Researchers from Tencent’s WeChat Vision team and Nanjing University have unveiled WeVisForensics, a cutting-edge framework for detecting AI-generated video artifacts. The system leverages a self-evolving agentic architecture and dual-timescale reinforcement learning to achieve a 95.2% detection accuracy, outperforming the previous best method by 9 percentage points.
Innovative Agentic Framework
WeVisForensics employs a two-tiered agentic system. A primary "investigator" agent coordinates a library of specialized "inspector" sub-agents, each trained to detect specific anomalies such as visual inconsistencies, physical implausibilities, warped text, and mismatched audio. The investigator dynamically decides which artifact to examine, over what time window, and when to conclude the analysis, optimizing the inspection process for efficiency and precision.
Dual-Timescale Policy Optimization
The framework's dual-timescale policy optimization (DTPO) is a key innovation. The investigator agent undergoes rapid reinforcement learning, while the inspector agents are refined more gradually. This co-evolutionary approach ensures that both levels of agents improve in tandem, enabling the system to identify subtle artifacts that might otherwise be overlooked.
ForenCue-30K Dataset
To support the training and evaluation of WeVisForensics, the researchers curated ForenCue-30K, a dataset comprising role-specific forensic annotations and interaction trajectories. This dataset facilitates both supervised learning for individual inspectors and reinforcement learning for their coordination.
Performance and Open-Source Release
WeVisForensics demonstrated its effectiveness across in-domain, out-of-domain, and agentic benchmarks, achieving a 95.2% average detection accuracy. The system’s ability to ground its verdicts in localized forensic evidence further enhances its reliability. The project has been open-sourced, with the code, LoRA weights, and a Gradio demo available on GitHub at https://github.com/Tencent/WeVisForensics.
With its state-of-the-art performance and innovative architecture, WeVisForensics represents a significant advancement in the field of AI-generated video detection, addressing the growing need for reliable forensic tools in an era of increasingly sophisticated video generation technologies.


