Research map / Large language models / AI and society
Usage Governance Advisor: From Intent to AI Governance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents the Usage Governance Advisor, a human-in-the-loop system that uses a knowledge graph, CoT auto-assisted questionnaires, LLM-as-a-judge risk prioritization, and model recommendation to surface AI use-case risks and mitigations.
Problem
Organizations deploying LLM-based AI systems must manually gather heterogeneous safety information, complete compliance questionnaires, and identify risks and mitigations, which is time-consuming and hard to scope across varying intended uses.
Approach
Elicits semi-structured compliance information from a free-text use-case intent, links questions/answers to risks via a LinkML-defined AI governance ontology and knowledge graph, recommends models using benchmark scores, and proposes guardrails and manual mitigation actions.
Key results
- Custom prompt with granite-3-8b-instruct and LLM judge reached F1 0.80 (P 0.76, R 0.84) for KG triple extraction; overfitted prompt reached F1 0.92.
- Few-shot/CoT with user choice raised dropdown question accuracy from 0.40 zero-shot to 0.915.
- Judge-based matching raised F1 substantially over exact matching (e.g., LLMGraphTransformer from 0.29 to 0.73).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Few-shot/CoT instead of zero-shot inference | Granite-3-8B-Instruct questionnaire auto-assist · Dropdown questions | accuracy: 0.40 -> 0.73 | Helped |
| Allow users to choose from suggested answers | Granite-3-8B-Instruct questionnaire auto-assist · Dropdown questions | accuracy: 0.73 -> 0.915 | Helped |
| Use Few-shot/CoT instead of zero-shot inference | Granite-3-8B-Instruct questionnaire auto-assist · Binary questions | accuracy: 0.6 -> 0.81 | Helped |
| Use Few-shot/CoT instead of zero-shot inference | Granite-3-8B-Instruct questionnaire auto-assist · Freeform questions | accuracy: 0.73 -> 0.81 | Helped |
| Allow users to choose from suggested answers | Granite-3-8B-Instruct questionnaire auto-assist · Freeform questions | accuracy: 0.81 -> 0.905 | Helped |
| Use judge matching instead of exact lexical matching | Knowledge graph triple extraction with mixtral-8x7b-instruct-v01 · Granite-8B-Code-Base-4K model card triples (RAG) | F1: 0.19 -> 0.66 | Helped |
| Accumulate extraction results across additional runs | LLMGraphTransformer with mixtral-8x7b-instruct-v01 · Granite-8B-Code-Base-4K model card triples | F1: 0.29 -> 0.35 | Helped |
| Accumulate extraction results across additional runs | GraphRAG with mixtral-8x7b-instruct-v01 · Granite-8B-Code-Base-4K model card triples | F1: 0.32 -> 0.40 | Helped |
| Use judge matching rather than exact matching | LLMGraphTransformer with mixtral-8x7b-instruct-v01 · Granite-8B-Code-Base-4K model card triples | F1: 0.35 -> 0.73 | Helped |
| Use judge matching rather than exact matching | GraphRAG with mixtral-8x7b-instruct-v01 · Granite-8B-Code-Base-4K model card triples | F1: 0.40 -> 0.73 | Helped |
| Use judge matching rather than exact matching | Custom prompt with mixtral-8x7b-instruct-v01 · Granite-8B-Code-Base-4K model card triples | F1: 0.25 -> 0.82 | Helped |
| Use an overfitted prompt instead of the best generic approach | Knowledge graph triple extraction · Granite-8B-Code-Base-4K model card triples | F1: 0.82 -> 0.92 | Helped |
Surprising findings
- Exact-string triple matching is too strict; LLM-as-a-judge matching was needed to capture semantically equivalent triples.
Limitations
- KG extraction from generative AI carries risk of hallucinated or incorrect facts, requiring human verification.
- Evaluation uses a single model card as ground truth and a small set of 42 synthetic intents.
- Mitigation actions are context-specific and need testing within each deployment context.
Experimental setup
Datasets: granite-8b-code-base-4k Hugging Face model card synthetically generated user intents (42)
Benchmarks: granite-8b-code-base-4k model card ground truth (KG triples) 42 synthetic intents with human-annotated questionnaire answers
Metrics: precision recall F1 accuracy
Baselines: zero-shot inference LLMGraphTransformer GraphRAG RAG
Builds on: LinkML Unitxt LangChain SKOS AIRO
Closest papers by meaning
- Position: Technical Research and Talent is Needed for Effective AI Governance (ICML 2024 · similarity 0.82)
- AI Governance and Lessons Learned as an AI Policy Advisor in the United States Senate (AAAI 2025 · similarity 0.82)
- AI Risk Profiles: A Standards Proposal for Pre-deployment AI Risk Disclosures (AAAI 2024 · similarity 0.80)
- Position: Standardization of Behavioral Use Clauses is Necessary for the Adoption of Responsible Licensing of AI (ICML 2024 · similarity 0.80)
- SafetyAnalyst: Interpretable, Transparent, and Steerable Safety Moderation for AI Behavior (ICML 2025 · similarity 0.79)
- AI Evaluation Authorities: A Case Study Mapping Model Audits to Persistent Standards (AAAI 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.