Research map / Multi-agent systems and games / LLM agents
Thinking as Society: Multi-Social-Agent Self-Distillation for Multimodal Misinformation Detection
Yifei Gao, Ning Xu, Wenhui Li, Hongshuo Tian, Lanjun Wang, Anan Liu
Read the paperCheck your idea against 137,217 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.
Proposes a multi-social-agent self-distillation framework for multimodal misinformation detection, using MLLM agents that simulate diverse social viewpoints.
Problem
Multimodal misinformation detection in mixed-source scenarios requires robust reasoning over social complexity and diverse forgery types; single-agent MLLM methods give only limited single-view analysis.
Approach
Uses multiple social agents to provide diverse analytic perspectives and distills their collective reasoning into a single model via self-distillation, addressing the single-view versus costly multi-agent trade-off (abstract is truncated, so details beyond this are not confirmed).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train Qwen2-VL on SCoT data with SCPO | Qwen2-VL-7B · MFC-Bench overall | accuracy: 57.24 -> 67.15 | Improved |
| add social correction value scaling to ORPO loss (SCPO vs ORPO) (ablation) | SCoT-trained Qwen2-VL-7B · MFC-Bench overall | accuracy: 66.30 -> 67.15 | Improved |
| add preference loss on top of SFT (ORPO vs SFT) (ablation) | SCoT-trained Qwen2-VL-7B · MFC-Bench overall | accuracy: 64.20 -> 66.30 | Improved |
| add DPO after SFT (ablation) | SCoT-trained Qwen2-VL-7B · MFC-Bench overall | accuracy: 64.20 -> 57.98 | Hurt |
| SFT on SCoT data instead of self-consistency prompting (ablation) | Qwen2-VL-7B · MFC-Bench overall | accuracy: 61.63 -> 64.20 | Improved |
| self-consistency prompting (ablation) | Qwen2-VL-7B · MFC-Bench OOC | accuracy: 63.85 -> 52.60 | Hurt |
| self-consistency prompting (ablation) | Qwen2-VL-7B · MFC-Bench overall | accuracy: 57.46 -> 61.63 | Improved |
| apply MMD-Agent multi-agent framework | Qwen2-VL-7B · MMFakeBench (closed prompting) | accuracy: 41.2 -> 35.6 | Hurt |
| apply MMD-Agent multi-agent framework | LLaVA-Next-34B · MMFakeBench (closed prompting) | accuracy: 46.6 -> 40.5 | Hurt |
| apply MMD-Agent multi-agent framework | Qwen2-VL-7B · MMFakeBench (closed prompting) | F1: 31.8 -> 32.8 | Improved |
| train with SCPO | Qwen2-VL-7B · MMFakeBench (open prompting, Top-1) | accuracy: 41.0 -> 59.2 | Improved |
| train with SCPO | Qwen2-VL-7B · MFC-Bench reasoning quality | misleadingness: 2.82 -> 2.51 | Improved |
Closest papers by meaning
- Beyond Detection: Exploring Evidence-based Multi-Agent Debate for Misinformation Intervention and Persuasion (AAAI 2026 · similarity 0.88)
- Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning (AAAI 2026 · similarity 0.86)
- MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs (ICLR 2025 · similarity 0.86)
- SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory (NeurIPS 2026 · similarity 0.86)
- iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference (AAAI 2026 · similarity 0.85)
- Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
