration	H@1	PairAcc	Joint@1
Labels in text	0.696	0.625	0.545
Final-score only	0.718	0.696	0.579
Prior + shared KV	0.744	0.763	0.634
Prior + relation KV	0.766	0.819	0.681

Table 3. Consolidated mechanism controls. Cond. H@1 conditions on a retrieved gold answer; PairAcc compares same-answer valid/invalid paths. FSR is acceptance of an all-invalid candidate pool; the one-path condition also uses invalid evidence.

Figure 4. Answer availability and evidence quality. Rows 1–12 identify methods; CE denotes Cross-Encoder. Higher is better except for WPR. A dash denotes an unreported path metric.

Figure 5. Candidate availability and evidence quality. Numeric method IDs follow Figure 4 ; methods 1 (Dense RAG) and 5 (GreaseLM) do not return paths and are omitted on the right.

Multi-Hop QA

Multi-hop question answering studies how systems combine evidence across documents, facts, or modalities to answer questions requiring linked reasoning.

See all 229 papers

Weaving Multi-Source Evidence for Biomedical Reasoning: The BioMedHop Benchmark and BioWeave Framework
Jun 15, 2026Xingyu Tan, Shiyuan Liu, Xiaoyang Wang +5Multi-Hop QA Evidence-Grounded Reasoning
University of New South Wales, Australia · 2CSIRO, Australia · University of Technology Sydney, Australia
MedHopQA: A Disease-Centered Multi-Hop Reasoning Benchmark and Evaluation Framework for LLM-Based Biomedical Question Answering
May 12, 2026Rezarta Islamaj, Robert Leaman, Joey Chan +13Multi-Hop QA LLM Evaluation
National Library of Medicine, Division of Intramural Research, Bethesda, MD, US · University of Illinois at Urbana-Champaign, Department of Computer Science, Urbana, IL, US · University of Michigan Medical School, Ann Arbor, Michigan, US
AdaPath: Query-Adaptive Path-Finding via Path-Bank for Multi-Hop Implicit Biomedical KGQA
Aug 31, 2026Jun Hyeong Kim, Dongki Kim, Yinhua Piao +1Multi-Hop QA Biomedical QA
KAIST

Jun 15, 2026J/K move · Enter open · S save

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

Xingyu Tan, Shiyuan Liu, Xiaoyang Wang, Qing Liu+4

University of New South Wales, Australia · 2CSIRO, Australia · University of Technology Sydney, Australia

Multi-Hop QA Evidence-Grounded Reasoning Graph Retrieval-Augmented Generation Biomedical QA Knowledge Graph Question Answering Biomedical KGs

Open paper AskPDF arXiv

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. BioMedHop contains 10,045 instances across KG, document, web, and hybrid evidence settings, covering shared-neighbor matching, intersection reasoning, path-based reasoning, and counting, with option-based, open-ended, and numeric count renderings. To support this benchmark, we further 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. Comprehensive experiments show that BioWeave achieves the best overall performance among compared methods on BioMedHop, outperforming the strong hybrid baseline ToG-2 by 10.5% in the overall average. Moreover, BioWeave consistently improves different LLM backbones and enables smaller models, such as Qwen3-4B, to achieve reasoning performance comparable to GPT-4-Turbo.

Abstract

Multi-Hop QA

Explore similar work

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

MedHopQA: A Disease-Centered Multi-Hop Reasoning Benchmark and Evaluation Framework for LLM-Based Biomedical Question Answering

AdaPath: Query-Adaptive Path-Finding via Path-Bank for Multi-Hop Implicit Biomedical KGQA