Hallucination Reduction

Give a language model one document, tell it to summarize using only the facts in that document, and it will still sometimes add a fact the document does not contain. Vectara's hallucination leaderboard measures exactly that, the most favourable case grounding will ever get, and scores each summary with its HHEM classifier. As of its 22 September 2026 update, the share of summaries with unsupported content ranged from 1.8% to 24.2% across 108 models, with a median of 9.55%. The source was in the context every time.