Thinking to recall: How reasoning unlocks parametric knowledge in LLMs
Research paper
Google Research presents a study on how reasoning helps LLMs recall simple facts even when step-by-step reasoning is unnecessary, identifying two mechanisms: a computational buffer effect and factual priming.
The research focuses on Gemini-2.5 (Flash and Pro) and Qwen3-32B models using closed-book QA datasets and will be presented at COLM 2026.