Research map / Multi-agent systems and games / LLM agents

Beyond Detection: Exploring Evidence-based Multi-Agent Debate for Misinformation Intervention and Persuasion

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ED2D, an evidence-retrieval extension of the Debate-to-Detect multi-agent debate framework for misinformation detection, and tests its persuasive effects against expert Snopes fact-checks in a 200-participant user study.

Problem

Multi-agent debate systems for misinformation detection focus on accuracy and rely on internal model knowledge, leaving unclear whether their debunking transcripts can persuade users and what risks arise when the system is wrong.

Approach

Augments a five-stage LLM debate (D2D) with an evidence module that extracts entities, retrieves Wikipedia content, and classifies evidence stance, then uses the debate transcripts as user-facing debunking explanations and compares them with expert fact-checks in a controlled study.

Key results

• ED2D reaches 83.59% accuracy on Weibo21, 84.45% on FakeNewsDataset, and 77.90% on Snopes25, the best among all compared methods (D2D: 82.17, 81.65, 74.11).
• In the user study, when ED2D is correct on false claims, accuracy is 80.40% vs 63.60% for control, close to Snopes at 85.60%.
• When ED2D is incorrect on false claims, participant accuracy drops to 42.00% vs 59.60% in control, and belief in false claims rises (4.44 vs 3.80).

Surprising findings

• Incorrect ED2D explanations shifted users toward wrong beliefs, lowering accuracy below the no-explanation control on both false and true claims.
• Adding evidence to the baseline prompting methods improved them uniformly, with the largest gain for zero-shot prompting.

Limitations

• Evidence retrieval is restricted to Wikipedia, which limits open-domain coverage.
• The user study uses 200 native English speakers and 10 claims per participant, so generalization is limited.
• Evaluation relies on a single base model (GPT-4o).

Experimental setup

Datasets: Weibo21 FakeNewsDataset Snopes25

Benchmarks: Weibo21 FakeNewsDataset Snopes25

Metrics: accuracy precision recall F1 belief rating sharing intention emotional agreement

Baselines: BERT RoBERTa Zero-shot prompting Chain-of-Thought Self-Reflect Standard Multi-Agent Debate Debate-to-Detect

Builds on: Debate-to-Detect (D2D) chain-of-thought self-reflection

Closest papers by meaning

• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.87)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.87)
• Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning (AAAI 2026 · similarity 0.87)
• When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.86)
• Collaborative Disagreement Resolution for Scalable Oversight (ICML 2026 · similarity 0.85)
• DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents (NeurIPS 2026 · similarity 0.85)