Reasoning LLMs Reduce Tokens with Stored Skills

A novel technique for reasoning with large language models (LLMs) suggests that summarizing and archiving reusable reasoning abilities gained from thorough exploration and experimentation can be beneficial. These skills can then be accessed during inference to enhance future reasoning. In contrast to the traditional 'reasoning from scratch' approach, this method retrieves pertinent skills for each inquiry, enabling the model to bypass unnecessary detours and concentrate on productive solution strategies. When tested on coding and mathematical reasoning challenges, this technique markedly decreases the number of reasoning tokens needed while boosting overall performance. The reduced cost per request highlights significant practical and economic advantages for real-world applications.