Reproduce “QuadSentinel: Sequent Safety for Machine-Checkable Control in Multi-agent Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiliu Yang, Yilei Jiang, Qunzhong Wang, Yingshui Tan, Xiaoyong Zhu, Sherman S.M. Chow, Bo Zheng, Xiangyu Yue
QUADSENTINEL, a framework from The Chinese University of Hong Kong and Alibaba Group, provides a structured supervisory approach to enforce human-written safety policies in multi-agent LLM systems at runtime. It translates natural language policies into formal, machine-checkable rules and achieves superior safety performance with significantly lower runtime and token overhead compared to existing guardrails.
Key insights
- A modular 'agents guarding agents' architecture, combining specialized components and formal logic, enables robust and auditable safety enforcement across inter-agent communications and actions.
- Efficient state management with selective predicate updates and a risk-adaptive referee is crucial for balancing safety performance with low runtime overhead.
- Translating ambiguous natural language policies into precise, machine-checkable sequents allows for transparent, verifiable safety decisions and human-readable rationales.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.