🧠How Reasoning Helps LLMs Retrieve Knowledge
Google researchers have discovered that Chain-of-Thought (CoT) helps models better retrieve facts from their own memory through a "computational buffer" and "factual priming."
🌍 This changes the approach to LLM optimization: the focus is shifting toward "trajectory selection" methods and training with process rewards.
👤 This explains why models like Gemini or Qwen become "smarter" in reasoning mode—they are essentially "warming up" their memory before providing an answer.
Source 1: https://research.google/blog/thinking-to-recall-how-reasoning-unlocks-parametric-knowledge-in-llms/