A recent study demonstrates that enabling large language models capable of producing chain-of-thought (CoT) traces can surface correct factual answers that are otherwise effectively unreachable from the models’ parametric memory. Through controlled experiments the authors identify two complementary mechanisms explaining this effect: an implicit computational buffer provided by extra generated tokens and a generative ‘‘factual priming’’ effect in which related facts produced by the model facilitate retrieval of the target fact.
Why this matters
Chain-of-thought has been shown to improve performance on complex, multi-step tasks. It was unclear, however, why generating reasoning traces would help on simple, single-hop factual questions where no logical decomposition is required—for example: "What year was Mary Engle Pennington inducted into the National Inventors Hall of Fame?" Either the model already stores the fact in its weights or it does not, so the benefit of reasoning was not obvious.
Experimental setup and datasets
The work evaluates reasoning-capable LLMs (R-LLMs) where reasoning can be toggled on or off. The experiments use Gemini-2.5 (Flash and Pro) and Qwen3-32B, and focus on two challenging closed-book QA datasets: SimpleQA Verified and EntityQuestions. The primary evaluation metric is pass@k, which measures whether a correct answer appears among multiple generated attempts rather than only the top-1 output. pass@k helps estimate the potential of reasoning for factual recall by examining the model’s output distribution.
Main finding
Across the models and datasets, enabling reasoning consistently allowed the models to recall answers that were virtually unrecoverable when reasoning was disabled. Because the datasets primarily contain simple, one-hop questions, the gain cannot be attributed to decomposing complex problems.
Mechanism 1: the computational buffer
The authors test the hypothesis that extra generated tokens function as extended computation—additional forward passes that let the model refine its internal state. To probe this, they replaced natural reasoning traces with a meaningless repeated string: "Let me think", repeated to match the original trace length, and then conditioned the model on that dummy text before predicting the final answer.
Conditioning on this meaningless trace substantially improved recall relative to the no-reasoning baseline, supporting the idea that the extra tokens provide computation time that helps surface hard-to-reach parametric facts. The compute-buffer effect, however, has limits: increasing the dummy text length yields diminishing returns, and the dummy traces never fully match the performance of natural reasoning traces—indicating that the semantic content of thoughts still contributes.
Mechanism 2: factual priming
Analysis of natural reasoning traces revealed that models often list related facts rather than produce formal logical proofs. This mirrors the psychological concept of spreading activation, where retrieving one concept primes related concepts. The authors term the analogous model behavior factual priming: the model generates topically related facts that create a contextual bridge aiding retrieval of the target fact.
To isolate this effect, they filtered reasoning traces to extract only concrete facts (removing filler, search plans, or explicit restatements of the final target) and conditioned the model on these short fact lists. Doing so recovered most of the gains attributable to full reasoning and improved performance even when reasoning was switched off. For instance, when asked for the 10th King of Nepal, listing the prior nine kings acts as a semantic warm-up that helps retrieve the tenth.
The hallucination risk
Because the model self-generates intermediate facts, factual priming brings the risk of hallucinated (incorrect) intermediate statements. The authors constructed a large-scale auditing pipeline with a search-enabled verifier to check the correctness of every intermediate fact across hundreds of thousands of reasoning traces.
The audit shows a clear pattern: if a reasoning trace contains even a single hallucinated intermediate fact, the model’s probability of producing the correct final answer drops substantially. Thus, factual priming—while powerful—is fragile if intermediate facts are not reliable.
Practical implications
Understanding these mechanisms suggests concrete ways to improve model reliability. The authors implemented a test-time selection strategy that generates multiple reasoning trajectories per question and keeps only those whose intermediate facts are verifiable and hallucination-free; this prioritization significantly improves accuracy. In training, similar objectives could be encouraged using process rewards that favor factually supported intermediate steps, reducing susceptibility to hallucination.
Conclusion
The study shows that reasoning traces in LLMs do more than decompose tasks: they expose and expand the model’s parametric knowledge boundary by providing extra computation and by priming related facts. Because factual priming depends on the correctness of intermediate facts, promoting verifiable reasoning traces—either at test time through selection or during training via appropriate rewards—appears a promising path toward more accurate and reliable models.
Acknowledgements
This research was performed by Zorik Gekhman, Roee Aharoni, Eran Ofek, Mor Geva, Roi Reichart, and Jonathan Herzig. The authors thank Eyal Ben-David and Avinatan Hassidim for reviewing the work and offering valuable suggestions.
