News in 2025
-
09/2025: 联合360研究团队发布 TinyR1-32B 模型——以仅20k数据实现安全性能里程碑式突破:超同尺寸 Qwen3-32B 25分、超 DeepSeek-R1-0528 17分,推理性能达后者的93%;提出 Control Token 动态安全控制,打破安全与可用二选一困境,同步推出轻量安全模型 TinyR1-Safety-8B。[模型链接-32B] [模型链接-8B] [新闻链接]
-
08/2025: 研究团队发布 iFairy ——首个将全部参数约束在 {±1, ±i} 的 2-bit 复数量化大模型,实现无乘法、仅加法的高效推理与约1/8存储压缩,并在语言建模与下游任务上超越同尺寸全精度 LLaMA;相关论文及代码全面开源,进一步为低功耗与边缘部署提供新路径。[GitHub链接] [模型链接] [论文链接] [新闻链接] [报告视频]
-
08/2025: 研究团队发布 Proof2Hybrid,为首个全自动数学基准合成架构。该架构以多大语言模型流水线为核心,实现数学评测集的自动、高效构建;并据此推出首个专注于代数几何领域的基准 AlgGeoTest。[论文链接] [数据集链接]
-
07/2025: 受邀担任 CCF Computility 2025 “智算集群创新与实践” 论坛主席,论坛将于2025年7月在甘肃兰州举行,聚焦大模型训练推理中的智算基础设施发展与挑战。[报名链接]
-
06/2025: 研究团队发布 ScholarSearch 数据集,为首个专注于评估大语言模型在学术研究中复杂信息检索能力的基准。结果显示,即便是最强搜索增强模型 GPT-4o-search-preview 的准确率也仅为 18.83%。[论文链接] [数据集链接]
-
05/2025: 研究团队发布 FairyR1-32B 模型,采用“分合蒸馏”思路,实现在仅 5% 参数量下,数学与代码能力超越 DeepSeek-R1 满血版,充分探索在有限资源下实现高性能模型的可行性。[模型链接]
-
02/2025: 联合360研究团队发布 Tiny-R1-32B-Preview 模型,以仅 5% 参数逼近 DeepSeek-R1-671B 的性能,在数学、代码、科学三大领域全面领先 Deepseek 70B模型。[模型链接]
News in 2024
- 06/2024: Congratulations to Ruixin Wang for being awarded the Outstanding Graduate of Beijing City!
- 06/2024: Congratulations to Yuhan Wu for being awarded the President's Scholarship of PKU!
- 06/2024: Congratulations to Fenghao Dong for being admitted to the PhD program at CMU!
- 06/2024: Congratulations to Xiuqi Zheng for being admitted to the Master's program at CMU!
- 06/2024: Congratulations to Jianan Ji for being admitted to the Master's program at CMU!
- 06/2024: Congratulations to Siyuan Dong for being admitted to the PhD program at University of Michigan!
- 06/2024: Congratulations to Wei Zhou for being admitted to the Master's program at USC!
- 06/2024: Congratulations to Zhouran Shi for being admitted to the Master's program at HKUST!