Enough context for a passage to answer the question

Google Research (ICLR 2025) separates relevant context from sufficient context: context that contains all the information needed for a definitive answer. A relevant passage can be insufficient. Their analysis found that large models (Gemini, GPT, Claude) answer well with sufficient context but fail to avoid wrong answers when it is insufficient, smaller open models hallucinate or abstain even when it is sufficient, and adding context reduces abstention: Gemma gave incorrect answers on 10.2% of questions with no context but 66.1% with insufficient context. The tested models predate 2026; the distinction, not the numbers, is what carries.