MEGA Hub
New

Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning

Published on Aug 13 Submitted by MEGAon Aug 17#1 Paper of the day

Authors

Do you know Kai Chen?You can claim authorship or link another user.Do you know Jifeng Ding?You can claim authorship or link another user.Do you know Ning Ding?You can claim authorship or link another user.Do you know Jiaye Ge?You can claim authorship or link another user.Do you know Lixin Gu?You can claim authorship or link another user.Do you know Yicheng Gu?You can claim authorship or link another user.Do you know Qipeng Guo?You can claim authorship or link another user.Do you know Ermo Hua?You can claim authorship or link another user.Do you know Haian Huang?You can claim authorship or link another user.Do you know Haozheng Hou?You can claim authorship or link another user.Do you know Jie Hou?You can claim authorship or link another user.Do you know Xiangyu Hong?You can claim authorship or link another user.Do you know Che Jiang?You can claim authorship or link another user.Do you know Minxi Jin?You can claim authorship or link another user.Do you know Cheng Liang?You can claim authorship or link another user.Do you know Dahua Lin?You can claim authorship or link another user.Do you know Dawei Liu?You can claim authorship or link another user.Do you know Kuikun Liu?You can claim authorship or link another user.Do you know Chengqi Lv?You can claim authorship or link another user.Do you know Haijun Lv?You can claim authorship or link another user.Do you know Han Lv?You can claim authorship or link another user.Do you know Ningsheng Ma?You can claim authorship or link another user.Do you know Biqing Qi?You can claim authorship or link another user.Do you know Jianmin Qian?You can claim authorship or link another user.Do you know Shiya Su?You can claim authorship or link another user.Do you know Youbang Sun?You can claim authorship or link another user.Do you know Huanze Tang?You can claim authorship or link another user.Do you know Zhongbo Tian?You can claim authorship or link another user.Do you know Hanjing Wang?You can claim authorship or link another user.Do you know Rui Wang?You can claim authorship or link another user.Do you know Ting Wang?You can claim authorship or link another user.Do you know Yi Wang?You can claim authorship or link another user.Do you know Baiting Wu?You can claim authorship or link another user.Do you know Jun Xu?You can claim authorship or link another user.Do you know Bowen Yang?You can claim authorship or link another user.Do you know Hui Wang?You can claim authorship or link another user.Do you know Weida Wang?You can claim authorship or link another user.Do you know Haochen Ye?You can claim authorship or link another user.Do you know Jiashuo Yu?You can claim authorship or link another user.Do you know Shan Yu?You can claim authorship or link another user.Do you know Xiaoyi Yu?You can claim authorship or link another user.Do you know Qirui Zeng?You can claim authorship or link another user.Do you know Qi Zhang?You can claim authorship or link another user.Do you know Ming Zhang?You can claim authorship or link another user.Do you know Wenwei Zhang?You can claim authorship or link another user.Do you know Bowen Zhou?You can claim authorship or link another user.Do you know Xinyu Zhou?You can claim authorship or link another user.

Abstract

We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iteratively achieve compositional reasoning. Using hidden states as cache and carrier, reasoners repeatedly query memory for required knowledge-vectors, while the knowledge is transmitted back to reasoning operators. Through this knowledge-reasoning-separation architecture, Mobius achieves better knowledge compression and reasoning efficiency. Built upon Mobius-v0 architecture: 1) Our 7B model trained-from-scratch achieves similar downstream score as a 7B Transformer baseline with 62.6% of baseline's training data. 2) Our Intern-S2-Mobius, continually-pretrained from Qwen3.5-35B, achieves similar downstream score while delivering nearly 4x end-to-end inference speedup.

Community

00