Web Retrieval Agents for Evidence-Based Misinformation Detection

TL;DR

This paper develops an agent-based automated fact-checking approach for detecting misinformation that is robust across multiple models, outperforming alternatives and increasing the macro F1 of misinformation detection by as much as 20 percent compared to LLMs without search.

Abstract

This paper develops an agent-based automated fact-checking approach for detecting misinformation. We demonstrate that combining a powerful LLM agent, which does not have access to the internet for searches, with an online web search agent yields better results than when each tool is used independently. Our approach is robust across multiple models, outperforming alternatives and increasing the macro F1 of misinformation detection by as much as 20 percent compared to LLMs without search. We also conduct extensive analyses on the sources our system leverages and their biases, decisions in the construction of the system like the search tool and the knowledge base, the type of evidence needed and its impact on the results, and other parts of the overall process. By combining strong performance with in-depth understanding, we hope to provide building blocks for future search-enabled misinformation mitigation systems.

Figures (14)

• Figure 1: F1 improvement from Cohere RAG Search by category of missing information for each model, for statements labeled as hard and impossible to evaluate. The columns represent the category of missing information in each statement.
• Figure 2: Improvement with Cohere vs. DuckDuckGo: gpt-4-turbo-0125. The rows correspond to Possibility label (i = Impossible, h = Hard, p = Possible).
• Figure 3: Improvement with Cohere vs. DuckDuckGo: gpt-3.5-turbo.
• Figure 4: Improvement with Cohere vs. DuckDuckGo: mixtral-8x7b-instruct-v0.1-1.
• Figure 5: F1 improvement: vicuna-13b-v1.5-1.