Thinking to recall: How reasoning unlocks parametric knowledge in LLMs

It is well-established that allowing large language models (LLMs) to generate step-by-step reasoning traces, commonly known as chain-of-thought (CoT), enhances performance on complex tasks.