Scholarlink

<p>The proliferation of misinformation on digital platforms has intensified the demand for accurate and scalable fact-checking. This paper introduces FactualRAG, a retrieval augmented generation (RAG) framework that grounds large language models (LLMs) in external evidence to improve factual accuracy and transparency. FactualRAG combines hybrid retrieval with modular reasoning, aiming to reduce the risk of ungrounded model outputs and support verifiable claim checking. Evaluations across multiple datasets show substantial improvements on AVeriTeC and MultiFC, while results on Climate-Fever highlight the limits of retrieval when coverage is weak. We further test a lightweight Debate extension among multiple LLM agents, which yields incremental gains under strong evidence conditions. Among models, Gemma&#xa0;2:9b achieves the strongest results on AVeriTeC and MultiFC, though no single model dominates across all benchmarks. Overall, these findings clarify when retrieval-driven methods are most effective and how multi-agent debate can complement them, offering practical guidance for building reliable, interpretable fact-checking tools in journalism, policy, and content moderation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FactualRAG: a RAG-Powered framework for LLM fact-checking via IFCN-Accredited sources

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Lien-Jung Chang,
  • Chun-Ming Lai

摘要

The proliferation of misinformation on digital platforms has intensified the demand for accurate and scalable fact-checking. This paper introduces FactualRAG, a retrieval augmented generation (RAG) framework that grounds large language models (LLMs) in external evidence to improve factual accuracy and transparency. FactualRAG combines hybrid retrieval with modular reasoning, aiming to reduce the risk of ungrounded model outputs and support verifiable claim checking. Evaluations across multiple datasets show substantial improvements on AVeriTeC and MultiFC, while results on Climate-Fever highlight the limits of retrieval when coverage is weak. We further test a lightweight Debate extension among multiple LLM agents, which yields incremental gains under strong evidence conditions. Among models, Gemma 2:9b achieves the strongest results on AVeriTeC and MultiFC, though no single model dominates across all benchmarks. Overall, these findings clarify when retrieval-driven methods are most effective and how multi-agent debate can complement them, offering practical guidance for building reliable, interpretable fact-checking tools in journalism, policy, and content moderation.