DeepMindLLMreasoningAI research
DeepMind shows reasoning improves factual recall in LLMs
personLab Desenvolvimento
schedule5 minA new study from Google DeepMind called Thinking to Recall brought an interesting discovery: reasoning chains don’t just help models think better — they unlock factual knowledge the model already has but can’t access without reasoning.
The research reveals two key mechanisms:
- Computational buffer: reasoning creates extra space for the model to ‘process’ before answering, like giving it time to query its own memory.
- Factual priming: while reasoning, the model activates relevant neurons that stay ‘dormant’ in direct answers.
This changes how we understand parametric knowledge retrieval in LLMs. In practice, it means models with reasoning aren’t just better at logic — they’re better at remembering facts they already know.