Thinking with Reasoning Skills: Fewer Tokens, More Accuracy

TL;DR

This paper introduces a method to improve reasoning in large language models by summarizing and reusing reasoning skills, reducing token usage and enhancing accuracy.

Contribution

It proposes a retrieval-based approach to reuse distilled reasoning skills, moving away from reasoning from scratch in LLMs.

Findings

Reduces reasoning tokens significantly.

Improves performance on coding and math tasks.

Lowers per-request cost for practical deployment.

Abstract

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.