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

Summary

Researchers at Google Research and Technion found that hallucination in large language models (LLMs) is often due to retrieval failure, not missing knowledge. Experiments show that frontier models like GPT-5 and Gemini-3 encode 95-98% of tested facts but sometimes fail to surface them during generation. By using inference-time computation, models can recover up to 65% of facts they can't directly recall, indicating that inference-time compute is key to improving factual accuracy.