❓ Can AI-generated videos always fool people?
👉 Nope -- we catch them in 94% of cases 😅
🎥 With DeeptraceReward (deeptracereward.github.io), we teach MLLMs to detect human-perceived fake traces in AI-generated videos — and release both training data & model!
🔥 Benchmark:
- 4.3K human annotations on 3.3K AI-generated videos + 3.3K real videos
- Pinpoints visually & temporally grounded traces that reveal "fakeness"
🤖 A 7B reward model trained on this data outperforms GPT-5 by +34.7% at detecting, grounding & explaining deepfake traces.
🧠 Key insight: saying “this is fake” is easy; locating where & when is far harder.
📄 Paper: arxiv.org/abs/2509.22646
This work is done with amazing collaborators at Penn, Princeton, and Stanford! @liusiyi64198 @yinuoxu @lupantech @guangqiusehu @tianboyang @tarananatasagar @christophershen @yikaimao @yuanzheliu @keyushshah @chungunlee @YejinChoinka @james_y_zou
@DanRothNLP @ccb 🙌 A huge shoutout to everyone!
#AI #Deepfakes #Multimodal #ComputerVision #MachineLearning #LLM #TrustworthyAI
