Tencent WeChat Vision releases WeVisForensics: a lead investigator agent with self-evolving forensic inspectors hits 95.2% mean detection across domains
Synopsis
AI-generated video detection is shifting from one-pass reasoning to a dispatchable forensic investigation. Earlier methods either had a single model judge the whole video in one pass or trained only a coordinator over a fixed expert pipeline, so brief or localized artifacts were easily missed. WeVisForensics has a lead investigator agent call forensic inspector sub-agents covering appearance, physical consistency, common sense, on-screen text and audio, and a dual-timescale policy optimization, DTPO, lets the investigation policy and the inspectors' competence co-evolve; mean detection accuracy is 95.2% across in-domain, out-of-domain and agentic benchmarks, 9.0% above the strongest prior detector, with verdicts grounded in localized evidence.
Interpretation
AI-generated video detection is organized as a dispatchable forensic investigation: a lead investigator agent decides which artifact type to examine, over which time window, and when to conclude, instead of one model issuing a single judgment over the whole clip. Chain-of-thought detectors covered every artifact type in one pass, so anomalies confined to a brief interval or a small region were easily missed or explained away by fabricated rationales; fixed expert pipelines or coordinator-only training were bounded by the local accuracy of frozen sub-agents. It reaches 97.7% on the in-domain GenBuster-Bench, 3.4% above VideoVeritas and 11.3% above ARGUS, 97.5% on out-of-domain GenBuster++, and an 88.0% mean over the six generator categories of MintVid.
The system consists of a lead agent and a forensic sub-agent library on a shared Qwen3.5-4B backbone, with inspectors for appearance, physical consistency, common sense and on-screen text, plus audio analysis when a soundtrack is present. Each inspection returns a verdict, a confidence and a timestamped explanation for a single forensic perspective, so every piece of evidence can be audited, supervised and credited on its own rather than entangled in one free-form rationale. The appearance, physical and commonsense inspectors read frames of the requested window at 2 FPS, the OCR inspector localizes text on the single frame at the requested timestamp and judges enlarged crops, and the audio tools profile speakers and transcribe the whole soundtrack.
DTPO closes the loop on two timescales: the lead agent's inspection policy updates quickly on every rollout batch, while inspectors slowly accumulate the calls the lead agent actually issues and refine their weights and reporting rule cards. Updating both levels at once makes each a moving target for the other, and optimizing only the investigation policy cannot directly improve fixed experts' forensic capability. Ablations show GRPO over a frozen library gives limited average gains, slow inspector evolution adds 1.6% on average, and adding call-level credit reaches the best overall average of 95.2%, 3.9% above SFT.
The work also builds ForenCue-30K, a dataset with supervision decomposed by forensic role, and AgenticCue, a benchmark that scores both verdicts and tool invocations. Existing forensic video datasets gave a single detector a verdict or a holistic rationale, providing neither role-specific evidence nor trajectories for coordinating sub-agents, and existing benchmarks scored only the final verdict. ForenCue-30K contains 117,552 role-specific records, 33,896 interaction trajectories and a 7,500-row reinforcement-learning pool; AgenticCue has 500 generated videos per category across five categories sharing 500 real videos, with no overlap with the training set.
Perspective
The demonstrated scope is real/fake judgment and evidence localization for AI-generated video, covering in-domain GenBuster-Bench, out-of-domain GenBuster++ and MintVid, and AgenticCue organized into visual, physical, OCR, commonsense and audio categories; the audio tools and the PaddleOCR detector stay fixed during training, and only the four visual inspectors and the lead agent are trained.
The improvement in inspector evidence rests on a paired human review, which preferred the evolved responses in 65 of 200 pairs against 20 and cut flawed responses from 30.5% to 8.0%; these conclusions are limited to the reported benchmarks and review setting, and the source gives no behavior under long-term generator drift.
