Auditing Cascading Risks in Multi-Agent Systems via Semantic-Geometric Co-evolution

Early warning systems can detect problems before they become obvious: The paper demonstrates that geometric patterns in how AI agents communicate can reveal instability several interaction turns before any explicit policy violations appear in the actual messages—meaning problems can be caught and addressed proactively rather than reactively.

Traditional message-by-message auditing is insufficient: Current approaches that only examine individual messages for policy compliance miss the underlying structural breakdown in how multiple agents interact with each other, allowing cascading risks to develop undetected beneath a veneer of acceptable communication.

Graph geometry offers interpretable accountability: By analyzing the communication network structure using Ollivier-Ricci Curvature, the framework can identify which specific agents or communication links are responsible for system breakdown, providing clear attribution rather than opaque "black box" warnings.

Multi-agent systems require fundamentally different oversight: Because risks in LLM-based multi-agent systems emerge from interaction dynamics rather than individual agent behavior, auditing methods must track how agents influence each other over time, not just what each agent says in isolation.

Semantic fluency masks real problems: Interactions can sound perfectly reasonable and policy-compliant on the surface while the underlying collaboration structure is deteriorating—highlighting the need for auditing methods that look beyond surface-level content to deeper patterns of system behavior.