Research map / Video / Video-language understanding
VidGuard-R1: AI-Generated Video Detection and Explanation via Reasoning MLLMs and RL
Kyoungjun Park, Yifan Yang, Juheon Yi, Shicheng Zheng, Muhammad Muaz, Yifei Shen, Dongqi Han, Caihua Shan, Lili Qiu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal-artifact reward (GRPO-TA) to GRPO | VidGuard-R1 (GRPO) · Own dataset (CogVideoX) | accuracy (%): 81.30 -> 82.17 | Improved |
| add diffusion-step quality reward (GRPO-Q) to GRPO | VidGuard-R1 (GRPO) · Own dataset (CogVideoX) | accuracy (%): 81.30 -> 84.32 | Improved |
| add diffusion-step quality reward (GRPO-Q) to GRPO | VidGuard-R1 (GRPO) · Own dataset (HunyuanVideo) | accuracy (%): 81.90 -> 86.17 | Improved |
| replace DPO with GRPO | VidGuard-R1 · Own dataset (CogVideoX) | accuracy (%): 79.13 -> 81.30 | Improved |
| add RL (DPO) after SFT cold start | VidGuard-R1 (CoT) · Own dataset (CogVideoX) | accuracy (%): 66.18 -> 79.13 | Improved |
| increase intermediate diffusion steps from 1 to 5 (ablation) | VidGuard-R1 (GRPO-Q) · Own dataset | accuracy (%): 81.63 -> 85.80 | Improved |
| increase diffusion steps from 1 to 3 (ablation) | VidGuard-R1 (GRPO-Q) · Own dataset | accuracy (%): 81.63 -> 83.21 | Improved |
| set alpha1=0.5, alpha2=0.3 versus 0.3/0.1 (ablation) | VidGuard-R1 (GRPO-TA) · Own dataset | accuracy (%): 81.31 -> 83.57 | Improved |
| raise reward weights to alpha1=0.7, alpha2=0.5 from 0.5/0.3 (ablation) | VidGuard-R1 (GRPO-TA) · Own dataset | accuracy (%): 83.57 -> 82.53 | Hurt |
| add GenVideo data to own-dataset training (ablation) | VidGuard-R1 (GRPO) · Own dataset | accuracy (%): 81.65 -> 82.97 | Improved |
| add own data to GenVideo training (ablation) | VidGuard-R1 (GRPO) · GenVideo | accuracy (%): 97.53 -> 97.98 | Improved |
| pretrain on GenVideo then zero-shot to GenVidBench vs training on it | VidGuard-R1 (GRPO) · GenVidBench | mean Top-1 accuracy (%): 97.53 -> 96.37 | Mixed |
Closest papers by meaning
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.86)
- Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning (NeurIPS 2025 · similarity 0.86)
- GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning (NeurIPS 2025 · similarity 0.86)
- BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation (NeurIPS 2026 · similarity 0.86)
- Vid-SME: Membership Inference Attacks against Large Video Understanding Models (NeurIPS 2025 · similarity 0.85)
- Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought (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.
