(논문 요약) Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs (Paper)

핵심 내용

• Reasoning Unlocks Latent Knowledge (reasoning on 성능 > reasoning off 성능)
• Less Capable Models Benefit More from Reasoning.
• Question Complexity is a Poor Predictor of Reasoning Effectiveness.
• How Reasoning Improves Parametric Recall?
  • Reasoning Tokens as a Computational Buffer: generating extra tokens during reasoning allows models to perform additional latent operations and to bypass the depth limits of a single forward pass on the input.
  • Factual Priming
• Hallucinating intermediate facts during reasoning increases the likelihood of hallucinations in the final answer.
• Directly improve model accuracy by prioritizing reasoning trajectories that contain hallucination-free factual statements.