New research shows that enabling step-by-step reasoning traces allows large language models to accurately retrieve facts, even for simple queries. This improvement occurs due to a computational buffer effect where increased generation time enhances the model’s access to internal parametric knowledge.