Multi-Agent Citation Integrity Verification with Chain-of-Evidence Reasoning and Gated Behavior Tree Policies

The integrity of scientific literature depends critically on accurate citations, yet miscitation—where references fail to support cited claims—remains a pervasive problem. Existing detection approaches based on semantic similarity or graph anomaly detection struggle with nuanced logical relationships and multi-hop reasoning, while LLM-based methods face hallucination risks and prohibitive computational costs. Moreover, current LLM agent architectures rely on unconstrained generation, lacking verifiable and safe execution guarantees. We propose CiteGuard, a multi-agent framework that unifies chain-of-evidence reasoning, graph-enhanced detection, and gated behavior tree policies for reliable and efficient miscitation detection. CiteGuard employs a Citation Tracing Agent for multi-hop verification, a Graph-Enhanced Detection Module with knowledge distillation for structural analysis, and a Gated Behavior Tree Policy that externalizes verification into executable, verifiable behavior trees. Experiments on three benchmarks show CiteGuard achieves state-of-the-art F1 of 0.84 while reducing LLM invocations by 30%.