Reliable Health Question Answering: From Quality-Aware Retrieval to Reasoning under Knowledge Conflicts

Problem. Artificial intelligence (AI) has advanced rapidly and changed how people access health-related information, yet the capability of AI systems has outpaced their reliability. The propagation of misinformation has evolved from misleading webpages, to large language models (LLMs) generating incorrect health-related answers, and further to agentic systems taking consequential clinical actions grounded in unreliable reasoning. Existing safeguards mitigate these limitations through credibility labeling of health sources, manual and automated fact-checking, and benchmark-based model evaluation. However, search engines continue to rank results by relevance over reliability, fact-checking relies mostly on common knowledge rather than scientific evidence, and models are evaluated on one-sided context instead of the conflicting evidence they could encounter in practice. The following three contributions address these limitations in turn. Approach and results. (1) To rank health information by reliability, we model its quality in a three-dimensional information retrieval (IR) system by asking three questions: is it useful for the query, does it support or refute the claim, and does it come from a credible source? We use reciprocal rank fusion to combine these dimensions into a ranking that outperforms the BM25 baseline by 7 points in ranking helpful documents over harmful documents, among the two best automatic systems on the TREC Health Misinformation 2021 benchmark [1]. (2) To ground claim verification in the scientific evidence, we introduce CliniFact, a scientific claim dataset built from the hypothesis-testing results of clinical research, covering 992 clinical trials across 20 disease categories. From each trial's study arms, primary outcome measures, and results, we derive and label 1,970 claims, each linked to its source trial publication. Evaluating a range of models on CliniFact, we find that small discriminative models such as BioBERT (74.7% macro-F1) outperform large generative models such as OpenBioLLM-70B (39.6% macro-F1). (3) To evaluate models on the conflicting evidence they encounter in practice, we introduce HealthContradict, 920 expert-verified health questions, each paired with conflicting evidence and a factual answer supported by scientific evidence. Biomedically adapted models are better at exploiting correct context while resisting incorrect context. Without context, the biomedical model and its general-purpose counterpart differ by only 0.7 macro-F1 points. When a document is provided, Meditron3-8B exceeds its general-purpose counterpart by 18.9 points under correct context, 16.9 under incorrect context, and 13.3 to 16.8 under conflicting context. But every model performs worse under incorrect or conflicting context. Context. Prior work treats these problems separately: adhoc retrieval optimizes topical relevance [1], scientific claim verification datasets pair claims with one-sided evidence [2], and knowledge-conflict benchmarks perturb context synthetically rather than drawing on realistic clinical passages [3, 4]. Together, these three contributions are steps toward trustworthy answers to health questions. A multidimensional ranking method places helpful documents above harmful ones with more reliability than relevance alone. The dataset CliniFact shows that a small fine-tuned model can verify clinical claims more accurately than generative models many times its size. The benchmark HealthContradict indicates that biomedical domain adaptation helps, but that human oversight remains essential wherever health misinformation can cause harm.