MERMAID: Memory-Enhanced Retrieval and Reasoning with Multi-Agent Iterative Knowledge Grounding for Veracity Assessment

TL;DR

MERMAID addresses veracity assessment by tightly integrating retrieval with reasoning in a memory-augmented, multi-agent framework. It introduces four components—Decomposer, Executor, Toolset, and a persistent Evidence Memory—to enable cross-claim evidence reuse and transparent Reasoning-Action traces via a ReAct loop with bounded steps $T_{ m max}$. Empirical results across five diverse benchmarks show state-of-the-art performance and meaningful reductions in search effort, with open-source models achieving competitive results. The work demonstrates that coupling memory with adaptive retrieval and reasoning yields scalable, interpretable veracity assessment suitable for real-world information ecosystems.

Abstract

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 retrieval as a static, isolated step and do not effectively manage or reuse retrieved evidence across claims. In this work, we propose MERMAID, a memory-enhanced multi-agent veracity assessment framework that tightly couples the retrieval and reasoning processes. MERMAID integrates agent-driven search, structured knowledge representations, and a persistent memory module within a Reason-Action style iterative process, enabling dynamic evidence acquisition and cross-claim evidence reuse. By retaining retrieved evidence in an evidence memory, the framework reduces redundant searches and improves verification efficiency and consistency. We evaluate MERMAID on three fact-checking benchmarks and two claim-verification datasets using multiple LLMs, including GPT, LLaMA, and Qwen families. Experimental results show that MERMAID achieves state-of-the-art performance while improving the search efficiency, demonstrating the effectiveness of synergizing retrieval, reasoning, and memory for reliable veracity assessment.

Figures (4)

• Figure 1: Overall comparison between our framework and existing approaches.
• Figure 2: Left) Overview of the MERMAID architecture. The framework comprises a Decomposer agent, an Executor agent, a Toolset, and a Memory module. The Decomposer agent transforms the input claim into a structured knowledge graph and topical context. The Executor agent then engages in an iterative ReAct loop, using the Toolset to retrieve evidence and evaluate the claim’s veracity. The retrieved evidence is stored in the Memory module. Right) An example of MERMAID Workflow.
• Figure 3: Tool calls with vs. without memory (GPT-5-mini).
• Figure 4: Resolution of label-prediction disagreements ($Y\neq \hat{Y}$), where $Y$ is the gold label, $\hat{Y}$ is the model prediction, and $L$ is the adjudicated label with (True, False, and Uncertain). Bars show Prediction-Correct, Label-Correct, and Uncertain; y-axis lists the disagreement configuration (e.g., $Y=T/F,\,\hat{Y}=T/F$) with group size $N$.