Thinking to recall: How reasoning unlocks parametric knowledge in LLMs

Chain-of-thought reasoning was supposed to be for hard problems: math, code, multi-step logic. So why does it also help a model answer something dead simple, like the year Mary Engle Pennington got inducted into the National Inventors Hall of Fame? There's no arithmetic there. The model either knows it or it doesn't. Google Research decided to find out why reasoning still helps, and the answer is stranger than "more thinking equals better answers."

Using Gemini 2.5 Flash and Pro, plus Qwen3-32B, the team tested closed-book QA sets like SimpleQA Verified and EntityQuestions, toggling reasoning on and off for the same models. With reasoning off, plenty of correct answers were essentially locked away, unreachable no matter how many times you asked. Flip reasoning on, and those same facts suddenly surface. That's not decomposition doing the work, since these are single-hop questions with nothing to decompose.

Two separate mechanisms turned out to be responsible. The first is almost embarrassingly mechanical: researchers replaced the model's actual reasoning text with a meaningless repeated phrase, "Let me think, let me think," padded out to the same length as a real reasoning trace. Accuracy went up anyway. Just having more forward passes, more compute time, helps the model settle into the right internal state, like giving your brain a few extra seconds before answering a trivia question. But this buffer effect plateaus, and it never catches up to real reasoning, so content clearly matters too.

That's where the second mechanism, which the researchers call factual priming, comes in. Real reasoning traces on simple questions are full of adjacent facts, not logical steps. Ask for the 10th king of Nepal, and the model tends to first recall kings one through nine, and that list of correct-but-irrelevant facts acts as a warm-up that helps it land on the tenth. It's the LLM version of spreading activation, the psychology term for how thinking about one concept makes related concepts easier to retrieve.

The catch is that these self-generated facts can be wrong. Google built a search-verified auditing pipeline across hundreds of thousands of reasoning traces and found that a single hallucinated intermediate fact meaningfully lowers the odds of getting the final answer right. So the same mechanism that unlocks buried knowledge can also derail it. The fix the team tested was selecting only reasoning paths with verifiable, hallucination-free intermediate facts, which boosted accuracy further, and they suggest training with process rewards for factual accuracy in reasoning steps as the logical next move.