Thinking to recall: How reasoning unlocks parametric knowledge in LLMs

Google Research explores how reasoning processes unlock and improve the retrieval of parametric knowledge within large language models.