Knowledge Graphs Boost LLM Compliance with AI Policy
A team of researchers has introduced an agentic framework aimed at generating knowledge graphs (KGs) from AI policy documents to enhance the reasoning capabilities of LLMs regarding policy adherence. They created KGs based on three AI risk-related policies utilizing two ontology schemas, assessing five LLMs across 42 policy QA tasks that encompassed six types of reasoning. The addition of KGs led to improved performance for all models, and a schema discovered by LLMs outperformed or equaled the established formal ontology. This research responds to the urgent demand for safe and secure AI as features are swiftly incorporated into software applications.
Key facts
- Framework constructs KGs from AI policy documents
- KGs built from three AI risk-related policies
- Two ontology schemas used
- Five LLMs evaluated on 42 policy QA tasks
- Six reasoning types tested: entity lookup to cross-policy inference
- KG augmentation improved scores for all five models
- Open, LLM-discovered schema matched or exceeded formal ontology
- Addresses risks from AI features in software applications
Entities
Institutions
- arXiv