<ul><li>Allocating more compute to large language models (LLMs) reasoning has been shown to improve their effectiveness but also increases inference time.</li><li>This paper explores whether LLMs can become faster at reasoning through recurrent exposure on relevant tasks.</li><li>The study formalizes the problem setting of LLM reasoning speedup in terms of task relevancy and compute budget calculation.</li><li>Experiments conducted show that LLMs can reason faster with past experience, achieving up to a 56% reduction in compute cost with suitable memory and reasoning methods.</li></ul>

Can Past Experience Accelerate LLM Reasoning?

Discover more