Multi-agent deepfake detector beats GPT and Gemini with small open-source models

Deepfake detection is stuck: benchmarks rarely cover newest generators, and single-model detectors miss subtle clues. To fix this, researchers from multiple institutions introduced FaceVid-Forensics-100K, a large-scale dataset with 100,000 videos across 33 synthesis methods—covering face swapping, reenactment, and full-face synthesis—including recent generators like Seedance 2.0. Each video comes with fine-grained textual annotations of visual artifacts and forensic explanations, automatically generated through a multi-model aggregation and conflict-resolution pipeline powered by advanced MLLMs.

On top of this, they propose a multi-agent forensic reasoning framework. Four specialized agents independently analyze forgery cues from texture, lighting, motion, and physics perspectives. A judge agent then reconciles their reports into a final prediction and explanation. Despite being composed entirely of small open-source MLLMs, the framework outperformed all baselines, including closed-source GPT and Gemini models, on out-of-domain test sets—ranking first across every reported metric. The paper and project page are available on arXiv, offering a new path toward generalizable and interpretable deepfake detection.