Google Research has published a study on how reasoning affects a model’s ability to recall parametric knowledge — information stored inside the model rather than retrieved from an external source. The work examines how structured thinking can unlock answers that a model may not produce directly.
The finding matters for developers deciding when to use reasoning, retrieval, or both. If reasoning can improve recall, some tasks may benefit from deliberate intermediate steps even when no external search is available.
It also adds nuance to evaluation: a model’s knowledge is not just what it answers on the first pass, but what it can recover when prompted to reason more carefully.