OpenCoF: Learning to Reason Through Video Generation

OpenCoF studies Chain-of-Frame (CoF) reasoning, where intermediate reasoning unfolds across temporally connected video frames. The framework includes OpenCoF-17K, a dataset covering 11 reasoning task families, and Wan-CoF, a model fine-tuned from Wan2.2-I2V-A14B. The paper reports substantial gains over the baseline on four video reasoning benchmarks. It also introduces visual and textual reasoning tokens to represent low-level visual cues and high-level semantic priors, then analyzes their behavior across model depth, denoising steps, spatial structure, and temporal structure. The dataset, model, and code are released for further research.