Daily papers

Friday, 20 November 2026 · 8 min read

WeVisForensics jointly trains an investigator agent and evolving forensic inspectors for AI video detection

WeVisForensics uses a main investigator agent to adaptively select forensic inspections from a self-evolving library of visual and audio sub-agents, trained via dual-timescale policy optimization to detect AI-generated videos with 95.2% accuracy and grounded evidence.

Based on the paper WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection by Yipeng Du, Yuxi Li, Ying Zhang, Chong Sun and 3 others

Challenges in detecting AI-generated videos require adaptive forensic analysis

Recent advances in AI video generation have made synthetic videos highly realistic, raising the need for reliable detection methods grounded in explicit forensic evidence across visual, textual, and audio modalities. Existing methods either perform single-pass chain-of-thought reasoning that can miss brief or localized artifacts or rely on fixed multi-agent workflows that lack flexibility in selecting relevant inspections.

Single-pass reasoning models analyze the entire video at once, often missing subtle or localized artifacts and producing free-form rationales that are hard to verify or supervise individually. Agent-based methods either use static expert pipelines or train only the coordinator agent, leaving the sub-agents fixed and limiting improvements in forensic capability.

Optimizing only the investigation policy without evolving the inspectors restricts detection performance, as inspectors that miss subtle cues bottleneck the system. There is also a lack of datasets providing role-specific forensic annotations and interaction trajectories to supervise both inspectors and their coordination.

Sign in to read the whole paper

You have read the opening. The rest is free with a Bezen account, and no card is needed.

In the rest of this paper

  • How the method works, step by step, with a flow chart
  • The key formula from the paper, explained symbol by symbol
  • Three figures from the paper, and how to read them
  • How it was tested, what was found, and where it falls short
WeVisForensics jointly trains an investigator agent and evolving forensic inspectors for AI video detection · Bezen