LongCat-Next: 将模态词化为离散标记
计算机视觉与模式识别
2026-03-31 v1 计算与语言
摘要
prevailing 的下一词预测（NTP）范式通过离散自回归建模驱动了大语言模型的成功。然而，当前的多模态系统仍然以语言为中心，通常将非语言模态视为外部附件，导致碎片化的架构和次优的集成。为了超越这一限制，我们提出了离散原生自回归（DiNA），一个统一框架，在共享的离散空间中表示多模态信息，使跨模态的一致且原则性的自回归建模成为可能。一项关键创新是离散原生任意分辨率视觉Transformer（dNaViT），它在任意分辨率下进行标记化和反标记化，将连续视觉信号转换为分层离散标记。在此基础上，我们开发了 LongCat-Next，一个原生多模态模型，在单一自回归目标下处理文本、视觉和音频，且模态特定设计 minimal。作为工业强度的基础模型，它在一个框架内擅长看、画和说，在广泛的多模态基准测试中取得了强劲性能。特别是，LongCat-Next 解决了离散视觉建模在理解任务上的长期性能瓶颈，并提供了一种统一的方法来有效调和理解与生成之间的冲突。作为迈向原生多模态的一次尝试，我们开源了 LongCat-Next 及其标记器，希望促进社区进一步的研究和开发。GitHub: https://github.com/meituan-longcat/LongCat-Next
引用
@article{arxiv.2603.27538,
title = {LongCat-Next: Lexicalizing Modalities as Discrete Tokens},
author = {Meituan LongCat Team and Bin Xiao and Chao Wang and Chengjiang Li and Chi Zhang and Chong Peng and Hang Yu and Hao Yang and Haonan Yan and Haoze Sun and Haozhe Zhao and Hong Liu and Hui Su and Jiaqi Zhang and Jiawei Wang and Jing Li and Kefeng Zhang and Manyuan Zhang and Minhao Jing and Peng Pei and Quan Chen and Taofeng Xue and Tongxin Pan and Xiaotong Li and Xiaoyang Li and Xiaoyu Zhao and Xing Hu and Xinyang Lin and Xunliang Cai and Yan Bai and Yan Feng and Yanjie Li and Yao Qiu and Yerui Sun and Yifan Lu and Ying Luo and Yipeng Mei and Yitian Chen and Yuchen Xie and Yufang Liu and Yufei Chen and Yulei Qian and Yuqi Peng and Zhihang Yu and Zhixiong Han and Changran Wang and Chen Chen and Dian Zheng and Fengjiao Chen and Ge Yang and Haowei Guo and Haozhe Wang and Hongyu Li and Huicheng Jiang and Jiale Hong and Jialv Zou and Jiamu Li and Jianping Lin and Jiaxing Liu and Jie Yang and Jing Jin and Jun Kuang and Juncheng She and Kunming Luo and Kuofeng Gao and Lin Qiu and Linsen Guo and Mianqiu Huang and Qi Li and Qian Wang and Rumei Li and Siyu Ren and Wei Wang and Wenlong He and Xi Chen and Xiao Liu and Xiaoyu Li and Xu Huang and Xuanyu Zhu and Xuezhi Cao and Yaoming Zhu and Yifei Cao and Yimeng Jia and Yizhen Jiang and Yufei Gao and Zeyang Hu and Zhenlong Yuan and Zijian Zhang and Ziwen Wang},
journal= {arXiv preprint arXiv:2603.27538},
year = {2026}
}
备注
LongCat-Next Technical Report