Frontier LLMs are not forgetting facts. They are failing to retrieve them. A new study from Google Research and Technion, tested across 13 models and 4 million responses, found that GPT-5 and Gemini-3 encode 95-98% of tested facts parametrically. Yet those same models fail to directly recall 26-34% of what they have stored. The bottleneck is access, not storage.
Inference-time computation recovers 40-65% of those inaccessible facts. The researchers call this recall facilitation: Chain-of-Thought reasoning structurally primes the model to surface answers it cannot produce on direct query. This mirrors the human tip-of-the-tongue phenomenon. Scaling model size does not fix it. Gemma3 scaled from 1B to 27B parameters cut encoding failures from 85% to 23%, but recall failures peaked at 40% without thinking, meaning bigger models shift the problem rather than solve it. Fine-tuning larger models to patch recall failures is an expensive architectural misstep.
The paper introduces five distinct knowledge profiles, from direct recall to inference without encoding, and a benchmark called WikiProfile covering 2,150 Wikipedia facts. That taxonomy is the part worth reading in full. It reframes rare facts, reverse-direction queries, and long-tail recall gaps not as missing knowledge but as retrieval failures, each with different engineering remedies. The implication for production pipelines: deploying RAG to fix hallucinations may be solving the wrong problem.
[READ ORIGINAL →]