VEPA: Verified Evolving Formal Policy for Lifelong LLM Agent Guardrails
VEPA evolves formal LLM-agent guardrails with verified symbolic policy patches that stay deterministic under new risks.
Researchers introduce VEPA, a framework that represents LLM agent guardrails as executable symbolic rules over typed events. An LLM proposes policy patches from deployment feedback, a verifier accepts only structurally valid conflict-free policies, and an evidence-aware layer withholds uncertain rules. On ASSEBench-Safety, ASSEBench-Security, and ATBench, VEPA beats baselines, including under mixed streams, corrupted feedback, and longer horizons.
- Guardrails are formal symbolic rules over typed agent events.
- LLM-proposed patches must pass a deterministic conflict-free verifier.
- Uncertain rules stay inactive until supporting evidence is sufficient.
- Gains hold on ASSEBench-Safety, ASSEBench-Security, and ATBench.
From the original article
1 min read at arxiv.orgLLM agents increasingly operate over sensitive resources, such as private documents and mutable system states, which can create serious safety and security risks. Runtime guardrails should therefore make inspectable, reproducible, and deterministic allow or refuse decisions before potentially irreversible agent actions are executed.
By Duy Nguyen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.