Thinking with Reasoning Skills: Fewer Tokens, More Accuracy
For practitioners deploying reasoning LLMs, this method reduces computational cost and latency without sacrificing performance, offering practical and economic benefits.
The paper proposes a method to summarize and store reusable reasoning skills from extensive deliberation, then retrieve them at inference time to guide future reasoning. This reduces reasoning tokens by 30-50% while improving accuracy by 2-5% on coding and math tasks.
Reasoning LLMs often spend substantial tokens on long intermediate reasoning traces (e.g., chain-of-thought) when solving new problems. We propose to summarize and store reusable reasoning skills distilled from extensive deliberation and trial-and-error exploration, and to retrieve these skills at inference time to guide future reasoning. Unlike the prevailing \emph{reasoning from scratch} paradigm, our approach first recalls relevant skills for each query, helping the model avoid redundant detours and focus on effective solution paths. We evaluate our method on coding and mathematical reasoning tasks, and find that it significantly reduces reasoning tokens while improving overall performance. The resulting lower per-request cost indicates strong practical and economic potential for real-world deployment.