← Key papers

Paper review · November 21, 2026

WeVisForensics: A self-evolving forensic agent for detecting AI-generated video

WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection

Computer visionAgentsReinforcement learningMedia Advanced Upvotes 47 cs.CV Tencent

An investigator agent and inspector subagents that are trained together examine AI-generated video step by step, reaching an average detection accuracy of 95.2% with verdicts grounded in localized evidence.

Problem

As video generation quality improves, it has become harder to detect fake videos and back up the verdict with evidence. Existing methods that judge in a single CoT pass miss short or localized artifacts. Fixed multi-agent workflows cannot flexibly choose which checks to run. Training only the investigation policy has another limitation: the detection ability of the fixed expert models themselves does not improve.

Method

After viewing an overview of the video, the main investigator agent decides which inspectors from the Forensic Specialist Library (FSL) to call and on which segments. The inspectors consist of visual inspectors that examine appearance, physical consistency, common sense and on-screen text (OCR), plus audio tools that perform voice profiling and ASR analysis. Up to 4 calls are allowed, and the investigator combines the reports to judge the video real or fake. Training uses DTPO (Dual-Timescale Policy Optimization). The investigator is updated quickly by reinforcement learning on every batch. The visual inspectors are improved on a slower cycle through LoRA Distillation and skill card revisions, using the calls the investigator actually made that match annotated artifact segments. The audio tools and OCR detector are frozen to keep the measurement criteria stable. For training data, the authors built ForenCue-30K, which contains role-specific forensic annotations and interaction Trajectories.

Results

Limitations

The provided text cuts off in the data construction section, so detailed experimental results and Ablations could not be checked. The audio tools are frozen because labeled audio forgery data is scarce, so audio-side detection ability does not improve through training. Calls are limited to 4, and for videos longer than 30 seconds only a segment of roughly 15 seconds is examined. The generated videos used for training are limited to the specified generation models, and generalization to new generators was verified only within the scope of the OOD benchmarks.

Why it matters

It records deepfake detection results as auditable evidence of "which segment is fake and on what grounds," a form well suited to media verification and platform operations. The structure that evolves the investigation policy and the inspector models together is also a useful reference for training other multi-agent systems.

Key terms

DTPO(Dual-Timescale Policy Optimization)
A two-timescale training method that updates the investigator quickly on every batch and the inspectors slowly on a periodic schedule. It reduces the instability that arises when each becomes a moving target for the other.
Forensic Sub-agent Library(FSL)
A collection of inspectors, each focused on one perspective such as appearance, physics, common sense, OCR or audio. Each inspector returns a verdict such as issue/no_issue/inconclusive, a confidence score and an explanation of its evidence in a structured format.
Skill card
A list of "report this / ignore this" rules appended after an inspector's role instructions. A separate reflector model revises these rules during training.

Original

Authors: Yipeng Du, Yuxi Li, Ying Zhang, Chong Sun, Chen Li, Jing Lyu, Ying Tai · Posted 2026-11-16 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.