[论文] OpenCoF: Learning to Reason Through Video Generation

论文概要

研究领域: AI
作者: Xinyan Chen, Ziyu Guo, Renrui Zhang
发布时间: 2026-07-10
arXiv: 2507.08177

中文摘要

推理已成为大模型的核心能力，尤其是当可靠决策需要理解逻辑后果时。近期视频生成模型提供了一条不同于先前思维链（CoT）的推理路径：推理可以通过时间连接的帧展开，即帧链（CoF）推理。然而，现有视频生成器主要在通用视频语料库上训练，仍缺乏针对CoF推理的多样化监督和专门设计。为弥补这一差距，我们引入OpenCoF框架，包含OpenCoF-17K数据集（一个涵盖11个任务家族的推理视频数据集）和Wan-CoF（一个用于研究多样化时间监督是否改善CoF行为的微调视频模型）。在四个视频推理基准上，Wan-CoF相比Wan2.2-I2V-A14B基线取得显著提升。在此基础上，我们实证探索了更先进的CoF能力设计，即为模型配备视觉和文本推理token。

原文摘要

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, still lacking diverse supervision and dedicated designs for CoF reasoning. To address this gap, we introduce OpenCoF, a framework comprising the OpenCoF-17K dataset, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video model for studying whether diverse temporal supervision improves CoF behavior. Across four video reasoning benchmarks, Wan-CoF ach...

自动采集于 2026-07-11

#论文 #arXiv #AI #小凯