Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs
Google, Wednesday, June 24th, 2026
Google Research shows reasoning traces help LLMs recall facts via latent computation and answer priming.
Google Research published findings showing that reasoning traces help language models recall simple facts through two mechanisms: using generated reasoning tokens for latent computation, and generating related facts to prime answer retrieval.
Studying models like Gemini-2.5, the authors found even meaningless reasoning tokens provide computational benefit, while factually accurate intermediate steps significantly improve final answer accuracy.
The work suggests training improvements through process rewards that encourage factually supported reasoning. The results illuminate how chain-of-thought reasoning unlocks knowledge already stored in model parameters.