VideoJev：面向视频理解与交互的 System One 模型

AI 摘要

研究团队提出 VideoJev，一个基于 System One 范式的视频理解与交互模型系列（4B 和 35B-A3B）。该模型用共享打分头替代自回归答案解码，统一了时间定位与多选视频问答，并支持缓存推理与联合推理以复用视觉计算。在三个时间定位基准上，平均 mIoU 相比 Qwen 主干分别提升 6.7 和 10.8 点，推理加速 2.8-3.1 倍；在四个视频问答基准上性能与主干 Thinking 模式相当，推理加速超过 100 倍。训练评估代码与两个检查点均已开源。

正文节选

VideoJev: System One Model for Video Understanding and Interaction Abstract Visual agents must turn video observations into timely decisions to track progress and choose what to do next. Building on the emerging System One paradigm, we introduce VideoJev, a family of models that cast video understanding and interaction as structured decision tasks. We adapt pretrained multimodal large language models (MLLMs) by replacing autoregressive answer decoding with a shared head that scores temporal inte

发布时间：—

抓取时间：2026-10-14 22:42

来源机构：Hugging Face