Thinking to recall: How reasoning unlocks parametric knowledge in LLMs (research.google)

Researchers Zorik Gekhman and Jonathan Herzig from Google Research explored an intriguing phenomenon in large language models (LLMs): how reasoning can enhance the recall of simple facts, even when complex logic isn't required. Their study reveals two key mechanisms at play: one acting as a computational buffer through generated reasoning tokens, and the other leveraging factual priming, where related facts are generated to aid recall. Employing the pass@k evaluation metric, the researchers demonstrated that enabling reasoning significantly boosted the models' ability to retrieve answers previously deemed unreachable while highlighting the effectiveness of generative self-retrieval in improving performance. This research carries important implications for the AI/ML community, emphasizing that reasoning in LLMs transcends mere task decomposition. It not only aids in accessing a model's internal memory but also opens new avenues to enhance model reliability by prioritizing factual accuracy. By minimizing hallucinations through structured reasoning trajectories that support verifiable intermediate facts, the findings suggest a pathway for developing more robust language models. These insights may inform future training methodologies, ultimately paving the way for LLMs that are better equipped to generate accurate and reliable knowledge responses.