论文概要
研究领域: ML 作者: Duling Xu, Zheng Chen, Zaifeng Pan 发布时间: 2025-05-15 arXiv: 2505.10889
中文摘要
最近，技能已被广泛采用于基于大语言模型（LLM）的智能体系统中。在现有框架中，技能通常在匹配到运行时任务后被注入智能体推理循环作为上下文指导，使专门的解题能力得以实现。我们发现这种执行范式引入了两个主要冗余来源：无关上下文注入和重复的技能特定推理与规划。为此，我们提出SkillSmith，一种边界优先的编译器-运行时框架，将技能包离线编译为最小可执行接口。通过从技能中提取细粒度的操作边界，SkillSmith使智能体能够在运行时动态访问和执行仅相关的组件，从而最小化不必要的上下文注入和冗余推理开销。在SkillsBench基准评估中，SkillSmith相比使用原始技能，减少了57.44%的求解阶段token使用量、42.99%的思考迭代次数、50.57%的求解时间（快2.02倍），以及57.44%的与token比例相关的货币成本。此外，由更强模型生成的编译产物可被更小或更高效的运行时模型复用，在原始技能解释失败的案例中提升任务准确率。源代码和数据可在该https URL获取。
原文摘要
Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing...
--- *自动采集于 2026-05-19*
#论文 #arXiv #ML #小凯