Recall, Not Knowledge Gaps, Causes Most Fact Errors in Cutting-Edge LLMs
Google’s new knowledge‑profiling framework shows that Gemini 3, GPT‑5 and similar frontier models store the vast majority of facts but fail to retrieve them, turning factual errors into ‘lost‑key’ problems. By classifying facts into encoding, recall and reasoning categories, the study points to post‑training retrieval tricks as the next lever for improving LLM reliability.