对比四种技能加载方式，发现按需加载可大幅降低令牌消耗
Skill Blocks: How Should an Agent Load Its Skill? A Caching-Correct Comparison of Pre-load, On-Demand Tool-Loading, Progressive Disclosure, and Hybrid
- 提出四种技能加载方法：全量加载、分块加载、引用加载和混合加载
- 混合加载在多轮任务中减少39.8%令牌使用，最高降耗73%
- 适合长流程任务，能显著节省计算成本
智能体技能通常在每次请求时完整注入，导致令牌开销增加。本文对比了四种保持内容完整的加载方法：全量加载、技能分块（Skill Block）、引用加载和混合加载。在SearchQA、SpreadsheetBench、ALFWorld、ScienceWorld和SynthProc数据集上，单轮任务以原始输入为基准，多轮任务采用缓存修正的有效输入进行度量。结果表明无通用最优方案。混合加载在SearchQA上减少27.4%输入，在SpreadsheetBench上减少39.8%。在大型多轮任务中，技能分块与混合加载表现优异，分别在ScienceWorld上降低62.5%和52.8%，在SynthProc上分别降低73.0%和66.6%。ALFWorld收益较小，因其流程短且重复调用频繁。成对结果测试未发现质量差异，但无法证明等价性。总体而言，当技能中大部分内容并非每轮都需时，条件性加载最有效。
原文摘要 · Abstract (English)
Agent skills are often injected in full on every request, increasing token cost. We compare four content-preserving loading methods: Full, Skill Block, Reference, and Hybrid. Across SearchQA, SpreadsheetBench, ALFWorld, ScienceWorld, and SynthProc, we measure token usage using raw input for single-turn tasks and cache-correct effective input for multi-turn tasks. Results show no universal winner. Hybrid reduces input by 27.4% on SearchQA and 39.8% on SpreadsheetBench. On large multi-turn skills, Skill Block and Hybrid achieve substantial reductions, reaching 62.5% and 52.8% on ScienceWorld and 73.0% and 66.6% on SynthProc. ALFWorld shows smaller gains because procedures are short and repeatedly needed. Paired outcome tests detect no quality differences, though they do not establish equivalence. Overall, conditional loading is most beneficial when large portions of a skill are not needed on every turn.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。