Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs

<p>This research explores how <strong>reasoning</strong> helps Large Language Models (LLMs) answer simple, <strong>single-hop factual questions</strong> that do not logically require step-by-step thinking. The authors demonstrate that enabling reasoning expands the model’s <strong>parametric knowledge boundary</strong>, allowing it to "unlock" correct answers that are otherwise unreachable. This improvement is driven by two primary mechanisms: a <strong>computational buffer</strong> effect where extra tokens allow for more latent processing, and <strong>factual priming</strong> where the model retrieves related facts to bridge toward the correct answer. However, the study warns that <strong>hallucinating</strong> facts during the reasoning phase significantly increases the risk of providing a false final answer. Ultimately, the paper suggests that accuracy can be improved by <strong>prioritizing reasoning paths</strong> that contain verified factual statements.</p>