资讯

news

ICCV高分论文｜可灵ReCamMaster在海外爆火，带你从全新角度看好莱坞大片

作者

浙江大学博士生团队提出ReCamMaster，一种可以沿新相机轨迹重新运镜的视频生成模型。该模型性能显著优于baseline方法，并发布高质量多相机同步拍摄数据集MultiCamVideo-Dataset。

这才是IMO奥赛战神：满分，5战3金，刚被MIT录取

作者

Canada’s Warren Bei is one of five IMO contestants who scored a perfect 42/42. He won three gold medals and two silver medals over the past few years, showcasing his exceptional talent in mathematics. His success was influenced by an early passion for logic puzzles and science, leading him to participate in math competitions. Despite being praised for his achievements, he is still undecided about pursuing academic research in future.

四款扩散大语言模型全部破防？上交&上海AI Lab发现致命安全缺陷

作者

扩散语言模型（dLLMs）因并行解码、双向上下文建模和灵活插入masked token而备受关注。然而，上海交通大学等团队在最新研究中指出，dLLMs存在根本性架构安全缺陷，几乎毫无防御能力。DIJA攻击无需训练或改写模型参数，就能生成有害内容，并揭示了扩散语言模型的弱点，为dLLMs的安全研究拉开序幕。