Abstract
Retrieval-augmented generation (RAG) evaluation checks if model claims are factually grounded in retrieved documents, but it does not verify whether the retrieved evidence is attributed to the correct entity. A clinical RAG response can pass all automated checks (zero hallucinations, near-perfect faithfulness, real citations) while misrepresenting clinical evidence for drug Y as evidence for queried drug X. We term this deceptive grounding (DG): a failure invisible to faithfulness, hallucination, and citation checks because every claim is sourced from a real document but pertains to the wrong entity.
Using a controlled factorial benchmark across 13 models, we find DG rates ranging from 8% to 87% under peak adversarial conditions. Medical and biomedical fine-tuned models reach up to 86.7%; domain specialization exacerbates rather than mitigates the failure. A controlled ablation identifies the mechanism: removing entity-specific clinical evidence from retrieved documents entirely eliminates entity-attribution failure, shifting all failures to confabulation. Both failure modes respond to the same trigger but take different paths.
Production measurement across 740 drug-disease pairs reveals an overall DG of 7.8% in a deployed RAG system, increasing to 13.6% for recently approved drugs. Entity-attribution verification (checking that cited evidence applies to the queried entity) detects DG with 97.0% precision and 98.7% DG recall (IPW-adjusted human gold standard); no existing framework implements this.
Blogger's Review: This study reveals significant flaws in entity attribution within medical generation models, highlighting the necessity of ensuring accuracy in clinical applications. Introducing entity attribution verification can significantly enhance the reliability of generation systems and reduce potential risks from misinformation.