跳转至

文章背景与核心概要

随着企业业务场景日益复杂，AI 智能体（Agent）系统极大地依赖长篇的 SKILL（技能）文档来规范工作流。这给合规性检测带来了双重挑战：大模型带来高昂的推理成本，而小模型往往难以保证准确性。

为了突破这一瓶颈，研究人员提出了 SkillCDG——一个创新的图结构框架，将长 SKILL 合规性视为逻辑推理任务。该框架将复杂的业务策略构建为双层约束依赖图，并采用闭包扎根检测（closure-grounded detection）与缩放引导的策略内蒸馏（scaling-guided on-policy distillation）。实验表明，SkillCDG 在显著减少 Token 消耗的同时，大幅超越了传统基线方法。

Long SKILL Compliance as Logical Reasoning: Closure-Grounded Detection with Scaling-Guided On-Policy Distillation

Executive Summary

As enterprise business scenarios grow increasingly complex, AI agent systems heavily rely on long SKILL documents to govern workflows. This introduces a dual challenge for compliance detection: large models incur prohibitive inference costs, while smaller models often struggle with accuracy.

To overcome this, researchers propose SkillCDG, an innovative graph-based framework that treats long SKILL compliance as a logical reasoning task. By structuring complex business policies into a two-layer constraint dependency graph and employing closure-grounded detection alongside scaling-guided on-policy distillation, SkillCDG significantly outperforms traditional baselines while dramatically reducing token consumption.

随着企业业务场景日益复杂，AI 智能体系统极大地依赖长 SKILL 文档来规范工作流。这给合规检测带来了双重挑战：大模型带来高昂的推理成本，而小模型往往难以保证准确性。

为了克服这一问题，研究人员提出了 SkillCDG，这是一个创新的图基础框架，它将长 SKILL 合规性视为逻辑推理任务。通过将复杂的业务策略结构化为双层约束依赖图，并采用闭包扎根检测以及缩放引导的策略内蒸馏，SkillCDG 在显著降低 Token 消耗的同时，大幅超越了传统的基线方法。

Paper Metadata

  • arXiv Identifier: arXiv:2608.08146 [cs.AI]
  • Submitted on: August 8, 2026
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Authors:
  • Shuaitao Zhao
  • Feng Ni
  • Lichao Ma
  • Jiaye Lin
  • Fei Han
  • Yang Wei
  • Lu Pan

论文元数据

  • arXiv 标识符： arXiv:2608.08146 [cs.AI]
  • 提交时间： 2026年8月8日
  • 主要学科： 人工智能 (cs.AI)
  • 作者：
  • Shuaitao Zhao
  • Feng Ni
  • Lichao Ma
  • Jiaye Lin
  • Fei Han
  • Yang Wei
  • Lu Pan

Abstract

The increasing complexity of enterprise business scenarios has promoted the widespread adoption of long SKILL documents in agent systems, posing new challenges for compliance detection: large models incur substantial inference costs, while small models may fail to maintain detection accuracy.

To address this gap, the authors propose SkillCDG, a graph-based framework for long SKILL compliance detection. SkillCDG represents complex business policies as a two-layer constraint dependency graph: 1. Upper Layer: Indexes SKILL descriptions for scenario routing. 2. Lower Layer: Captures dependencies among atomic constraints within each SKILL.

During inference, a two-level retrieval process followed by dependency closure supports compliance judgment and source traceability.

Comprehensive evaluations across three enterprise datasets and two controlled public benchmark variants demonstrate that SkillCDG outperforms baseline methods by up to 12.8 percentage points in detection F1 score while reducing token consumption by a maximum of 64.3%.

摘要

企业业务场景复杂性的不断增加，推动了长 SKILL 文档在智能体系统中的广泛应用，这也为合规检测带来了新的挑战：大模型会产生巨额推理成本，而小模型则难以维持检测准确率。

为了填补这一空白，作者提出了 SkillCDG，这是一个用于长 SKILL 合规检测的图结构框架。SkillCDG 将复杂的业务策略表示为双层约束依赖图： 1. 上层（Upper Layer）： 对 SKILL 描述进行索引，用于场景路由。 2. 下层（Lower Layer）： 捕获每个 SKILL 内部原子约束之间的依赖关系。

在推理过程中，通过两级检索流程结合依赖闭包，支持合规性判断与来源可追溯性。

在三个企业数据集和两个受控公共基准变体上的全面评估表明，SkillCDG 的检测 F1 分数比基线方法高出最多 12.8 个百分点，同时将 Token 消耗最多降低了 64.3%。

Key Methodology & Findings

  • Complexity-Differentiated Scaling: The study investigates the intrinsic relationships among policy-graph complexity, model scale, and detection performance. Comparative experiments across four checkpoints from a single model family validate a concise and effective scaling trend: end-to-end detection correctness exhibits a complexity-differentiated scaling pattern.
  • Instance Difficulty Quantification: The complexity metric derived from the constraint dependency graph effectively quantifies instance difficulty and a model's performance improvement potential.
  • Efficient Small-Model Enhancement: Leveraging this scaling trend, the framework executes adaptive training sample selection and applies on-policy distillation to efficiently enhance the compliance detection capabilities of smaller-scale models.

核心方法与发现

  • 复杂度差异化缩放（Complexity-Differentiated Scaling）： 本研究探讨了策略图复杂度、模型规模和检测性能之间的内在关系。通过对单一模型家族的四个检查点进行对比实验，验证了一个简洁且有效的缩放趋势：端到端检测的正确性表现出复杂度差异化的缩放模式。
  • 实例难度量化（Instance Difficulty Quantification）： 从约束依赖图中推导出的复杂度指标，能够有效地量化实例难度以及模型的性能提升潜力。
  • 高效的小模型增强（Efficient Small-Model Enhancement）： 利用这一缩放趋势，该框架执行自适应训练样本选择，并应用策略内蒸馏，从而高效地增强小规模模型的合规检测能力。

链接与资源