Formal Methods and LLMs Combined for AI Compliance Monitoring
A new arXiv paper proposes combining formal methods with state-of-the-art machine learning to audit and monitor AI systems, particularly large language models (LLMs), throughout their development lifecycle. The techniques enable offline auditing and online runtime monitoring of behavioral constraints like safety, norms, and regulations. The paper introduces predictive monitoring via sampling-based methods and intervening monitors that act at runtime to preempt violations. Experimental results demonstrate the effectiveness of exploiting formal syntax and semantics of linear temporal logic for compliance.
Key facts
- Paper available on arXiv with ID 2605.16198
- Focuses on AI governance dimension of monitoring and auditing
- Covers pre-deployment testing to post-deployment auditing
- Combines formal methods with state-of-the-art machine learning
- Targets black-box advanced AI systems, notably LLMs
- Introduces predictive monitoring using sampling-based methods
- Proposes intervening monitors that act at runtime
- Exploits formal syntax and semantics of linear temporal logic
Entities
Institutions
- arXiv