Reasoning Enhances LLMs' Recall of Simple Facts

Google Research has discovered that reasoning traces can help language models recall simple facts more effectively. This phenomenon is driven by two mechanisms: a computational buffer effect and factual priming, where related facts are generated to aid recall. The study found that even meaningless reasoning traces can improve recall, though the actual content still matters. However, hallucinated facts in reasoning traces can negatively impact accuracy. These findings suggest that focusing on factually accurate reasoning can enhance model reliability and open new avenues for training improvements.