Empty shelves or lost keys? Recall is the bottleneck for parametric factuality

Google Research · 12 August 2026, 09:51

Knowledge profiling found that many factual mistakes in frontier LLMs are recall failures (encoded facts not reliably accessible) rather than encoding failures (facts not encoded). 95–98% of facts were encoded in Gemini-3-Pro and GPT-5, yet direct recall failed on 26–34% of facts. The implication is a shift toward interventions that improve knowledge utilization (including thinking/recovery) instead of primarily scaling to add missing knowledge.

Why it matters

Generative AI