BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation
BusterX leverages MLLM and reinforcement learning to detect AI-generated videos, utilizing a 200K high-quality dataset and producing interpretable reasoning chains, significantly improving accuracy and explainability.
Key Findings
Methodology
This paper introduces BusterX, a deep video forgery detection framework based on multimodal large language models (MLLMs) combined with reinforcement learning (RL). The core process involves: • Constructing GenBuster-200K, a dataset of over 200,000 high-resolution real and AI-generated videos, ensuring diversity, fairness, and realism; • Designing GenBuster-Bench, a progressive benchmark divided into In-Domain, Out-of-Domain, and In-the-Wild stages, to systematically evaluate models’ generalization across different generational shifts and real-world degradations; • Employing RL training (using DAPO algorithm) to optimize BusterX, transforming forgery detection into a visual reasoning task where the reasoning chain itself acts as the detector; • Incorporating MLLM-as-a-Judge protocol to assess the quality of generated explanations, emphasizing depth, factual accuracy, and forensic relevance. The experimental results demonstrate that BusterX outperforms leading models like Qwen3.5 and Claude-Sonnet-4.6 in detection accuracy and interpretability, especially under challenging real-world scenarios.
Key Results
- On the In-Domain stage, BusterX achieves an overall accuracy of 86.7%, surpassing models like Qwen3.5-122B (67.1%) and Claude-Sonnet-4.6 (72.4%), indicating strong baseline performance; in Out-of-Domain tests, it maintains 87.8% accuracy on unseen generators, demonstrating excellent cross-generator generalization; in the Wild scenario, it reaches 84.7% accuracy, significantly outperforming general detectors and showing robustness against real-world social media degradations; meanwhile, its explanation scores in Wild reach 0.87, reflecting high-quality, physically grounded rationales.
- The model's performance scales positively with size, with larger parameter counts (up to 122B) correlating with improved detection in diverse, real-world conditions, confirming the benefit of scale; additionally, the generated rationales are rated highly by experts, validating the model's deep reasoning capabilities; ablation studies highlight the importance of RL training and progressive benchmark design in achieving these results.
- Furthermore, BusterX demonstrates resilience against sophisticated deepfakes, accurately localizing physical inconsistencies at the frame level, and providing detailed, causal explanations that align with forensic standards, thus offering a trustworthy tool for practical applications.
Significance
This research marks a significant advancement in AI-driven video forensics, addressing key limitations of existing methods by integrating deep reasoning, high-quality datasets, and rigorous progressive evaluation. Its ability to produce interpretable explanations enhances trustworthiness, crucial for legal and societal acceptance. The combination of large-scale data, domain-adaptive training, and multi-stage benchmarking sets a new standard for future research, enabling detection systems to keep pace with rapidly evolving generative technologies. The framework’s robustness in real-world scenarios paves the way for deployment in media verification, law enforcement, and cybersecurity, fostering a safer information environment. Moreover, the methodological innovations open avenues for broader applications in multimodal reasoning and AI explainability.
Technical Contribution
The paper’s key technical contributions include: • Introducing a reinforcement learning paradigm (via DAPO) to train a visual reasoning-based detection model, enabling the model to generate and evaluate reasoning chains dynamically; • Developing GenBuster-200K, a large, fair, and high-fidelity dataset that incorporates the latest generation models, addressing data bias and scarcity issues; • Designing a multi-stage benchmark (GenBuster-Bench) that systematically evaluates models’ robustness against domain shifts and generational evolution, fostering future-proof detection strategies; • Establishing the MLLM-as-a-Judge protocol for rigorous, standardized assessment of explanation quality, integrating human-like forensic reasoning into model evaluation; • Demonstrating that larger model scales (up to 122B parameters) correlate with improved detection and interpretability, confirming the scalability of the approach.
Novelty
This work is pioneering in integrating reinforcement learning into multimodal large language models for video forgery detection, emphasizing the generation of explicit reasoning chains as the core detection mechanism. Unlike prior methods that rely solely on feature extraction or binary classifiers, BusterX’s approach leverages deep causal reasoning grounded in physical and spatiotemporal analysis. Additionally, the staged benchmark framework, which evaluates models across evolving generational and real-world conditions, is a novel contribution that anticipates future challenges. The combination of high-fidelity, fairness-aware datasets with RL-driven training and interpretability assessment represents a comprehensive, forward-looking solution that sets a new paradigm in AI-based video forensics.
Limitations
- Despite its robustness, BusterX’s reliance on extensive high-quality data and computational resources limits its immediate deployment in resource-constrained environments;
- The model’s performance may degrade against future, more sophisticated generative techniques not represented in the current dataset;
- Interpretability, while improved, still depends on the quality of generated rationales, which can be affected by model size and training stability; ongoing updates and domain adaptation are necessary to maintain effectiveness.
Future Work
未来的研究将集中在：• 提升模型的训练效率，减少对大规模数据和计算资源的依赖，以实现更广泛的应用；• 持续扩展和更新数据集，涵盖最新的生成模型，确保检测技术的前瞻性；• 深入研究推理链的可解释机制，增强模型在法律和伦理场景中的可信度；• 探索多模态信息融合（如音频、文本）以增强检测的多维度能力；• 开发轻量化模型，适应边缘设备和实时检测需求，推动行业落地。
AI Executive Summary
在数字内容快速泛滥的时代，虚假视频的生成技术不断突破，逼真度日益提升，给社会带来了严峻的信息安全挑战。传统的检测方法多依赖于特征提取和浅层模型，面对新兴的深度伪造技术，表现出明显的鲁棒性不足。为应对这一局面，本文提出了BusterX，一种结合多模态大语言模型（MLLM）与强化学习（RL）的深度视频伪造检测框架，旨在实现高精度、可解释的检测能力。
首先，研究团队构建了GenBuster-200K数据集，涵盖超过20万高质量真实与伪造视频，确保多样性、公平性和真实性。这一数据集采用最新的生成模型，过滤掉偏见和低质量内容，为模型训练提供了坚实基础。随后，提出分阶段的GenBuster-Bench评估体系，将检测任务划分为In-Domain、Out-of-Domain和In-the-Wild三个阶段，逐步测试模型在不同生成技术和真实环境中的泛化能力。该体系不仅评估检测准确率，还引入MLLM-as-a-Judge协议，严格衡量模型生成的伪造解释的合理性和深度。
在技术实现上，BusterX将检测任务转化为视觉推理问题，通过生成推理链作为判别依据，避免传统二分类模型的黑箱局限。利用强化学习（采用DAPO算法）训练模型，使其在没有大量标注的情况下，自我校准推理能力。实验结果显示，BusterX在GenBuster-Bench的所有阶段均优于Qwen3.5和Claude-Sonnet-4.6等领先模型，检测准确率在In-Domain达到86.7%，在Out-of-Domain和In-the-Wild阶段分别达到87.8%和84.7%，展现出极强的泛化和鲁棒性。此外，模型在Wild场景中的伪造解释得分也达到了0.87，获得专家和自动评判的一致认可。
这项研究的意义在于：它不仅提升了视频伪造检测的技术水平，更通过引入可解释性，为行业提供了可信赖的工具，有助于打击虚假信息，维护社会信息安全。其创新的训练策略和渐进式评估体系，为未来深度伪造检测提供了标准化平台和研究方向。尽管如此，模型在极端复杂场景和低资源环境下的表现仍需进一步优化，未来将关注模型的效率和多模态融合能力，推动行业迈向更智能、更可信的未来。
Deep Dive
Abstract
As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability. However, applying Multimodal Large Language Models (MLLMs) to video forensics is currently limited by outdated datasets, simplistic evaluation protocols, and a reliance on black-box classification. To address these issues, we introduce a comprehensive dataset, benchmark, and baseline model for video forgery detection. First, we present \textbf{GenBuster-200K}, a fair dataset of over 200,000 high-quality videos sourced from state-of-the-art generators, featuring diverse real-world scenarios. Second, we propose \textbf{GenBuster-Bench}, a diagnostic benchmark spanning three progressive tracks (In-Domain, Out-of-Domain, and In-the-Wild) to evaluate models across \textit{domain shifts} and \textit{generational shifts}. It also introduces an MLLM-as-a-Judge protocol to assess the quality of the generated forensic explanations. Finally, we develop \textbf{BusterX}, an MLLM baseline with RL training. Instead of direct binary classification, BusterX formulates detection as a visual reasoning task, where the generated reasoning chain serves as detector itself. Experimental results demonstrate that BusterX outperforms several leading MLLMs (e.g., Qwen3.5, Claude-Sonnet-4.6) in both detection accuracy and rationale quality.
References (20)
GenVidBench: A 6-Million Benchmark for AI-Generated Video Detection
Zhenliang Ni, Qi Yan, Mouxiao Huang et al.
Exploring Temporal Coherence for More General Video Face Forgery Detection
Yinglin Zheng, Jianmin Bao, Dong Chen et al.
DeMamba: AI-Generated Video Detection on Million-Scale GenVideo Benchmark
Haoxin Chen, Yan Hong, Zizheng Huang et al.
FakeAVCeleb: A Novel Audio-Video Multimodal Deepfake Dataset
Hasam Khalid, Shahroz Tariq, Simon S. Woo
Protecting Celebrities from DeepFake with Identity Consistency Transformer
Xiaoyi Dong, Jianmin Bao, Dongdong Chen et al.
ID-Reveal: Identity-aware DeepFake Video Detection
D. Cozzolino, Andreas Rössler, Justus Thies et al.
VideoWorld 2: Learning Transferable Knowledge from Real-world Videos
Zhongwei Ren, Yunchao Wei, Xiao Yu et al.
Visual Instruction Tuning
Haotian Liu, Chunyuan Li, Qingyang Wu et al.
BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation
Junnan Li, Dongxu Li, Caiming Xiong et al.
Dynamic Graph Learning with Content-guided Spatial-Frequency Relation Reasoning for Deepfake Detection
Yuan Wang, Kun Yu, Chen Chen et al.
X3D: Expanding Architectures for Efficient Video Recognition
Christoph Feichtenhofer
Sa2VA: Marrying SAM2 with MLLM for Dense Grounded Understanding of Images and Videos.
Haobo Yuan, Xiangtai Li, Tao Zhang et al.
Detecting AI-Generated Video via Frame Consistency
Long Ma, Zhiyuan Yan, Qinglang Guo et al.
AI-Generated Video Detection via Spatial-Temporal Anomaly Learning
Jianfa Bai, Man Lin, Gang Cao et al.
AntifakePrompt: Prompt-Tuned Vision-Language Models are Fake Image Detectors
You-Ming Chang, Chen Yeh, Wei-Chen Chiu et al.
FaceForensics++: Learning to Detect Manipulated Facial Images
Andreas Rössler, D. Cozzolino, L. Verdoliva et al.
ViViT: A Video Vision Transformer
Anurag Arnab, Mostafa Dehghani, G. Heigold et al.
Quality-based Artifact Modeling for Facial Deepfake Detection in Videos
S. Concas, S. Cava, Roberto Casula et al.
DeCoF: Generated Video Detection via Frame Consistency
Long Ma, Jiajia Zhang, Hongping Deng et al.
GLFF: Global and Local Feature Fusion for AI-Synthesized Image Detection
Yan Ju, Shan Jia, Jia Cai et al.