Study: Large Language Models Recall More Than Initially Apparent
New research reveals advanced language models encode most facts but struggle with recall. Improving models' reasoning abilities can enhance reliability, potentially reducing reliance on larger architectures.
Researchers at Google Research and Technion have found that large language models (LLMs), including frontier models like GPT-5 and Gemini-3, encode 95-98% of tested facts. This knowledge is parametrically stored but often fails to surface during generation.
Traditionally, LLM errors such as hallucinations were attributed to missing knowledge, leading to efforts to increase model size or training data. However, the new study suggests the primary bottleneck is information recall, not its absence.
The research introduces "knowledge profiling," a method to evaluate a model's knowledge of individual facts across various conditions. This distinguishes between information being "encoded" (stored in parameters) and "known" (reliably retrievable). Models can recover up to 40-65% of facts they initially fail to recall directly, when given additional computation time to "think" through the process.
This recall through "thinking," termed "recall facilitation," may reduce the need for ever-larger models or external databases. The study indicates that simply scaling up model size does not automatically solve the access problem and can even exacerbate the challenge of retrieving information from a vast dataset.