Peking U, Kling open Verus-Insight-8B for AI-video artifacts

A team from Peking University, Kuaishou's Kling group, HKUST (Guangzhou), Fudan and Tsinghua has released Verus-Insight-8B, an open multimodal detector that does not just label a video real or fake but walks through what it inspected and why. The paper on Hugging Face and the accompanying Verus-Bench dataset argue that current detectors can land the right verdict for the wrong reason.

The authors formalize artifact detection as three capabilities: Inspection Planning (what to look at), Evidence Verification (does that content actually show an artifact), and Evidence Aggregation (combine the evidence into the real/AI call). "Starting from Qwen3-VL-8B-Instruct, we jointly train region tracking and rubric planning with a 1:1 sampling ratio," the paper reports, before full-process SFT and reinforcement learning.

The headline numbers separate detection from understanding. Against its Qwen3-VL-8B backbone, Verus-Insight lifts video-level balanced accuracy on RewardData from 48.81 to 94.79 and on Verus-Bench from 66.67 to 89.06. More telling, Finding F1, the metric for recovering human-annotated artifacts, climbs from 3.38 to 35.42 on RewardData and from 7.56 to 15.97 on Verus-Bench. The authors contrast this with another specialized detector: VideoVeritas "achieves 91.12 BA with only 8.47 Finding F1," a near-ceiling real/fake score built on little artifact evidence.

Finding F1 scores in the 15 to 35 range mean even the strongest open detector still misses most of what human annotators mark. Weights, code and the benchmark are posted; training videos were drawn from Kling-2.5, Veo 3, HunyuanVideo-1.5, LTX-2.3 and Wan2.2. The release lands amid a wider run of open-source AI stories we have tracked this week, including DeepSeek and MiniMax clearing Chinese regulatory approval to license models overseas.