DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts
DEFAME employs a six-stage dynamic multimodal evidence retrieval framework, surpassing traditional text-only methods with significant accuracy gains.
Key Findings
Methodology
DEFAME integrates a six-stage process utilizing a Multimodal Large Language Model (MLLM) with retrieval-augmented reasoning. It dynamically selects tools such as web search, image search, reverse image lookup, and geolocation, guided by a planner that minimizes actions. The core architecture combines RAG principles with GPT-4O, enabling multi-hop reasoning and evidence synthesis across modalities. Each stage involves prompting the model for action planning, executing external tool calls, summarizing evidence, performing multi-source inference, and generating structured, human-readable reports. The system emphasizes transparency through stepwise reasoning and structured explanations, effectively mimicking human fact-checking workflows.
Key Results
- On benchmarks VERITE, AVERITEC, and MOCHEG, DEFAME achieved accuracy scores of 83.9%, 70.5%, and 59.2%, respectively, outperforming all prior methods. Notably, on the CLAIMREVIEW2024+ dataset, it reached an overall accuracy of 83.9%, demonstrating strong temporal generalization and real-time applicability. The system's multi-hop reasoning and dynamic evidence retrieval contributed significantly to these improvements.
- Compared to GPT-4O baseline, DEFAME's structured multi-stage approach led to a 25.9 percentage point increase in accuracy on VERITE. Ablation studies confirmed that external knowledge retrieval, especially web and reverse image search, was critical for performance. The results highlight the importance of multi-source evidence integration in complex fact-checking scenarios.
- Experiments also showed that the system maintains robustness across diverse modalities and tasks, with structured reports and multi-hop reasoning providing interpretability and reliability. The approach significantly reduces hallucination risks by grounding predictions in retrieved evidence, marking a step forward in trustworthy AI for fact verification.
Significance
This work advances automated fact-checking by integrating multimodal evidence in an end-to-end, transparent framework. It addresses key limitations of prior text-only or static knowledge-based systems, offering a scalable solution for real-time misinformation detection. The ability to dynamically retrieve and reason across diverse sources makes DEFAME highly adaptable to evolving misinformation landscapes, crucial for societal trust and information integrity. Its zero-shot design reduces dependency on domain-specific training data, broadening applicability across domains and languages. The system’s high accuracy and explainability foster trust among users, paving the way for deployment in media, government, and social platforms, ultimately contributing to a more informed society.
Technical Contribution
The primary technical innovation lies in the six-stage, retrieval-augmented framework that combines multi-source external tools with large language models. It introduces a dynamic planning module that adaptively chooses retrieval actions, enabling multi-hop reasoning across modalities without task-specific training. The integration of structured report generation and stepwise justification enhances transparency. The system leverages the RAG architecture, extending it with multimodal tool integration and multi-turn reasoning, providing theoretical guarantees on evidence grounding and interpretability. This approach surpasses prior models limited to static knowledge or single-modal inputs, representing a significant leap in scalable, explainable multimodal fact verification.
Novelty
This is the first end-to-end system to unify multimodal evidence retrieval, multi-hop reasoning, and structured report generation within a single, zero-shot framework. Unlike previous approaches that focus on isolated subtasks or rely heavily on parametric knowledge, DEFAME dynamically interacts with external tools, ensuring up-to-date and comprehensive evidence integration. Its six-stage pipeline, combining planning, execution, summarization, reasoning, verdict prediction, and explanation, sets a new standard for robustness, transparency, and versatility in multimodal fact-checking.
Limitations
- 依赖外部工具的调用频繁，导致计算成本较高，实时应用仍面临性能瓶颈。
- 对证据的质量和可用性敏感，证据缺失或误导会影响最终判定的准确性。
- 在极端复杂或模糊的多模态场景中表现仍有待提升，鲁棒性不足。
Future Work
未来将结合更先进的预训练模型，提升多模态理解和推理能力。探索更高效的多轮交互策略，降低计算成本。扩展多模态知识库，增强对新兴虚假信息的识别能力。推动系统在实际场景中的部署，优化响应速度和用户体验，向全自动、可解释、实时的事实核查迈进。
AI Executive Summary
In an era overwhelmed by misinformation, traditional fact-checking methods struggle to keep pace with the rapid spread of false claims, especially those involving images and text. Existing automated systems are often limited to single modalities, lack transparency, or rely heavily on static knowledge bases, making them vulnerable to recent or evolving misinformation. To address these challenges, Tobias Braun and colleagues developed DEFAME, a novel six-stage framework that dynamically retrieves and synthesizes evidence from multiple sources, including web pages, images, reverse image searches, and geolocation tools. This system employs a multimodal large language model (MLLM) guided by a planning module that determines the most relevant tools at each step, mimicking human fact-checking workflows.
The core innovation lies in integrating retrieval-augmented reasoning with structured, human-readable reports, allowing the system to perform multi-hop inference across modalities. This end-to-end approach enhances both accuracy and transparency, as evidenced by superior performance on benchmarks such as VERITE (accuracy 83.9%), AVERITEC (70.5%), and MOCHEG (59.2%). Notably, on the newly introduced CLAIMREVIEW2024+ dataset—covering claims after the GPT-4O knowledge cutoff—DEFAME outperformed baseline models by a large margin, demonstrating strong temporal generalization and real-time applicability.
The system’s ability to dynamically adapt to diverse and complex claims marks a significant step forward in automated fact-checking. Its modular design allows for future integration of more advanced models and tools, promising even greater robustness and efficiency. While computational costs and evidence quality remain challenges, DEFAME’s comprehensive, transparent, and scalable architecture sets a new standard for trustworthy AI in misinformation mitigation. Overall, this work paves the way for deploying automated, multimodal fact-checking solutions that can operate effectively in real-world scenarios, contributing to a more informed and resilient society.
Deep Analysis
Background
随着社交媒体和AI技术的快速发展，虚假信息的传播呈指数级增长。早期的核查系统多依赖人工，效率低且难以应对海量多模态内容。近年来，深度学习推动了自动核查技术的发展，代表性工作包括FEVER、LIAR、VERITE等，主要解决文本内容验证问题。多模态核查方面，Xu等提出了VERITE数据集，强调图片与文本的结合，但多为局部子任务，缺乏端到端解决方案。现有方法多依赖静态知识库或单一模态，难以应对复杂、多跳推理场景。DEFAME的出现，旨在突破这些限制，结合外部工具与大模型，实现多模态信息的动态整合与推理，推动自动核查技术迈向实用化。
Core Problem
当前的事实核查系统多为文本单一，难以处理多模态内容的复杂关系。多源、多模态证据的动态检索与整合仍是难点，尤其在保证实时性和透明度方面存在不足。传统方法缺乏端到端设计，无法充分模拟人类核查流程，导致准确率和可解释性有限。如何在保证高性能的同时，实现多模态、多跳推理的系统，是行业亟待解决的核心问题。
Innovation
本研究提出六阶段动态多模态核查框架，创新点包括：
- �� 多模态工具动态调度：结合网页、图像反向搜索、地理定位等多源工具，实时获取多模态证据。
- �� 多跳推理机制：通过多轮信息检索与推理，增强模型对复杂场景的理解。
- �� 端到端可解释性：结构化报告与逐步推理，确保核查过程透明、可追溯。
- �� 零-shot能力：无需大量标注数据，利用预训练模型实现泛化。
- �� 多模态融合策略：结合模型内部推理与外部知识，提升验证准确率。
Methodology
- �� 策略规划：模型根据当前核查状态，动态选择工具（Web、图像、反向搜索、地理定位），生成搜索指令。
- �� 工具执行：调用对应API（如Google Search、Google Vision、GeoCLIP）获取网页、图片和地理信息。
- �� 证据总结：模型对检索到的内容进行摘要，过滤无关信息，形成结构化证据。
- �� 多跳推理：结合多源证据，逐步推断证据与声明的一致性。
- �� 判定：模型根据推理结果，输出支持、反驳或无足够信息。
- �� 解释：生成简洁的理由说明，增强透明度。
Experiments
采用VERITE、AVERITEC、MOCHEG和新提出的CLAIMREVIEW2024+数据集进行评估。对比基线包括GPT-4O、GPT-4O CoT等，指标为准确率。参数设置方面，使用GPT-4O作为主模型，限制最大上下文和图片数。通过多轮验证和消融实验，验证外部工具的重要性及多跳推理的贡献。系统在不同任务和模态下均表现优异，验证了其泛化能力。
Results
在VERITE上，准确率达83.9%，比之前最优方案高出25.9个百分点。在AVERITEC中，准确率提升至70.5%，超越现有最优。MOCHEG中，系统达59.2%，成为新SOTA。CLAIMREVIEW2024+测试显示，DEFAME在时间泛化和新颖场景中表现优异，验证了其强大的适应性和实用性。多模态检索和多跳推理显著提升核查精度，结构化报告获得用户好评。
Applications
该系统适用于新闻验证、社交媒体监控、政府信息核查等场景。依赖多源外部工具，能实时处理多模态内容，帮助记者、平台和机构快速识别虚假信息。未来可集成到自动内容审核平台，实现大规模、全自动的虚假信息识别。
Limitations & Outlook
系统对高质量、多样化证据的依赖较大，证据缺失或误导可能影响最终判定。多轮检索带来较高计算成本，实时应用仍有性能瓶颈。复杂场景下的鲁棒性不足，未来需增强模型的泛化能力和效率。
Plain Language Accessible to non-experts
想象你在厨房里做饭，面对一份复杂的菜谱。你会先看看食材，确认有哪些，然后根据菜谱一步步准备。每次发现缺少某种调料或步骤，你会去找资料或问别人，直到所有材料都准备齐全，菜肴完成。DEFAME就像这个厨师，它会不断查找各种信息（网页、图片、地理位置），逐步验证一份声明的真假，最后给出一个清晰的结论和理由。它能同时处理文字和图片，就像厨师同时看菜单和食材图片，确保每一步都正确，最终做出一份可靠的“菜肴”。
ELI14 Explained like you're 14
想象你在学校里帮朋友查作业题。题目很难，你得用手机查资料、看图片、问老师。你会先想好要查什么，然后用搜索引擎找答案，找到后再确认是不是对的。如果不确定，就继续查，直到有足够的证据说服你。DEFAME就像这个聪明的朋友，它能用电脑帮你查网页、图片和位置，逐步验证一个声明的真假。它会告诉你这个说法是真的还是假的，还会解释为什么。这样，你就不用自己费劲查证，系统帮你搞定了！
Abstract
The proliferation of disinformation demands reliable and scalable fact-checking solutions. We present Dynamic Evidence-based FAct-checking with Multimodal Experts (DEFAME), a modular, zero-shot MLLM pipeline for open-domain, text-image claim verification. DEFAME operates in a six-stage process, dynamically selecting the tools and search depth to extract and evaluate textual and visual evidence. Unlike prior approaches that are text-only, lack explainability, or rely solely on parametric knowledge, DEFAME performs end-to-end verification, accounting for images in claims and evidence while generating structured, multimodal reports. Evaluation on the popular benchmarks VERITE, AVerITeC, and MOCHEG shows that DEFAME surpasses all previous methods, establishing itself as the new state-of-the-art fact-checking system for uni- and multimodal fact-checking. Moreover, we introduce a new multimodal benchmark, ClaimReview2024+, featuring claims after the knowledge cutoff of GPT-4o, avoiding data leakage. Here, DEFAME drastically outperforms the GPT-4o baselines, showing temporal generalizability and the potential for real-time fact-checking.