Constraint Dependency Graph Retrieval for Compliance Detection on Long SKILL Documents
| Source: arXiv AI
Tags: SkillCDG, compliance detection, constraint dependency graph, GRPO, on-policy distillation, SOP
SkillCDG turns long enterprise SOP documents into constraint dependency graphs so small auditor models retrieve every mandatory prerequisite rule, improving compliance-check accuracy and F1 over full-document and retrieval baselines.
Details
LLM agents in businesses operate under long, cross-referenced SOP documents the authors call SKILL documents. These exceed small auditors' context budgets, and similarity retrieval pulls in irrelevant rules while missing mandatory prerequisites. SkillCDG represents each document as a graph of atomic rules anchored to source text, with prerequisite relations as edges. Keywords activate seed rules and their dependency closure gives a compact, conditionally complete rule context within budget. Retrieval is trained with group-relative policy optimization using a verifiable reward, and selective on-policy distillation targets high-entropy reasoning tokens. Tests on three enterprise datasets and two public benchmarks show higher accuracy and F1 than baselines.