Google: Reasoning Helps Language Models Recall Facts Through Buffer and Priming
Google/DeepMind
Alibaba/Qwen
Google Research finds that enabling reasoning in LLMs helps recall simple facts via two mechanisms: extra tokens act as a computational buffer, and generating related facts primes correct recall. The study tests Gemini-2.5 and Qwen3-32B on QA datasets.
Google Research studied why reasoning helps LLMs recall simple facts when no complex steps are needed. They found two mechanisms: a computational buffer where extra meaningless tokens provide more processing, and factual priming where generating related facts primes correct recall. Experiments used Gemini-2.5 (Flash and Pro) and Qwen3-32B on SimpleQA Verified and EntityQuestions datasets. The buffer effect improved recall but had diminishing returns, while natural reasoning traces often surfacing related facts gave the best performance. Hallucinated intermediate facts reduced accuracy. A test-time selection strategy that keeps hallucination-free trajectories improved results.
- Сокращения
- LLM = Large Language Model — Большая языковая модель
- CoT = Chain-of-Thought — Цепочка рассуждений
- R-LLM = Reasoning Large Language Model — Рассуждающая большая языковая модель
- COLM = Conference on Language Modeling — Конференция по языковому моделированию
- QA = Question Answering — Ответы на вопросы
Source: Google Research —
original