Thinking to Recall: How Reasoning Unlocks Hidden Knowledge in LLMs
Google Research reveals why reasoning helps LLMs recall simple facts — even when no reasoning is needed. Two mechanisms: computational buffer and factual priming.
Google Research reveals why reasoning helps LLMs recall simple facts — even when no reasoning is needed. Two mechanisms: computational buffer and factual priming.
It's a well-established fact: letting language models generate step-by-step reasoning traces dramatically improves their performance on complex tasks. Chain-of-thought reasoning is the gold standard for math, coding, and multi-hop questions. But here's the puzzle that doesn't make intuitive sense — reasoning also helps LLMs answer simple factual questions where no reasoning is actually needed.
Consider a question like: "What year was Mary Engle Pennington inducted into the National Inventors Hall of Fame?" An LLM either has this fact encoded in its weights or it doesn't. There's no arithmetic to perform, no logical deduction to execute. So why would generating a reasoning trace help?
Google Research scientists Zorik Gekhman and Jonathan Herzig set out to answer this question. Their paper, "Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs", reveals two complementary mechanisms that explain this counterintuitive phenomenon — and what they found challenges our understanding of how LLMs store and retrieve knowledge.
Before investigating mechanisms, the researchers needed to establish that the effect is real and measure its magnitude. They used the pass@k metric — instead of checking just one model-generated answer, pass@k checks whether the correct answer appears within k generated attempts. This reveals the model's true knowledge boundary, rather than just its top-1 behavior.
The experimental setup was clean: they tested reasoning LLMs where reasoning can be toggled on or off, comparing the same model in both modes. The models: Gemini-2.5 Flash, Gemini-2.5 Pro, and Qwen3-32B. The datasets: SimpleQA Verified and EntityQuestions — both focused on simple, single-hop factual questions.
The results were striking. When reasoning was enabled, models successfully recalled answers that were virtually unrecoverable when reasoning was off. The pass@k curves showed a clear gap — enabling reasoning didn't just improve the ranking of correct answers, it expanded the boundary of what the model could retrieve at all.
This isn't about decomposing complex questions. The datasets were deliberately chosen for simple, single-hop questions. The improvement comes from reasoning unlocking knowledge that exists in the model's weights but is otherwise inaccessible through direct generation.
The first mechanism is rooted in the mechanics of autoregressive generation. Every time an LLM generates a token, it performs a forward pass through all its layers. More tokens means more forward passes, which means more computation time for the model to refine its internal state.
The researchers tested this with an elegant experiment. They intercepted the model's reasoning process and replaced its generated trace with a meaningless string — "Let me think" — repeated over and over until it matched the length of the original reasoning trace. Then they let the model predict the final answer conditioned on this dummy text.
Here's what happened: conditioning on meaningless text substantially improved the model's ability to recall correct answers compared to the baseline where reasoning was completely off. This provides strong evidence that simply giving the model more computational runway — more forward passes — helps it refine its internal state and retrieve hard-to-reach facts.
But the computational buffer effect has clear limits. Pushing the dummy text to longer lengths produced diminishing returns, and it never fully matched the performance of the model's natural reasoning traces. The gap between "meaningless text of the same length" and "actual reasoning" tells us something important: while extra computation helps, the actual content of the thoughts still matters.
The researchers quantified this with a reasoning effectiveness metric — a weighted average of the relative difference in pass@k between reasoning ON and OFF modes across all k values. The curve showed that natural reasoning consistently outperformed dummy text, confirming that semantic content contributes beyond raw computation.
The second mechanism is where things get really interesting. When the researchers analyzed the natural reasoning traces generated for simple factual questions, they noticed a pattern. The models weren't writing out logical proofs — they were surfacing related facts.
This mirrors a well-known phenomenon in human cognition called spreading activation. When you think about a concept, related concepts in your semantic memory become more active and easier to retrieve. Think about it: if someone asks you "What's the capital of Australia?" and you're not sure, listing Australian cities or thinking about Australian geography often brings the answer to mind. The related facts act as stepping stones.
The researchers call this factual priming, and they provide strong experimental evidence for it in LLMs.
They extracted just the concrete facts from the model's reasoning traces, applying strict filtering to strip away filler text, search plans, or explicit mentions of the final target answer. Then they conditioned the model on this short list of recalled facts and measured the effect.
The result: conditioning on a short list of recalled facts recovered most of reasoning's gains. Even more striking, factual priming helped even when reasoning was disabled — providing the fact list as additional input context to a non-reasoning model still improved recall.
When asked for the name of the 10th King of Nepal, a reasoning model might first list the previous nine kings. Recalling those first nine acts as a semantic warm-up, priming the network to successfully recall the 10th. The facts themselves are the stepping stones — not the reasoning process, but the facts the reasoning process surfaces.
Factual priming is powerful, but it introduces a fundamental vulnerability. Because the model generates these intermediate facts itself, they might be wrong. If the stepping stones are hallucinated, where does the retrieval chain lead?
The researchers built a large-scale auditing pipeline using a search-enabled verifier to independently check the correctness of every intermediate fact generated across hundreds of thousands of reasoning traces. The audit revealed a clear pattern:
If a reasoning trace contains even a single hallucinated intermediate fact, the model is significantly less likely to arrive at the correct final answer.
This suggests that factual priming, while effective, is fragile. A single poisoned step in the retrieval chain can derail the entire process. This has important implications for how we train and evaluate reasoning models.
Understanding these mechanisms opens up concrete avenues for improving model reliability. The researchers demonstrate one approach: test-time selection. Generate multiple reasoning trajectories for a single question, then retain only those containing verifiable, hallucination-free facts. Prioritizing clean trajectories considerably improves accuracy.
In practice, this could be implemented during training through process rewards that specifically encourage factually supported intermediate steps. Rather than rewarding only correct final answers, the training signal would reinforce reasoning traces that build on verified facts.
This research fundamentally changes how we should think about reasoning in LLMs. It's not just a tool for task decomposition or mathematical logic — it's a mechanism for exposing the model's internal memory. Reasoning expands the parametric knowledge boundary, allowing models to access facts they "know" but can't directly retrieve.
For practitioners, this means:
The "thinking to recall" phenomenon reveals something profound about how LLMs work. Their parametric knowledge isn't a simple lookup table — it's a structured, interconnected semantic space where the path to retrieval matters as much as the destination. Reasoning doesn't just help models think; it helps them remember.
As Gekhman and Herzig note, these insights open exciting directions for future research. By understanding that factually accurate reasoning traces yield better answers, we can design training recipes that produce models which are inherently more reliable and less prone to hallucination — not just by penalizing wrong answers, but by rewarding the right kind of thinking.
This post is based on the Google Research blog post "Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs" by Zorik Gekhman and Jonathan Herzig, published June 24, 2026. The full paper is available on arXiv.