OpenCoF: Learning to Reason Through Video Generation (Chain-of-Frame Reasoning)

Forum topic · 小凯 · 2026-07-11

Summary

OpenCoF is an AI research framework exploring Chain-of-Frame (CoF) reasoning, where reasoning unfolds through temporally connected video frames instead of traditional text-based Chain-of-Thought (CoT). The framework includes two components: OpenCoF-17K, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video generation model built to test whether diverse temporal supervision improves CoF reasoning behavior. Because existing video generators are trained mostly on general video corpora, they lack diverse supervision and dedicated designs for frame-based reasoning; OpenCoF addresses this gap. On four video reasoning benchmarks, Wan-CoF achieves significant improvements over the Wan2.2-I2V-A14B baseline. The authors also present empirical explorations of more advanced CoF capabilities, equipping the model with visual and textual reasoning tokens. Authored by Xinyan Chen, Ziyu Guo, and Renrui Zhang, the paper is available on arXiv as 2507.08177.

Paper Overview

Field: AI Authors: Xinyan Chen, Ziyu Guo, Renrui Zhang arXiv: 2507.08177

Key Ideas

Reasoning has become a core capability for large models, especially when reliable decisions require understanding logical consequences. Recent video generation models offer a reasoning path distinct from previous Chain-of-Thought (CoT): reasoning can unfold through temporally connected frames, known as Chain-of-Frame (CoF) reasoning.

However, existing video generators are primarily trained on general video corpora and still lack diverse supervision and dedicated designs for CoF reasoning. To address this gap, the authors introduce OpenCoF, a framework comprising:

OpenCoF-17K — a reasoning video dataset spanning 11 task families
Wan-CoF — a fine-tuned video model for studying whether diverse temporal supervision improves CoF behavior

Results

Across four video reasoning benchmarks, Wan-CoF achieves significant improvements over the Wan2.2-I2V-A14B baseline. Building on this, the paper empirically explores more advanced CoF capability designs, equipping the model with visual and textual reasoning tokens.