Frontier Models Can Recover Up to 65% of Facts by 'Thinking Longer,' Study Shows

Summary

A new study by Google Research and Technion reveals that large language models (LLMs) often possess facts but struggle with recall, leading to hallucinations. The research introduces "knowledge profiling" to distinguish between encoding and recall failures, demonstrating that frontier models can recover up to 65% of seemingly forgotten facts by employing inference-time computation like Chain-of-Thought.