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

GPT-5 and Gemini-3 encode 95-98% of tested facts, but a Google Research and Technion study finds they fail to recall a third ...