Thinking to recall: How reasoning unlocks parametric knowledge in LLMs

Researchers investigated why large language models recall simple facts better when allowed to generate reasoning traces, even though these facts require no complex step-by-step problem solving. The study identified two mechanisms at work: models use generated reasoning tokens as a computational buffer for latent processing, and they generate related facts that prime the recall of correct answers. This matters because it reveals that reasoning helps language models access knowledge that would otherwise be unreachable, though the mechanism differs fundamentally from how reasoning aids complex problem solving. The findings also highlight a risk where models may generate false intermediate facts in their reasoning process.