Constraint Dependency Graph Retrieval for Compliance Detection on Long SKILL Documents

What it claims

SkillCDG is a compliance auditor from Meituan and university co-authors (arXiv v1, 31 October 2026). Given an interaction and an agent's response, it returns compliant or non-compliant. A non-compliant verdict also cites the violated rule and its source span. The core argument is that a verdict depends on the set of rules that are jointly required, not on rules that read as similar. A prerequisite can use words the scenario never mentions, so similarity retrieval cannot surface it. The authors target SKILL documents that can exceed ten thousand tokens and a small deployed auditor that can hold only a fraction of one.

How it works

Each document is compiled offline into a graph. Nodes are atomic rules tied to their source spans. The single edge type, require, means one rule must be judged before another. At inference the policy emits keywords that activate seed rules. The auditor then reads the seeds plus everything reachable along require edges.

Both stages belong to one policy and are trained jointly with only verdict labels. There are no retrieval-level labels, no reward model and no critic.

• Retrieval stage (GRPO): the reward combines output-format validity with downstream verdict correctness. A selection counts as valid only if it parses and reaches at least one rule.
• Judgment stage (selective on-policy distillation): the teacher scores the student's own tokens rather than generating its own. Supervision is gated by old-policy entropy, so it covers the highest-entropy reasoning tokens. Verdict and termination tokens are always included.

flowchart LR
  A[Interaction + response] --> B[Retrieval stage: keywords]
  B --> C[Seed rules]
  G[Compiled dependency graph] --> D
  C --> D[Closure along require edges]
  D -->|empty| X[Invalid, no reward]
  D -->|non-empty| E[Judgment stage: verdict + cited span]
  E -->|verdict correctness| R[GRPO reward]
  R --> B
  T[Teacher scores high-entropy tokens] --> E

Related systems differ in specific ways. GraphRAG, LightRAG and HippoRAG use semantic edges and maximize relevance. PolicyGuide is the closest prior work, but it compiles policy into a workflow graph and checks at user-turn boundaries.

Evidence

The fetched text states results only qualitatively. Accuracy and F1 improve over full-document and retrieval baselines on three enterprise datasets and two public benchmarks. Most of the gain over frozen-retrieval baselines comes from training the retrieval stage. The dependency closure is the largest single contributor. The ablation removes four components: the closure, retrieval training (frozen retrieval), the entropy gate and the teacher. The fetched text gives no figures, model sizes, dataset sizes or benchmark names, apart from a mention of CompliBench as prior work.

Limits

Completeness holds only relative to the activated seeds and the extracted require edges. A missed seed keyword or a missed edge during compilation drops a prerequisite silently. The fetched text stops at the start of the graph construction method, so the edge extraction procedure, the numeric results and the compute cost could not be checked. The fetched text mentions no code, weights or licence, and the three enterprise datasets appear to be internal.