Why bigger language models remember facts better—and predictably so

Large language models hallucinate confidently, but their ability to recall actual facts follows a mathematical pattern. Researchers tested 38 models on over 8,900 scholarly references and found that factual accuracy scales predictably with model size and training-data frequency: a bigger model learning about uncommon topics performs like a smaller model learning about common ones. The relationship follows a sigmoid curve in the combination of these two factors, suggesting recall works like a signal-to-noise problem—more frequent topics cut through the noise, larger models reduce the noise floor.