A video can look convincing overall and still contain a telling artifact in just a few frames—or in its audio. That makes detection less like a single classification and more like an investigation: decide what to inspect, where to look, and whether the evidence is strong enough to conclude.
Key Findings
- A main investigator selects a forensic specialty and temporal window, then gathers structured reports from specialists covering appearance, physics, commonsense, on-screen text, and audio. This separates the choice of what to examine from the judgment about what the evidence shows.
- Dual-timescale policy optimization trains the investigator more frequently while refining visual inspectors more gradually, using interaction feedback and annotated artifact intervals. The aim is to improve not just the routing of inspections, but also the inspectors’ ability to recognize subtle evidence.
- The authors report 95.2% average accuracy across in-domain, out-of-domain, and agentic benchmarks, with localized evidence supporting verdicts. ForenCue-30K supplies role-specific annotations and interaction trajectories for training these components.
How It Works
Instead of relying on one free-form rationale or a fixed sequence of checks, the investigator can choose an artifact type, focus on a time window, review the specialist’s confidence and explanation, and decide whether to continue or conclude. This makes individual evidence reports easier to audit and gives the system’s interactions a role in improving both planning and inspection.
Who It Suits—and the Trade-off
A good fit for researchers studying verifiable video forensics, agentic detection, or evaluation across unseen generators. The multi-agent approach offers more targeted evidence than a single-pass detector, but its benchmark results do not by themselves establish reliability on every real-world video or against future generators; coordinating multiple inspectors may also require more inference resources.