Vista Equipo: :: Library Catalog

Guardado en:

Detalles Bibliográficos
Autores principales:	Ling Team, Shen, Anqi, Li, Baihui, Hu, Bin, Jing, Bin, Chen, Cai, Huang, Chao, Zhang, Chao, Yang, Chaokun, Lin, Cheng, Wen, Chengyao, Li, Congqi, Zhao, Deng, Yuan, Dingbo, You, Donghai, Mao, Fagui, Meng, Fanzhuang, Xu, Feng, Li, Guojie, Wang, Guowei, Dai, Hao, Zheng, Haonan, Liu, Hong, Guo, Jia, Liu, Jiaming, Liu, Jian, Fu, Jianhao, Shi, Jiannan, Wang, Jianwen, Lai, Jianxin, Yang, Jin, Mei, Jun, Zhou, Jun, Zhao, Junbo, Zhao, Junping, Xu, Kuan, Su, Le, Chen, Lei, Tang, Li, Jiang, Liang, Fu, Liangcheng, Xu, Lianhao, Shi, Linfeng, Liao, Lisha, Zheng, Longfei, Li, Meng, Chen, Mingchun, Zuo, Qi, Cheng, Qiang, Cao, Qianggang, Shi, Qitao, Guo, Quanrui, Zhu, Senlin, Wang, Shaofei, Zheng, Shaomian, Li, Shuaicheng, Gu, Shuwei, Chen, Siba, Wu, Tao, Zhang, Tao, Zhang, Tianyu, Zhou, Tianyu, Bie, Tiwei, Yang, Tongkai, Hong, Wang, Ren, Wang, Chen, Weihua, Yu, Wenbo, Zheng, Wengang, Wang, Xiangchun, Yan, Xiaodong, Wan, Xiaopei, Zhao, Xin, Kong, Xinyu, Tang, Xinyu, Han, Xudong, Wang, Xudong, Yang, Xuemin, Hu, Xueyu, Zhang, Yalin, Sun, Yan, Shan, Yicheng, Wang, Yilong, Xu, Yingying, Liu, Yongkang, Guo, Yongzhen, Wang, Yuanyuan, Yan, Yuchen, Wang, Yuefan, Guo, Yuhong, Li, Zehuan, Xu, Zhankai, Li, Zhe, Zhang, Zhenduo, Gui, Zhengke, Pan, Zhenxuan, Huang, Zhenyu, Lan, Zhenzhong, Ding, Zhiqiang, Zhang, Zhiqiang, Li, Zhixun, Liu, Zhizhen, Wang, Zihao, Wen, Zujie
Formato:	Preprint
Publicado:	2025
Materias:	Computation and Language Artificial Intelligence
Acceso en línea:	https://arxiv.org/abs/2510.18855
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

_version_	1866912669635706880
author	Ling Team Shen, Anqi Li, Baihui Hu, Bin Jing, Bin Chen, Cai Huang, Chao Zhang, Chao Yang, Chaokun Lin, Cheng Wen, Chengyao Li, Congqi Zhao, Deng Yuan, Dingbo You, Donghai Mao, Fagui Meng, Fanzhuang Xu, Feng Li, Guojie Wang, Guowei Dai, Hao Zheng, Haonan Liu, Hong Guo, Jia Liu, Jiaming Liu, Jian Fu, Jianhao Shi, Jiannan Wang, Jianwen Lai, Jianxin Yang, Jin Mei, Jun Zhou, Jun Zhao, Junbo Zhao, Junping Xu, Kuan Su, Le Chen, Lei Tang, Li Jiang, Liang Fu, Liangcheng Xu, Lianhao Shi, Linfeng Liao, Lisha Zheng, Longfei Li, Meng Chen, Mingchun Zuo, Qi Cheng, Qiang Cao, Qianggang Shi, Qitao Guo, Quanrui Zhu, Senlin Wang, Shaofei Zheng, Shaomian Li, Shuaicheng Gu, Shuwei Chen, Siba Wu, Tao Zhang, Tao Zhang, Tianyu Zhou, Tianyu Bie, Tiwei Yang, Tongkai Hong, Wang Ren, Wang Chen, Weihua Yu, Wenbo Zheng, Wengang Wang, Xiangchun Yan, Xiaodong Wan, Xiaopei Zhao, Xin Kong, Xinyu Tang, Xinyu Han, Xudong Wang, Xudong Yang, Xuemin Hu, Xueyu Zhang, Yalin Sun, Yan Shan, Yicheng Wang, Yilong Xu, Yingying Liu, Yongkang Guo, Yongzhen Wang, Yuanyuan Yan, Yuchen Wang, Yuefan Guo, Yuhong Li, Zehuan Xu, Zhankai Li, Zhe Zhang, Zhenduo Gui, Zhengke Pan, Zhenxuan Huang, Zhenyu Lan, Zhenzhong Ding, Zhiqiang Zhang, Zhiqiang Li, Zhixun Liu, Zhizhen Wang, Zihao Wen, Zujie
author_facet	Ling Team Shen, Anqi Li, Baihui Hu, Bin Jing, Bin Chen, Cai Huang, Chao Zhang, Chao Yang, Chaokun Lin, Cheng Wen, Chengyao Li, Congqi Zhao, Deng Yuan, Dingbo You, Donghai Mao, Fagui Meng, Fanzhuang Xu, Feng Li, Guojie Wang, Guowei Dai, Hao Zheng, Haonan Liu, Hong Guo, Jia Liu, Jiaming Liu, Jian Fu, Jianhao Shi, Jiannan Wang, Jianwen Lai, Jianxin Yang, Jin Mei, Jun Zhou, Jun Zhao, Junbo Zhao, Junping Xu, Kuan Su, Le Chen, Lei Tang, Li Jiang, Liang Fu, Liangcheng Xu, Lianhao Shi, Linfeng Liao, Lisha Zheng, Longfei Li, Meng Chen, Mingchun Zuo, Qi Cheng, Qiang Cao, Qianggang Shi, Qitao Guo, Quanrui Zhu, Senlin Wang, Shaofei Zheng, Shaomian Li, Shuaicheng Gu, Shuwei Chen, Siba Wu, Tao Zhang, Tao Zhang, Tianyu Zhou, Tianyu Bie, Tiwei Yang, Tongkai Hong, Wang Ren, Wang Chen, Weihua Yu, Wenbo Zheng, Wengang Wang, Xiangchun Yan, Xiaodong Wan, Xiaopei Zhao, Xin Kong, Xinyu Tang, Xinyu Han, Xudong Wang, Xudong Yang, Xuemin Hu, Xueyu Zhang, Yalin Sun, Yan Shan, Yicheng Wang, Yilong Xu, Yingying Liu, Yongkang Guo, Yongzhen Wang, Yuanyuan Yan, Yuchen Wang, Yuefan Guo, Yuhong Li, Zehuan Xu, Zhankai Li, Zhe Zhang, Zhenduo Gui, Zhengke Pan, Zhenxuan Huang, Zhenyu Lan, Zhenzhong Ding, Zhiqiang Zhang, Zhiqiang Li, Zhixun Liu, Zhizhen Wang, Zihao Wen, Zujie
contents	We present Ring-1T, the first open-source, state-of-the-art thinking model with a trillion-scale parameter. It features 1 trillion total parameters and activates approximately 50 billion per token. Training such models at a trillion-parameter scale introduces unprecedented challenges, including train-inference misalignment, inefficiencies in rollout processing, and bottlenecks in the RL system. To address these, we pioneer three interconnected innovations: (1) IcePop stabilizes RL training via token-level discrepancy masking and clipping, resolving instability from training-inference mismatches; (2) C3PO++ improves resource utilization for long rollouts under a token budget by dynamically partitioning them, thereby obtaining high time efficiency; and (3) ASystem, a high-performance RL framework designed to overcome the systemic bottlenecks that impede trillion-parameter model training. Ring-1T delivers breakthrough results across critical benchmarks: 93.4 on AIME-2025, 86.72 on HMMT-2025, 2088 on CodeForces, and 55.94 on ARC-AGI-1. Notably, it attains a silver medal-level result on the IMO-2025, underscoring its exceptional reasoning capabilities. By releasing the complete 1T parameter MoE model to the community, we provide the research community with direct access to cutting-edge reasoning capabilities. This contribution marks a significant milestone in democratizing large-scale reasoning intelligence and establishes a new baseline for open-source model performance.
format	Preprint
id	arxiv_https___arxiv_org_abs_2510_18855
institution	arXiv
publishDate	2025
record_format	arxiv
spellingShingle	Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model Ling Team Shen, Anqi Li, Baihui Hu, Bin Jing, Bin Chen, Cai Huang, Chao Zhang, Chao Yang, Chaokun Lin, Cheng Wen, Chengyao Li, Congqi Zhao, Deng Yuan, Dingbo You, Donghai Mao, Fagui Meng, Fanzhuang Xu, Feng Li, Guojie Wang, Guowei Dai, Hao Zheng, Haonan Liu, Hong Guo, Jia Liu, Jiaming Liu, Jian Fu, Jianhao Shi, Jiannan Wang, Jianwen Lai, Jianxin Yang, Jin Mei, Jun Zhou, Jun Zhao, Junbo Zhao, Junping Xu, Kuan Su, Le Chen, Lei Tang, Li Jiang, Liang Fu, Liangcheng Xu, Lianhao Shi, Linfeng Liao, Lisha Zheng, Longfei Li, Meng Chen, Mingchun Zuo, Qi Cheng, Qiang Cao, Qianggang Shi, Qitao Guo, Quanrui Zhu, Senlin Wang, Shaofei Zheng, Shaomian Li, Shuaicheng Gu, Shuwei Chen, Siba Wu, Tao Zhang, Tao Zhang, Tianyu Zhou, Tianyu Bie, Tiwei Yang, Tongkai Hong, Wang Ren, Wang Chen, Weihua Yu, Wenbo Zheng, Wengang Wang, Xiangchun Yan, Xiaodong Wan, Xiaopei Zhao, Xin Kong, Xinyu Tang, Xinyu Han, Xudong Wang, Xudong Yang, Xuemin Hu, Xueyu Zhang, Yalin Sun, Yan Shan, Yicheng Wang, Yilong Xu, Yingying Liu, Yongkang Guo, Yongzhen Wang, Yuanyuan Yan, Yuchen Wang, Yuefan Guo, Yuhong Li, Zehuan Xu, Zhankai Li, Zhe Zhang, Zhenduo Gui, Zhengke Pan, Zhenxuan Huang, Zhenyu Lan, Zhenzhong Ding, Zhiqiang Zhang, Zhiqiang Li, Zhixun Liu, Zhizhen Wang, Zihao Wen, Zujie Computation and Language Artificial Intelligence We present Ring-1T, the first open-source, state-of-the-art thinking model with a trillion-scale parameter. It features 1 trillion total parameters and activates approximately 50 billion per token. Training such models at a trillion-parameter scale introduces unprecedented challenges, including train-inference misalignment, inefficiencies in rollout processing, and bottlenecks in the RL system. To address these, we pioneer three interconnected innovations: (1) IcePop stabilizes RL training via token-level discrepancy masking and clipping, resolving instability from training-inference mismatches; (2) C3PO++ improves resource utilization for long rollouts under a token budget by dynamically partitioning them, thereby obtaining high time efficiency; and (3) ASystem, a high-performance RL framework designed to overcome the systemic bottlenecks that impede trillion-parameter model training. Ring-1T delivers breakthrough results across critical benchmarks: 93.4 on AIME-2025, 86.72 on HMMT-2025, 2088 on CodeForces, and 55.94 on ARC-AGI-1. Notably, it attains a silver medal-level result on the IMO-2025, underscoring its exceptional reasoning capabilities. By releasing the complete 1T parameter MoE model to the community, we provide the research community with direct access to cutting-edge reasoning capabilities. This contribution marks a significant milestone in democratizing large-scale reasoning intelligence and establishes a new baseline for open-source model performance.
title	Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model
topic	Computation and Language Artificial Intelligence
url	https://arxiv.org/abs/2510.18855

Ejemplares similares