LLMs Guess Instead of Admitting Ignorance

🔥 What happened Researchers at RPI and SUNY found that when RAG systems retrieve only coarse passages, LLMs answer anyway instead of abstaining. On PrimeKG biomedical questions, family-only passages cut abstention by 13.8 points – and roughly half those guesses are wrong. 💡 Why it matters Swapping an exact disease name for its broader family name drops abstention by 17.4 points versus deleting the sentence entirely. Confidence filtering, bigger models, and explicit abstain instructions all fail to fix it – and the pattern replicates on GloBI species data. ⚡ Our take This isn't a prompt bug, it's a structural flaw: LLMs confuse plausibility with knowledge. Building RAG without ontology grounding means shipping confident wrong answers at scale.