Hallucination as Context Drift: Synchronization Protocols for Multi-Agent LLM Systems

The paper introduces the concept of context drift in multi-agent LLM systems, identifying it as a key factor in hallucination occurrences due to discrepancies in agents' internal knowledge states. It presents the Context Divergence Score (CDS) for quantifying these discrepancies and proposes the Shared State Verification Protocol (SSVP) to facilitate periodic state exchanges, which significantly reduces hallucination rates by 30% while utilizing 58% fewer API calls compared to naive synchronization methods. This work reframes hallucination mitigation as a distributed systems challenge, emphasizing the importance of context synchronization in the design of multi-agent LLM architectures.