Condition-Gated Reasoning for Context-Dependent Biomedical Question Answering

Jash Rajesh Parekh; Wonbin Kweon; Joey Chan; Rezarta Islamaj; Robert Leaman; Pengcheng Jiang; Chih-Hsuan Wei; Zhizheng Wang

cs.CLFeb 20, 2026

Condition-Gated Reasoning for Context-Dependent Biomedical Question Answering

Authors: Jash Rajesh Parekh, Wonbin Kweon, Joey Chan, Rezarta Islamaj, Robert Leaman, Pengcheng Jiang, Chih-Hsuan Wei, Zhizheng Wang,

Organizations: University of Illinois Urbana-Champaign · National Institutes of Health

Abstract

Current biomedical question answering (QA) systems often assume that medical knowledge applies uniformly, yet real-world clinical reasoning is inherently conditional: nearly every decision depends on patient-specific factors such as comorbidities and contraindications. Existing benchmarks do not evaluate such conditional reasoning, and retrieval-augmented or graph-based methods lack explicit mechanisms to ensure that retrieved knowledge is applicable to given context. To address this gap, we propose CondMedQA, the first benchmark for conditional biomedical QA, consisting of multi-hop questions whose answers vary with patient conditions. Furthermore, we propose Condition-Gated Reasoning (CGR), a novel framework that constructs condition-aware knowledge graphs and selectively activates or prunes reasoning paths based on query conditions. Our findings show that CGR more reliably selects condition-appropriate answers while matching or exceeding state-of-the-art performance on biomedical QA benchmarks, highlighting the importance of explicitly modeling conditionality for robust medical reasoning.

Figures & tables

Figure 1. Example of condition-gated reasoning in the biomedical domain. Existing KG-RAG extracts triples and retrieves contraindicated treatments. CGR extracts n-tuples with patient-specific conditions, gating unsafe paths and retrieving only safe alternatives .

Reasoning Category	%	Example(s)
Comorbidity & Organ-Based Contraindications	57	Doc 1: Treatment for major depressive disorder includes antidepressants such as bupropion , sertraline , and fluoxetine . Bupropion acts as a norepinephrine-dopamine reuptake inhibitor.
		Doc 2: Bupropion lowers the seizure threshold and is contraindicated in patients with bulimia nervosa or other eating disorders due to elevated seizure risk.
		Q: Which antidepressant for MDD is contraindicated in patients with bulimia ?
		A: Bupropion (w/o condition: Sertraline )
Diagnostic Modality Selection	23	Doc 1: For suspected appendicitis , diagnostic imaging options include CT scan , ultrasound , and MRI . CT scan has the highest sensitivity in adult populations.
		Doc 2: In pediatric patients , ultrasound is the preferred first-line imaging modality to avoid ionizing radiation exposure , with CT reserved for inconclusive cases.

Table 1. Types of conditional multi-hop reasoning required in CondMedQA. We show orange bold italics for bridge entities, blue italics for supporting facts, underline for the conditions, green bold for the conditional answer, and magenta for the general (non-conditional) answer. Each example requires synthesizing information across both documents, neither document alone contains sufficient information to answer the conditional question.

Dimension	Gwet’s AC	% Agree (all)	% Agree (pair)
Conditionality	0.96 (AC1)	96.7	97.8
Answer Accuracy	0.67 (AC1)	66.7	77.8
Question Quality	0.60 (AC2)	56.7	71.1

Table 2. Inter-annotator agreement across 30 items.

Figure 2. Overview of Condition-Gated Reasoning (CGR) compared to traditional KG-RAG. Given a clinical query about a patient with hypertension and bilateral renal artery stenosis (BRAS), traditional KG-RAG extracts standard relation triples and traverses all paths indiscriminately, retrieving evidence for contraindicated treatments (e.g., Lisinopril, an ACE inhibitor). CGR extends triples to n-tuples that include patient-specific conditions as gating constraints (e.g., $\neg$ BRAS, $\neg$ pregnancy), masking contraindicated paths during graph traversal and assembling only condition-appropriate evidence, yielding the correct answer (Amlodipine).

GPT-5.2
Method	CondMedQA		MedHopQA		MedHopQA (Cond)		BioASQ-B
Method	EM	F1	EM	F1	EM	F1	EM	F1
Zero-Shot	22.00	41.29	20.50	47.87	28.27	53.52	13.00	40.21
RAG	44.00	52.36	51.25	62.61	45.71	55.04	31.80	54.80
CoT	40.00	52.81	49.50	57.55	50.00	59.38	25.60	47.81
HippoRAG2	46.00	67.56	45.00	60.81	48.57	72.82	22.60	43.57

Table 3. Performance comparison across biomedical multi-hop QA datasets and our custom benchmark. Bold indicates the best performance within each foundation model group.

Method	CondMedQA		MedHopQA		MedHop (C)
Method	EM	F1	EM	F1	EM	F1
CGR w/o gating	57.0	60.2	72.5	76.8	62.9	73.5
CGR	82.0	86.7	86.8	89.9	80.0	85.4

Table 4. Ablation study on the effect of condition gating.

Panel of Figure 3 — Figure 3. Hyperparameter sensitivity analysis. Top row: varying $k_{\text{paths}}\in\{1,3,5,10\}$ . Bottom row: varying $k_{\text{nodes}}\in\{1,3,5,10\}$ . Based on these results, we set $k_{\text{paths}}=3$ and $k_{\text{nodes}}=5$ for all experiments.

Extractor	CondMedQA		MedHopQA		MedHop (C)		Tok/s	s/doc
Extractor	EM	F1	EM	F1	EM	F1
Flan-T5-Large	65.00	66.67	85.00	90.00	70.00	77.33	1053.95	21.69
LLaMA-3.2-3B	65.00	66.67	85.00	90.00	85.00	88.33	932.12	15.04
Qwen2.5-14B	65.00	66.67	85.00	92.50	85.00	88.33	1077.55	15.08

Table 5. Ablation study on n-tuple extraction model, showing that higher quality n-tuple extraction leads to better CGR performance. Tok/s = tokens per second, s/doc = seconds per document (throughput from vLLM)

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Document 1: HIV Treatment
“A large study in Africa and India found that a PI-based regimen was superior to an NNRTI-based regimen in children less than 3 years who had never been exposed to NNRTIs in the past. Thus the WHO recommends PI-based regimens for children less than 3 . In pregnant women , viral load is proportional to transmission risk; ART reduces transmission to mothers and infants before, during, and after delivery .”
Extracted N-Tuples:
$\langle$ PI-based regimen, superior_to , NNRTI-based regimen, [ children <3 years , not exposed to NNRTI ] $\rangle$
$\langle$ WHO, recommends , PI-based regimen, [ children <3 years ] $\rangle$
$\langle$ viral load, proportional_to , transmission risk, [ pregnant women ] $\rangle$
$\langle$ ART, reduces , transmission risk, [ pregnant women , before/during/after delivery ] $\rangle$

Appendix

Table 6. Examples of condition-aware n-tuple extraction. Green text indicates patient-specific conditions extracted from source text and captured in the n-tuple conditions field.

Example 1: “Which gene causes cardiomyopathy in pediatric patients but not in adults?”
K (keywords)	[gene, cardiomyopathy, causes]
C (required)	[pediatric, in children]
C (excluded)	[in adults, adult patients]
N (negated)	[]
Interpretation: K matches nodes related to genes and cardiomyopathy. C gates traversal: only edges with pediatric/child conditions are followed; edges conditioned on adults are blocked.
Example 2: “What drug treats hypertension in pregnant women, excluding ACE inhibitors?”

Appendix

Table 7. Query parsing examples showing decomposition into keywords ( K ), conditions ( C ), and negated entities ( N ).

Example 1: “Which gene causes cardiomyopathy in pediatric patients but not in adults?”
Conditions	[pediatric, in children, in adults, adult patients]
Output	{‘‘pediatric’’: true, ‘‘in children’’: true, ‘‘in adults’’: false, ‘‘adult patients’’: false}
Gating result: Edges conditioned on “pediatric” or “in children” are traversable ( $\mathcal{G}=1$ ); edges conditioned on “in adults” or “adult patients” are blocked ( $\mathcal{G}=0$ ).
Example 2: “What antibiotic is safe for a 5-year-old boy with pneumonia and no known allergies?”
Conditions	[in children, in adults, penicillin allergy, renal impairment, pregnancy]
Output	{‘‘in children’’: true, ‘‘in adults’’: false, ‘‘penicillin allergy’’: false, ‘‘renal impairment’’: null, ‘‘pregnancy’’: false}

Appendix

Table 8. Condition evaluation examples showing how a single LLM call populates the lookup table

\mathcal{L}

for edge gating.

Error Type	QID	Question	Gold	Predicted	Diagnosis
Retrieval (11/18)	23	What antibiotic is recommended for treating scrub typhus during pregnancy ?	Azithromycin	Doxycycline	Gold tuple with pregnancy condition exists in KG but not in top-ranked paths; typhus–doxycycline edge dominates ranking.
	74	Which multikinase inhibitor is most effective for hepatocellular carcinoma in advanced (BCLC-C) unresectable disease ?	Sorafenib	Insufficient evidence	Gold tuple with BCLC-C condition in KG; top paths traverse unrelated edges. Gold never reaches evidence assembly.
Extraction (5/18)	65	Which csDMARD that has the side effect of liver cancer is the most effective for ankylosing spondylitis in peripheral arthritis ?	Methotrexate	Insufficient evidence	Methotrexate not extracted from source documents; KG lacks the gold entity entirely.
	96	Which anxiolytic is preferred for generalized anxiety disorder in a patient who has myasthenia gravis ?	Buspirone	Sertraline	Buspirone absent from KG. Model correctly avoids benzodiazepines and selects SSRI; normalization maps SSRI to an incorrect canonical form.
Reasoning (2/18)	28	What imaging test is preferred for suspected pulmonary embolism in a pregnant patient due to lower radiation exposure?	V/Q scan	MRI w/o contrast	V/Q is the guideline alternative to CTPA in pregnancy; model over-optimizes for radiation avoidance and selects MRI.
	99	Which antimycobacterial drug should be replaced in the RIPE regimen for a patient who is HIV-positive and receiving protease inhibitors ?	Rifabutin	Rifampin	Gold present in retrieved paths. Model answers the drug to be replaced (rifampin) rather than the replacement (rifabutin).

Appendix

Table 9. Error analysis of 18 incorrect CGR predictions on CondMedQA. Patient-specific conditions are highlighted in color . Two representative examples are shown per error type.

Biomedical QA

Biomedical question answering develops systems that answer health and life-science questions using clinical knowledge, research literature, and structured data.

See all 51 papers

Hybrid-IR: Dual-Path Hybrid Retrieval with Iterative Reasoning for Complex Medical Question Answering
Jun 24, 2026Sheng Wan, Jiahui Zhang, Zicheng Zhao +1Retrieval-Augmented Generation Iterative RAG
College of Artificial Intelligence, Nanjing Agricultural University, Nanjing, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China
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
MedJudgeRAG: Option-Wise Evidence Judgment with Dynamic Knowledge Graphs for Medical MCQA
Jul 24, 2026Seongwon Seo, Seung Hwan Cho, Young-Min KimGraph Retrieval-Augmented Generation Medical QA
Department of Industrial Data Engineering, Hanyang University, Seoul, South Korea · School of Interdisciplinary Industrial Studies, Hanyang University, Seoul, South Korea

Jun 24, 2026J/K move · Enter open · S save

Hybrid-IR: Dual-Path Hybrid Retrieval with Iterative Reasoning for Complex Medical Question Answering

Sheng Wan, Jiahui Zhang, Zicheng Zhao, Shougang Ren

College of Artificial Intelligence, Nanjing Agricultural University, Nanjing, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China

Retrieval-Augmented Generation Iterative RAG Graph Retrieval-Augmented Generation Biomedical QA Hybrid Retrieval

Open paper AskPDF arXiv

Large language models (LLMs) have shown promising performance across a wide range of biomedical applications, including medical question answering (QA), yet they remain prone to hallucinations and outdated knowledge. Although retrieval-augmented generation (RAG) can alleviate this issue by incorporating external documents, there still exist two fundamental limitations. First, medical knowledge is often fragmented across documents, while most RAG methods rely on a single retrieval path, which makes it challenging to jointly preserve fine-grained semantic information and structured global associations. Second, static retrieval strategies are typically insufficient to support deep reasoning that is important in complex medical QA. In this paper, we present a dual-path retrieval framework with an iterative retrieval-reasoning mechanism termed "Hybrid-IR" for complex medical QA. The proposed Hybrid-IR integrates graph-based retrieval for exploration of structured knowledge and dense retrieval for fine-grained semantic matching. Moreover, the reasoning trajectory can be progressively refined through an iterative retrieve-reason loop. Experiments on three widely used medical QA benchmarks demonstrate the effectiveness of our Hybrid-IR.

Abstract

Biomedical QA

Explore similar work

Hybrid-IR: Dual-Path Hybrid Retrieval with Iterative Reasoning for Complex Medical Question Answering

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

MedJudgeRAG: Option-Wise Evidence Judgment with Dynamic Knowledge Graphs for Medical MCQA