代理技能提升工業小型語言模型

🔑 增強重點摘要

• •Agent Skill framework is widely supported by major tools like GitHub Copilot, LangChain, and OpenAI, excelling with proprietary models in context engineering, hallucination reduction, and task accuracy[1][2][3].
• •Formal mathematical definition of Agent Skill process introduced, with evaluation on SLMs across open-source tasks (e.g., FiNER) and real-world insurance claims dataset[1][2].
• •Tiny models fail at reliable skill selection in large skill hubs (50–100 skills), while 12B-30B SLMs show substantial accuracy gains, e.g., Qwen3-80B-Instruct improves from 0.198 to 0.654 on FiNER[1][2].
• •80B code-specialized SLMs match closed-source baselines in performance with better GPU efficiency, enabling secure industrial deployments without public APIs[1][2][3].
• •Evaluation metrics include Cls ACC, Cls F1 for classification, and Skill ACC for routing quality, using 4–5 distractor skills per task[2].

🛠️ 技術深入

• •Formal mathematical definition of Agent Skill process provided, focusing on skill selection (routing) and execution correctness[1][2].
• •Experiments use temporary skill repositories with 4–5 distractor skills from public hubs combined with ground-truth skills[2].
• •Three context-engineering strategies tested for impact on agent performance and efficiency in decision-making[2].
• •Metrics: Classification Accuracy (Cls ACC), F1 score (Cls F1), Skill-selection Accuracy (Skill ACC)[2].
• •Example result: Qwen3-80B-Instruct Skill ACC high, performance boosts from 0.198 (Direct Instruction) to 0.654 on FiNER task[2].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線