arrow_backBack to Blog
DeepMindLLMreasoningAI research

DeepMind shows reasoning improves factual recall in LLMs

personLab Desenvolvimento
schedule5 min

You know when someone tells you “think before you answer”? Turns out it works the same way for LLMs — and now DeepMind has the science to prove it.

The Thinking to Recall study from Google DeepMind reveals something fascinating: reasoning chains can unlock correct answers that the model simply can’t reach when responding directly.

How does it work?

Think of it like this: the model has the knowledge stored, but it can’t access it properly without guided “thinking.” The study reveals two mechanisms:

  1. Computational buffer: reasoning creates extra space to “think” before answering
  2. Factual priming: the act of reasoning itself activates relevant knowledge

This explains why models that “think” (like OpenAI’s o1 or Claude with extended thinking) perform so much better on factual tasks.

Bottom line: asking a model to think before answering isn’t just a gimmick — it’s science.