MERMAID：基于多智能体迭代知识溯源的记忆增强检索与推理系统

Assessing the veracity of online content has become increasingly critical. Large language models (LLMs) have recently enabled substantial progress in automated veracity assessment, including automated fact-checking and claim verification systems. Typical veracity assessment pipelines break down complex claims into sub-claims, retrieve external evidence, and then apply LLM reasoning to assess veracity. However, existing methods often treat evidence as isolated pieces, failing to fully leverage potential relationships between evidence and accumulated knowledge from multi-round interactions. MERMAID (Memory-Enhanced Retrieval and Reasoning with Multi-Agent Iterative Knowledge Grounding) addresses these limitations by introducing a memory-enhanced mechanism and a multi-agent iterative knowledge grounding framework. Central to MERMAID is its memory module, which stores and integrates evidence and reasoning outcomes acquired at different assessment stages, forming an evolving knowledge graph. The multi-agent system simulates a team of human experts, with each agent responsible for specific tasks such as evidence retrieval, sub-claim verification, or cross-validation. Agents collaborate through an iterative knowledge grounding mechanism, where the reasoning results of one agent serve as the basis for further retrieval and verification by another, enabling deep integration and correction of knowledge. This iterative process not only enhances the efficiency of evidence utilization but also identifies and resolves evidence conflicts, thereby improving reasoning robustness. MERMAID's innovation lies in its combination of memory, multi-agent collaboration, and iterative knowledge grounding, aiming for a more comprehensive and precise assessment of online content veracity, overcoming the limitations of traditional methods in evidence processing, and providing a more powerful framework for automated fact-checking systems.