Scholarlink

The rise of misinformation on digital platforms creates a pressing need for robust fact-checking (FC) methods. While large language models (LLMs) have achieved remarkable success in natural language processing, fact-checking complex claims remains a critical challenge. Current state-of-the-art approaches, such as ProgramFC and FOLK, suffer from significant limitations, including inefficiency, high costs, and restricted expressiveness in reasoning. In this work, we propose FOLDCoVe, a novel method that addresses these challenges. FOLDCoVe employs a dual-LLM framework, leveraging a powerful LLM to translate claims into First-Order Logic (FOL) and a cost-effective LLM to verify predicates against lengthy sources. Our approach significantly outperforms state-of-the-art methods on the HOVER multi-hop reasoning benchmark in terms of fact-checking accuracy, while offering concise prompts and interpretable outputs, making it accessible to non-technical users.

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

Fact-Checking with Large Language Models via Cost-Effective First-Order Logic Reformulation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sara Asghari,
  • Laks V. S. Lakshmanan,
  • Venkatesh Srinivasan,
  • Alex Thomo

摘要

The rise of misinformation on digital platforms creates a pressing need for robust fact-checking (FC) methods. While large language models (LLMs) have achieved remarkable success in natural language processing, fact-checking complex claims remains a critical challenge. Current state-of-the-art approaches, such as ProgramFC and FOLK, suffer from significant limitations, including inefficiency, high costs, and restricted expressiveness in reasoning. In this work, we propose FOLDCoVe, a novel method that addresses these challenges. FOLDCoVe employs a dual-LLM framework, leveraging a powerful LLM to translate claims into First-Order Logic (FOL) and a cost-effective LLM to verify predicates against lengthy sources. Our approach significantly outperforms state-of-the-art methods on the HOVER multi-hop reasoning benchmark in terms of fact-checking accuracy, while offering concise prompts and interpretable outputs, making it accessible to non-technical users.