🤔 Even for simple factual questions that need no logic, making a model "think" (CoT) somehow raises accuracy. Google Research explains why. Title: Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs URL: Reasoning is not just task decomposition; it is also a device for pulling knowledge out of the weights. Three highlights. 🧮 Computational buffer The reasoning tokens themselves act as extra room for latent internal processing. In fact, replacing a meaningful trace with meaningless repeated text ("Let me think") of the same length still improved recall over doing nothing, though it never matched natural reasoning and plateaued when stretched too far. 🔗 Factual priming Generating related facts first primes recall of the correct answer, much like spreading activation in human cognition. Conditioning on just the facts extracted from a trace recovered most of reasoning's gains and helped even with reasoning disabled. ⚠️ Fragile to hallucination The mechanism is a double-edged sword: a single hallucinated fact in the trace sharply reduces correct final answers. Conversely, simply selecting hallucination-free trajectories at test time considerably improved accuracy. The practical takeaway: training with process rewards that encourage factually-grounded intermediate steps could boost reliability and cut hallucination vulnerability. #LLM# #Reasoning#