VidGuard-R1: AI-Generated Video Detection and Explanation via Reasoning MLLMs and RL

Kyoungjun Park, Yifan Yang, Juheon Yi, Shicheng Zheng, Yifei Shen, Dongqi Han, Caihua Shan, Muhammad Muaz, Lili Qiu

Introduction

The rapid evolution of video generation models has created a pressing need for robust detection tools. However, existing "black-box" classifiers often lack transparency, providing binary "real" or "fake" labels without justification. VidGuard-R1 addresses this by introducing a framework that not only detects synthetic content with high accuracy but also provides human-interpretable explanations grounded in physical and visual logic.

The Problem: Interpretability and Shortcuts

Current video detectors face two significant hurdles. First, they lack interpretability, which is critical for legal and moderation applications where the "why" behind a decision matters. Second, many models rely on shortcut learning, identifying synthetic videos based on metadata artifacts like resolution or bitrate rather than visual content. VidGuard-R1 overcomes these by forcing the model to reason about physical violations, motion consistency, lighting, and texture.

Methodology: Reasoning-Driven Detection

The VidGuard-R1 pipeline utilizes a two-stage training process to instill forensic reasoning capabilities into a Multi-Modal Large Language Model (MLLM).

  1. SFT Cold Start: The base model (Qwen2.5-VL-7B) undergoes Supervised Fine-Tuning on 30,000 videos with high-quality Chain-of-Thought (CoT) annotations, establishing a baseline for visual description.
  2. RL Fine-Tuning with GRPO: The model is refined using Group Relative Policy Optimization (GRPO). This approach encourages the model to explore multiple reasoning paths and is enhanced by two specific variants:
  3. GRPO-TA (Temporal Artifacts): Rewards the model for identifying injected temporal glitches, forcing attention to temporal flow.
  4. GRPO-Q (Quality Evolution): Trains the model to recognize artifacts at different stages of the diffusion process, helping it understand how generative artifacts evolve.
The VidGuard-R1 reasoning pipeline illustrates how the model analyzes motion, lighting, and physics to determine video authenticity.

The VidGuard-R1 Dataset

To prevent shortcut learning, the authors curated a dataset of 140,000 real/fake video pairs. Real videos from InternVid and ActivityNet are paired with synthetic versions from HunyuanVideo and CogVideoX. Crucially, all videos are standardized to the same resolution, frame rate, and bitrate, ensuring the model learns to identify visual realism rather than technical metadata.

Key Results and Generalization

VidGuard-R1 demonstrates state-of-the-art performance across multiple benchmarks:

  • High Accuracy: It achieves up to 97.53% accuracy on GenVidBench and an F1 score of 0.96 on GenVideo.
  • Zero-Shot Robustness: The model maintains high performance (80-96%) on videos generated by unseen models like Sora, Pika 2.2, and Luma Ray2.
  • Human-Centric Explanations: In blind evaluations, VidGuard-R1’s explanations were rated as more informative and accurate than those generated by GPT-4o.
A sequence of frames analyzed by VidGuard-R1, highlighting temporal inconsistencies and physical violations in a synthetic video.

Conclusion

By combining the reasoning power of MLLMs with the optimization of Reinforcement Learning, VidGuard-R1 moves video forensics from simple classification to nuanced analysis. While challenges like computational intensity and potential hallucinations remain, the framework represents a significant step toward verifiable and trustworthy AI detection.

Summary updated

BibTeX
@inproceedings{park2026vidguard,
  author = {Kyoungjun Park and Yifan Yang and Juheon Yi and Shicheng Zheng and Yifei Shen and Dongqi Han and Caihua Shan and Muhammad Muaz and Lili Qiu},
  title = {{VidGuard-R1: AI-Generated Video Detection and Explanation via Reasoning MLLMs and RL}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=gXjOsBcXIR},
}
Figure