Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer
When large language models (LLMs) hallucinate, developers typically assume the model lacks the required facts. Engineering teams diagnose the error as missing knowledge. The standard response is to increase model size, expand training data, or build complex retrieval architectures. A new study by researchers at Google Research and Technion demonstrates that the knowledge is often not missing. The…
Frontier models like GPT-5 and Gemini-3 encode 95-98% of tested facts, even when they fail to directly recall them. This indicates that the primary bottleneck for factual accuracy is often recall, not encoding. Researchers from Google Research and Technion propose a new framework called fact-level profiling to distinguish between facts that are encoded and those that are actually known by the model.
Encoding failures require pre-training interventions like scaling model size or data coverage, while recall failures suggest post-training interventions, such as using inference-time computation to unlock stored knowledge. By training models to use additional computational effort to retrieve facts, developers can build more reliable applications without relying on larger models or external databases.
Written by urgent.news from VentureBeat's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
