Research map / Vision-language and multimodal / Multimodal LLMs

MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs

ICLR 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 2 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 MMFakeBench, a 11,000-pair mixed-source multimodal misinformation benchmark with 12 forgery subtypes, evaluates 6 detectors and 15 LVLMs zero-shot, and proposes MMD-Agent, a reasoning-action-tool-use LVLM framework that boosts F1.

Problem

Existing multimodal misinformation detection methods assume a single forgery source per sample, which does not reflect real-world settings where multiple sources of forgery coexist, and no benchmark exists for mixed-source misinformation.

Approach

Build a benchmark that mixes textual veracity distortion, visual veracity distortion, and cross-modal consistency distortion (12 subtypes) using AI generators and curated real data, and propose MMD-Agent, which decomposes detection into three subtasks with interleaved thought/observation steps and Wikipedia retrieval, used zero-shot without weight updates.

Key results

• GPT-4V with standard prompting reaches 51.0 validation macro-F1 on 4-class MMFakeBench; MMD-Agent raises it to 61.5 on the test set.
• LLaVA-1.6-34B with MMD-Agent improves F1 from 25.7 to 49.9 on validation, approaching GPT-4V's standard-prompting score.
• Human participants achieve 35.9 F1 on the multi-class validation task, below the best LVLM results.

Measured effects reported in the paper

Surprising findings

• Human evaluators (80 participants) predicted 62.1% of items incorrectly, below GPT-4V and MMD-Agent performance.
• Fine-tuning existing detectors on MMFakeBench causes catastrophic forgetting on DGM4 while only partially generalizing.

Limitations

• External knowledge from Wikipedia API may not help with challenging natural and GPT-generated rumors.
• Data generation code is withheld for safety; access to the dataset is restricted to verified researchers under a usage agreement.
• Many small LVLMs show high prediction consistency, indicating biased single-class answers.

Experimental setup

Datasets: Politifact Gossipcop FEVER VisualNews MS-COCO Fakeddit NewsCLIPpings DGM4 COCO-Counterfactuals

Benchmarks: MMFakeBench NewsCLIPpings Fakeddit DGM4

Metrics: macro-F1 macro-precision macro-recall accuracy binary F1

Baselines: GPT-4V LLaVA-1.6-34B BLIP2-FlanT5-XXL InstructBLIP VILA HAMMER FKA-Owl FakingFakeNews LNP FakeNewsGPT4 CNNSpot UnivFD

Builds on: ReAct Fakeddit NewsCLIPpings DGM4 FEVER COCO-Counterfactuals

Closest papers by meaning

• Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.87)
• Drifting Away from Truth: GenAI-Driven News Diversity Challenges LVLM-Based Misinformation Detection (AAAI 2026 · similarity 0.87)
• SNIFFER: Multimodal Large Language Model for Explainable Out-of-Context Misinformation Detection (CVPR 2024 · similarity 0.87)
• Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.86)
• Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector (CVPR 2025 · similarity 0.86)
• A Hitchhiker's Guide to Fine-Grained Face Forgery Detection Using Common Sense Reasoning (NeurIPS 2024 · similarity 0.86)