Paper Overview
- Field: Computer Vision (CV)
- Authors: Bowei Liu, Zheng Lu, Yuhan Bian, Xinchen Zhang, Xingming Shui, Yuesheng Huang, Xuhuan Li, Zihao Liu, Yifan Yang, Jun Zhou, Xiu Li
- Published: 2026-08-11
- arXiv: 2608.11201
- Paper: https://arxiv.org/abs/2608.11201
Abstract (translated from Chinese summary)
Recent advances in video generation models have significantly improved the realism of synthetic videos, blurring the boundary between generated and authentic content and raising concerns about misinformation. Existing MLLM-based detectors mainly rely on supervised fine-tuning or label-level reinforcement learning, where coarse supervision limits generalization to unseen scenarios and emerging video generators.
To overcome these limitations, the authors are the first to introduce meta-detection into AI-generated video detection, enabling reliable forgery detection by jointly optimizing predicted labels and supporting evidence within reinforcement learning. This paradigm requires reliable evidence signals and effective mechanisms to integrate them into label-level optimization.
Textual rationales provide semantic descriptions of forgery artifacts, but their generation and verification depend on external models, making supervision vulnerable to hallucination and semantic bias. In contrast, temporal grounding offers more objective and verifiable evidence, since manipulated intervals can be precisely controlled during forgery construction.
Based on this insight, the paper proposes:
1. An automated data construction pipeline that generates paired real-forged videos by replacing temporal segments with outputs from a boundary-frame-conditioned video generation model. 2. Evidence-guided reward redistribution, which reallocates rewards among label-correct responses according to evidence quality, enforcing evidence-aware credit assignment while maintaining reliable label supervision. This encourages the detector to acquire fine-grained, verifiable forgery localization ability.
Extensive experiments demonstrate that VidForensics-M1 effectively leverages verifiable temporal evidence to achieve robust and generalizable AI-generated video detection.