A new study from Google Research examines a puzzling gap in large language models: they sometimes cannot recall a fact when asked directly, yet can produce it when prompted to reason step by step. The researchers describe this as reasoning unlocking parametric knowledge—information stored in the model's weights during training.
In their experiments, they found that chain-of-thought prompting improves recall on a range of knowledge-intensive tasks, even when the reasoning trace is not strictly necessary for the answer. The benefit seems to come from the intermediate steps reorienting the model's internal attention toward the relevant stored associations.
The work suggests that reasoning is not just a way to compute new conclusions, but also a mechanism for navigating the model's own memory. This has practical implications for prompt design and for understanding why some knowledge remains latent until the right line of thinking is elicited.