MoonshotAI
精选
Kimi-K3
Kimi K3 is an open-weight, 2.8T-parameter native multimodal agentic model with a 1M-token context window, designed for frontier coding, knowledge work, and reasoning tasks.
AI 仓库情报摘要
FR-AI / ANALYSIS为什么值得关注
它开创了一种不同于Chain-of-Thought的新型推理范式（Chain-of-Frame），提供了包含四种数据管道的多样化推理视频数据集，并通过数据监督和推理标记设计在外部视频推理基准上展示了显著性能提升。
适合谁使用
- 视频生成和多模态推理的研究人员
- 从事大型模型推理的AI实践者
- 合成数据集管道的开发者
- 研究Chain-of-Thought替代方案的学者
典型使用场景
- 改进视频生成模型以处理逻辑推理任务
- 在多个任务上对视频推理能力进行基准测试
- 通过实例渲染、专家引导、程序化场景合成和外部视频再利用生成合成推理视频
- 探索视频生成中基于标记的推理机制
项目优势
- 提出了一种新颖的Chain-of-Frame推理范式，利用时间帧连接进行推理
- 构建了大规模推理视频数据集OpenCoF-17K，涵盖11个任务家族和四种互补数据管道
- Wan-CoF模型在四个外部视频推理基准上优于基线Wan2.2
- 视觉和文本推理标记设计通过捕捉低层和高层线索进一步提升了性能
使用前须知
- 代码、数据集和模型检查点尚未公开（待内部审查）
- 框架仅基于单一基础模型Wan2.2构建，未验证对其他架构的通用性
- 评估仅限于四个特定基准，实际应用可能需进一步测试
README 快速开始
OpenCoF: Learning to Reason Through Video Generation
Official repository for the paper "OpenCoF: Learning to Reason Through Video Generation".
💥 News
- [2026.07] Code, dataset, and model release is pending internal company review.
👀 About OpenCoF
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 built around OpenCoF-17K, a reasoning video dataset of 17,312 videos spanning 11 task families, curated through four complementary pipelines: instance-based rendering, expert-guided rendering, procedural scene synthesis, and external video repurposing.
We use OpenCoF-17K to fine-tune Wan2.2-I2V-A14B into Wan-CoF, which achieves considerable gains over the baseline across four external video-reasoning benchmarks (MME-CoF, Gen-ViRe, VIPER, RULER-Bench) purely from data supervision. Building on this, we further explore two complementary reasoning-token designs, Visual Reasoning Tokens (vt) and Textual Reasoning Tokens (tt), which respectively capture low-level visual cues and high-level semantic priors, yielding Wan-CoFvt and Wan-CoFtt with further gains.
Our results suggest that stronger video reasoning requires both broad temporal supervision and explicit mechanisms for organizing intermediate reasoning state.
🚧 Code, Model & Dataset
The code, model checkpoints, and dataset are currently going through internal company review before public release.
- Code — Coming soon
- OpenCoF-17K dataset — Coming soon
- Wan-CoF model checkpoints — Coming soon
📖 Citation
If you find OpenCoF useful for your research, please consider citing our paper:
@article{chen2026opencof,
title = {OpenCoF: Learning to Reason Through Video Generation},
autho
项目描述
OpenCoF: Learning to Reason Through Video Generation
相关仓库与替代方案
根据分类、Topic 和编程语言匹配的相似项目。
xuchonglang
精选
investing-for-beginners
A structured investing guide for Chinese beginners covering US stocks, options, and cryptocurrency, with focus on foundational concepts and risk awareness.
Krishnagangwal
精选
CS-Fundamentals
A curated collection of Computer Science fundamentals (PDFs, notes, cheatsheets, interview question banks) for placement preparation, covering seven core subjects plus general resources.