OpenCoF Ships Chain-of-Frame Reasoning and 17K Video Dataset

A new paper out of the Wan video-model ecosystem takes the chain-of-thought idea and asks a genuinely different question: what if the intermediate reasoning trace was not text tokens at all, but a sequence of video frames. The arxiv preprint OpenCoF: Learning to Reason Through Video Generation, from Xinyan Chen, Ziyu Guo, Renrui Zhang, Dongzhi Jiang, and Hongsheng Li, calls the approach Chain-of-Frame, and argues that 'reasoning can unfold through temporally connected frames' — the video model itself generates the working-out, rather than a text decoder bolted onto it.

To back the claim, the group ships three things together. There is the OpenCoF framework, a training set they call OpenCoF-17K that spans 11 reasoning task families, and Wan-CoF, a fine-tune of the Wan2.2-I2V-A14B model that consumes both visual and textual reasoning tokens. Their reported result is 'considerable gains over the Wan2.2-I2V-A14B baseline' across four video reasoning benchmarks, and code, model weights, and dataset are open-sourced.

Why the framing matters is the direction it points. For the last two years post-training has been dominated by text-token chain-of-thought and the RL scaffolding around it — verifiers, process reward models, judge rubrics. If video-grounded tasks reason better in their native modality, that tooling stack does not port over cleanly and someone has to build the video-native counterparts. Whoever owns those tools owns a chunk of how the next generation of video and embodied models get trained.

Take the specifics as reported, not settled. The gains are self-reported against a single baseline in one model family, 17K samples is a small dataset, and the paper does not benchmark against closed frontier video generators. What the reporting doesn't give you is the inference-cost story versus text CoT, or how well the 11 task families transfer outside their own taxonomy. Still, the direction — reasoning as generated frames, not generated words — is the part worth watching, and because the release is fully open, other labs can pressure-test it themselves this quarter.