The Truth Becomes Clearer Through Debate! Multi-Agent Systems with Large Language Models Unmask Fake News

• Fake-news detection matters because its spread affects individuals, vulnerable communities, and society, while existing approaches have interpretability, generalization, or reasoning limitations.Traditional models often have low interpretability and limited generalization; direct LLM prompting may underuse LLM reasoning, and single-perspective analysis may miss detection signals.Source: §1 · pp. 1–2
• TED uses multiple LLM agents in a formal debate sequence to examine and refine arguments about news truth from different perspectives.The process is inspired by Lincoln-Douglas debate and includes presenting, challenging, and refining arguments.Source: §1 · p. 2
• DebateFlow assigns agents to opposing teams for staged debate, while InsightFlow synthesizes arguments and analyzes role-argument interactions for the final judgment.The Analysis Agent uses role embeddings, a debate graph, and an attention mechanism.Source: §1 · p. 2
• TED outperformed traditional methods across multiple metrics on ARG-EN and ARG-CN, while maintaining robust performance across different LLM backbones.The authors also describe TED as supporting accurate and transparent decisions.Source: §1 · pp. 2–3
• TED is presented as an interpretable LLM-based detector that provides justifications to help readers understand its results.The framework’s structured debates and analysis are intended to illuminate logical reasoning leading to conclusions.Source: §1 · p. 2

2.1 Fake News Detection

• Earlier fake-news detection work spans text-only early detection, convolutional methods, fine-grained reasoning, multimodal data, and LLM-based methods.The authors state that prior LLM approaches did not fully use LLM potential; TED requires neither external tools nor knowledge-based retrieval.Source: §2.1 · p. 3

2.2 Multi-agent Discussion with LLMs

• Prior LLM discussion work explores role diversity, role assignment, weighted consensus, and debate, but the authors argue it does not fully depict debate or multiple perspectives.Examples include ChatEval, ReConcile [4], and multi-agent debate for mathematical ability [9].Source: §2.2 · p. 3

2.3 LLM-Based Multi-Agent System

• Although multi-agent systems commonly pass messages among agents, TED establishes staged social debate from different perspectives to determine fake-news truth.The authors describe this as the first social debate simulation in the context of fake news.Source: §2.3 · p. 3

3.1 Task Formulation

• The framework formulates fake-news detection as a multi-agent LLM debate in which opposing teams assess a news item’s veracity through structured interaction.TED aims to use LLM reasoning within this debate framework to improve detection effectiveness and interpretability.Source: §3.1 · p. 3
• TED’s InsightFlow Agents comprise a Synthesis Agent that summarizes debate records and an Analysis Agent that uses the report and original news item to predict truth.Figure 2 depicts the debate stages feeding into synthesis and then truth prediction.Source: §3.1 · p. 3, Figure 2 · p. 4, §3.4 · p. 5

3.3 DebateFlow Agents

• DebateFlow assigns LLM agents to Proponent and Opponent teams, which argue for and against a news item’s truth through opening, cross-examination, rebuttal, and closing stages.Agents challenge opposing arguments during cross-examination and rebuttal, then summarize and reinforce their team’s position in closing statements.Source: §3.3 · p. 4

3.4 InsightFlow Agents

• The Synthesis Agent turns the debate log into a concise, multi-perspective summary of both teams’ key arguments and counterarguments.This report provides a coherent overview of each side’s strengths and weaknesses for subsequent analysis.Source: §3.4.1 · p. 5
• The Analysis Agent predicts a news item’s truth value by combining a role-aware encoder, debate graph, and news-debate interactive attention.The encoder represents debate interactions with text and role information; the graph models their relations, and attention integrates debate and news representations.Source: §3.4.2 · pp. 5–6

4.1 Experiment Setup

• The evaluation uses ARG-EN and ARG-CN, derived from GossipCop and Weibo21 and deduplicated and temporally split to reduce potential data leakage.Table 2 summarizes dataset statistics.Source: §4.1.1 · p. 6, Table 2 · p. 7

4.1.2 Baselines.

• Baselines include LLM-only, BERT-based SLM-only, and LLM+SLM methods, including ARG and its rationale-free distilled variant ARG-D.The LLM-only baselines use GPT-3.5-turbo and GPT-4o-mini; the compared SLM-only approaches use BERT for text encoding.Source: §4.1.2 · p. 6

4.1.3 Implementation Details.

• The implementation uses GPT-4o-mini agents and fine-tuned BERT analysis models, with fixed debate graphs whose nodes represent interactions and edges encode their logical flow.The agents are implemented with Mesa; experiments use language-specific BERT models and run on NVIDIA 3090 GPUs.Source: §4.1.3 · pp. 6–7

4.2 Main Results

• TED achieves the highest macF1 scores and accuracy across ARG-EN and ARG-CN, outperforming the LLM-only, SLM-only, and LLM+SLM methods.Table 1 reports these comparisons; its caption notes that bolded improvements over baseline models are statistically significant at p<0.01.Source: §4.2 · p. 7, Table 1 · p. 7
• TED consistently outperforms ChatEval across all metrics on the fake news detection task.ChatEval uses One-by-One and Simultaneous-Talk strategies for multi-agent debates.Source: §4.2 · p. 7

5.1 Ablation Study

• Removing DebateFlow Agents lowers macF1 by approximately 2.3% on ARG-EN and 2.8% on ARG-CN; removing the Synthesis Agent lowers it by 1.1% and 2.4%, and the Analysis Agent by 8.0% and 4.5%, respectively.Source: §5.1 · p. 8

5.2 Performance on Different Backbones

• TED improves on the LLM-Only baseline across all evaluation metrics with both GPT-4o-mini and Deepseek backbones.The reported TED results with GPT-4o-mini and Deepseek are comparable, with a slight advantage for GPT-4o-mini.Source: §5.2.1 · p. 8, Figure 3 · p. 8
• TED improves performance across the tested open-source models, and Qwen 2.5 slightly outperforms the Llama 3.1 variants.The reported open-source-model results are lower than those obtained with GPT-4o-mini.Source: §5.2.2 · pp. 8–9, Table 4 · p. 8

5.3 Case Study

• In the case study, judges favor the pro team, and the debate report serves as a basis for determining whether the news is fake.The illustrated DebateFlow proceeds through opening, cross-examination, and closing statements.Source: §5.3 · p. 9, Figure 4 · p. 9

6 Conclusion

• On ARG-EN and ARG-CN, TED outperforms traditional machine-learning methods and standalone LLMs across multiple performance metrics.The conclusion describes its structured debate design as combining diverse perspectives and argumentation to assess news truth.Source: §6 · p. 9

A Appendix

• For ARG-CN, the authors use a Chinese version of the previously described prompt set and point to the code files for other function prompts.Source: A Appendix · p. 9