What it is
Framework and 17K-video dataset for training video generators to reason through temporally connected frames, with gains on four video-reasoning benchmarks. Train image-to-video models for reasoning using a 17,312-video supervision set and optional visual or textual reasoning tokens that organize intermediate reasoning state.
Key results
- OpenCoF-17K contains 17,312 videos across 11 reasoning task families
- Improves Wan2.2-I2V-A14B on MME-CoF, Gen-ViRe, VIPER, and RULER-Bench
- Adds two token designs, vt and tt, for visual cues and semantic priors
Shows video generation can serve as a reasoning process when trained with temporal supervision and explicit intermediate-state tokens.