Source-linked AI summary
VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning
Hao Tan, Jun Lan, Senyuan Shi, Zichang Tan, Zijian Yu, Huijia Zhu, Weiqiang Wang, Jun Wan, Zhen Lei
TL;DR
Reliable AI-generated video detection is needed as video generation creates security concerns, while existing MLLM detectors have limited fine-grained perception. VideoVeritas combines Joint Preference Alignment with perception pretext reinforcement learning and evaluates the resulting detector with MintVid. It reports balanced performance across diverse benchmarks and highlights the value of foundational perceptual skills for complex detection tasks.
Problem
AI-generated videos raise security concerns, while existing MLLM-based detectors struggle to combine fine-grained perception with reasoning across challenging scenarios.
Method
VideoVeritas combines Joint Preference Alignment with PPRL, using spatiotemporal grounding and self-supervised object counting rather than directly optimizing only the detection task.
Results
VideoVeritas achieves more balanced performance across diverse benchmarks than methods biased toward superficial reasoning or mechanical analysis.
Takeaways & Limitations
The results highlight the value of learning foundational perceptual skills for complex AI-generated video detection tasks.
Takeaways & Limitations
Artifact-grounding annotations from Molmo2 are coarse-grained and may be less compatible with fine-grained signals, while higher-quality annotations could produce different results.
Abstract
from arXiv · showhide
The growing capability of video generation poses escalating security risks, making reliable detection increasingly essential. In this paper, we introduce VideoVeritas, a framework that integrates fine-grained perception and fact-based reasoning. We observe that while current multi-modal large language models (MLLMs) exhibit strong reasoning capacity, their granular perception ability remains limited. To mitigate this, we introduce Joint Preference Alignment and Perception Pretext Reinforcement Learning (PPRL). Specifically, rather than directly optimizing for detection task, we adopt general spatiotemporal grounding and self-supervised object counting in the RL stage, enhancing detection performance with simple perception pretext tasks. To facilitate robust evaluation, we further introduce MintVid, a light yet high-quality dataset containing 3K videos from 9 state-of-the-art generators, along with a real-world collected subset that has factual errors in content. Experimental results demonstrate that existing methods tend to bias towards either superficial reasoning or mechanical analysis, while VideoVeritas achieves more balanced performance across diverse benchmarks.
1. Introduction
AI-generated video detection is increasingly important as generated videos spread and raise security concerns. VideoVeritas addresses limitations in existing MLLM detectors through a two-stage alignment and perception-focused training pipeline, evaluated with MintVid.
- AI-generated video detection has emerged as a critical topic because generated videos raise significant security concerns.
- Existing detectors face challenges in combining MLLM reasoning with reliable fine-grained perception across detection scenarios.
- VideoVeritas uses Joint Preference Alignment for cold-start training and PPRL to enhance detection with perception pretext tasks.
- Existing methods show imbalanced behavior, whereas VideoVeritas achieves more balanced performance across challenging subsets and multiple datasets.
- MintVid contains 3K videos from 9 state-of-the-art generators and supports evaluation of general-content, facial, and fact-based scenarios.
2. Related Work
Prior work spans image and video forgery detection, explainable MLLM detectors, and generic video-language modeling. The paper positions VideoVeritas against limitations in temporal consistency and training strategies while introducing a new evaluation dataset.
- Earlier studies focused mainly on image-level forgery detection, while newer efforts address AI-generated video detection.
- Existing video datasets often rely on outdated generators with limited temporal consistency, motivating more robust evaluation data.
- Generic MLLMs improve temporal encoding through mechanisms including absolute temporal encoding, timestamp tokens, and slow–fast frame encoding.
- Explainable video detectors use approaches including SFT, human-perceivable reasoning with timestamps and boxes, manual grounding data, and pure RL.
- VideoVeritas differs by introducing Perception Pretext RL to train perception alongside fact-based reasoning for video detection.
3. Method
VideoVeritas uses Joint Preference Alignment and perception-pretext reinforcement learning to combine fine-grained perception with fact-based reasoning for video detection. Its PPRL stage trains spatiotemporal grounding, tracking, and self-supervised counting before detection, while MintVid evaluates general-content, facial, and fact-based scenarios.
- Joint Preference Alignment: The method presents a two-stage pipeline combining Joint Preference Alignment with Perception Pretext RL for video detection.Joint Preference Alignment uses response-level and video-level preferences, while PPRL adds perception-oriented reinforcement learning before detection.
- Joint Preference Alignment: Response-level alignment prefers annotated fine-grained reasoning for perception videos but base-model reasoning for fact-based videos.For perception videos, annotated reasoning is preferred over the base model’s output; for fact-based videos, the preference is reversed.
- Perception Pretext RL: PPRL trains general spatiotemporal grounding, object tracking, and self-supervised object counting instead of directly optimizing only detection.Grounding and tracking require temporal spans or bounding boxes, while counting requires exact quantities for generated shape categories.
- Perception Pretext RL: The sequential RL setup uses 3K grounding samples, 2K counting samples, and 10K AIGC-detection samples with GSPO training.Detection combines accuracy and format rewards, with α = 0.2 controlling the format-reward contribution.
- MintVid Dataset: MintVid supports evaluation across general-content, facial, and fact-based videos, including 1.5K videos from six proprietary models and 2K facial videos from three public models.Its fact-based subset contains over 200 real and fake videos collected from TikTok, YouTube, and Bilibili and manually filtered for objective verifiability.
4. Experiments
Experiments show that VIDEOVERITAS performs robustly across diverse video-detection settings, while ablations link its gains to cold-start alignment and perception pretext training.
- Main Results: VIDEOVERITAS achieves 15.1% averaged gains over the previous best binary detector.
- Main Results: VIDEOVERITAS improves over its base model by 25.2% on average and surpasses similarly sized MLLMs.
- Main Results: Against MLLM-based detectors, VIDEOVERITAS gains +11.8% on GenBuster++, +7.2% on LOKI, and +9.1% on Jimeng3.0-Pro.
- Ablation Studies: Pure RL remains limited on challenging subsets, reaching 69.1% versus 79.2% for VIDEOVERITAS, supporting the necessity of cold-start training.
- Ablation Studies: Each perception task independently improves performance, while combining SSL and G-G yields further MintVid gains.
- Ablation Studies: 1K perception data adds +1.1% Acc and 1.3% F1 on MintVid, while around 5K provides the best efficiency-performance balance.
- Ablation Studies: Harder self-supervised counting tasks improve detection most, although the model can accurately count targets that are non-trivial for humans to perceive.
- Analysis: PPRL-trained models more often analyze fine-grained entities, object relations, and temporal changes, while case studies show VIDEOVERITAS identifying tiny-car distortion.
5. Conclusion
The paper presents VIDEOVERITAS for AI-generated video detection, combining fine-grained perception with fact-based reasoning and using PPRL to improve perceptual capacity. MintVid and extensive experiments support the value of foundational perceptual skills for complex detection.
- VIDEOVERITAS integrates fine-grained perception and fact-based reasoning for AI-generated video detection.
- PPRL improves perceptual capacities by incorporating simple perception pretext tasks.
- MintVid is a challenging dataset containing three evaluation aspects.
- Experiments highlight the value of learning foundational perceptual skills for complex detection tasks.
A. More Details of MintVid Dataset
MintVid is designed for broader evaluation through realistic generated videos, facial videos, and fact-based videos collected from online platforms.
- Facial-video generation uses curated conditional frames and prompts combining first-frame captions, action control, and background descriptions.
- MintVid general-content videos use six proprietary models, while facial videos use three specialized public models.
- Fact-based videos are collected from online platforms and manually filtered.
- Figure 10 provides an example input prompt for facial video generation.
B. Artifacts Taxonomy
The artifact taxonomy organizes common AI-generated video failures across perceptual, physical, material, and semantic dimensions.
- The taxonomy covers three main perspectives and eleven detailed artifact aspects.
- Physical-level Anomalies: Physical-level anomalies include inconsistent lighting and optics plus violations of causality or object properties.
- Flawed material simulation describes unrealistic behavior in cloth, smoke, fire, or water.
- Contextual and semantic mismatch occurs when scenes, objects, or actions contradict logic or commonsense despite realistic individual elements.
- Perceptual-level Anomalies: Perceptual anomalies include texture flicker, clarity fluctuations, unnatural trajectories, object permanence failures, structural distortion, interaction anomalies, and biological-motion irregularities.
C. Extended Details of Methodology: Prompts and Implementation Details
The methodology evaluates reasoning across five dimensions and uses perception-task prompts alongside detection prompts. Implementation details describe GSPO-based reinforcement learning and a modified objective that improves perception integration.
- Reasoning behavior evaluation: The reasoning evaluation uses component granularity, spatiotemporal continuity, physics depth, forensic objectivity, and relational logic.These dimensions assess fine-grained decomposition, temporal-spatial anchoring, physical explanations, evidence-based language, and element interactions.
- Prompts: The framework provides separate system and input prompts for detection and perception, with perception instructions varying across subtasks.The perception subtasks include spatiotemporal grounding, object tracking, object counting, and artifact grounding.
- Implementation details: The RL implementation uses GSPO, with sequence-likelihood importance ratios and group-estimated advantages.The training setup uses 8 PPUE GPUs, while the objective and advantage calculations are described through the implementation equations.
- Implementation details: Removing output-length normalization from the RL objective greatly improves performance when integrating perception learning.The authors report that batch-level perception integration is non-trivial under the original normalized formulation, while the modified objective outperforms the baseline.
D. More Experimental Results
Additional experiments show contrasting reasoning behaviors among detectors and report that some models have uneven recall across datasets. The authors also summarize unsuccessful training and reward-design attempts.
- Additional metrics: NSG-VD has high recall on several subsets but near-chance overall accuracy, while RestraV and zero-shot MLLMs have low recall on most datasets.Gemini-2.5-Pro and Gemini-3-Pro-Preview show clear leads in recall, which the authors associate with stronger perceptual capabilities.
- Reasoning comparisons: Skyra-RL produces grounded but mechanical and imprecise analyses, whereas BusterX++ is usually correct but focuses mainly on macro-level semantic concepts.These behaviors contribute to failures on fact-based videos for Skyra-RL and superficial reasoning for BusterX++.
- Explored attempts: Removing objective normalization improves perception integration, while conditional batch-level rewards and temporal anomaly grounding were less effective than phase-level or spatiotemporal tasks.More detailed detection rewards encouraging grounding formats were vulnerable to reward hacking.
E. Limitation and Future Work
The paper identifies a scope limitation in PPRL’s perception data and documents the evaluation prompts and datasets used for broader analysis. Future work centers on artifact-specific grounding data and explanation fidelity.
- Limitation and future work: PPRL mainly uses broadly available public perception resources rather than systematically investigating higher-quality artifact-specific grounding annotations.The authors propose curating and integrating such data into pretext reinforcement learning as future work.
- Evaluation details: Table 6 reports accuracy, recall, and F1 across in-domain, out-of-domain, and OOD-MintVid settings, with final averages computed across these three groups.Average Precision is used where required by the official guideline.
- Evaluation details: Table 7 presents a detailed ablation of perception tasks on MindVid using accuracy, recall, and F1.The reported format is Acc/Recall/F1.
- Reasoning evaluation: The reasoning evaluation compares reports using a specified dimension and requires judgments based only on textual evidence.The evaluation prompt asks judges to compare Assistant A and Assistant B side by side.
- Reasoning examples: The examples contrast analyses that identify physical artifacts with responses relying on generic realism cues or appearance-based reasoning.The supplied outputs include motion-grounded evidence, household-scene cues, and clothing-based judgments.