Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer

When large language models (LLMs) hallucinate, developers typically assume the model lacks the required facts. Engineering teams diagnose the error as missing knowledge.