时间：2026年9月19日（周六）下午2:30-5:30

地点：岳麓山校区计算机楼313

报告一（2:30-3:30）

报告人：孙贺教授，中国科台湾av女优深圳先进技术研究院

题目：Multi-Modal Graph Clustering with Theoretical Guarantee

报告摘要：This talk studies clustering algorithms for multi-modal graphs, in which each modality may have a different number of clusters caused by the unique information of each modality, and we assume that data points in the same cluster if and only if they appear in the same optimal cluster for every single modality. We prove that the modality-consensus clusters can be well approximated by performing k-means on concatenated modality-specific spectral embeddings. Experimental results on both synthetic and real-world datasets further verify the effectiveness of our designed algorithm.

The talk is based on joint work with Zhenglai Li (Center for Algorithms and Learning Theory, SIAT@CAS).

报告人简介：Professor He Sun is the Director of Center for Algorithms and Learning Theory, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, and the Associate Director of the State Key Laboratory of Biomedical Imaging Science and System with over 150 researchers. He received PhD from Fudan University in 2010 and worked at the MPI for Informatics, UC Berkeley, University of Bristol, and University of Edinburgh. He is a recipient of the Chinese High-Level Talent Recruitment Program for Oversea Experts and is a fellow selected under the CAS 100 Talents Program (Category A). He has served as an area chair and PC member of leading international conferences like ICML (2026), NeurIPS (2026), ICLR (2027), and STOC (2026), and as a regular reviewer for research councils including the ERC, UKRI and ISF.

报告二（3:30-4:30）腾讯会议：212-809-720

报告人：严骏驰教授，上海交通大学

题目：大模型新架构探索：JTok与xHC

报告摘要：本报告介绍团队在大模型新架构方面的近期探索和初步进展，包括JTok 和xHC两部分。两者出于同一个判断：除宽度与深度之外，模型能力还可以在几乎不增加算力的前提下继续大幅提升。JTok 沿token 维度扩展，让每个 token 在每一层携带一条可学习的记忆向量，以一次查表加一次轻量调制的旁路方式接入，主干完全不动；在 61B 总参数（17B 底座 + 44B 记忆参数）、570B tokens 规模上验证，达到同等性能所需训练算力节省约 35%，训练吞吐仅损失 6.8%，推理零额外显存开销，且可为已训成的存量模型直接加装。xHC 沿残差流维度扩展，针对 Hyper-Connections 家族长期止步于 N=4 的现象，定位出写回信息供给不足与残差映射生成代价随 N 立方增长两个瓶颈，分别以时序特征增强与稀疏残差流（稠密读取、稀疏写入）加以解决，首次将扩展率推进到 N=16；在 18B / 28B MoE 上，十二项下游评测平均分较 DeepSeek 的 mHC 提升 4.0 与 3.1 分，训练 FLOPs 仅增加 4.1% 与 3.0%，规模定律上达到同等性能所需训练算力较原始 Transformer 节省约 33%。报告同时指出，这类结构改动的真正约束在于访存带宽而非 FLOPs，并给出 xHC-Flash 的工程方案，将每个 sublayer 的访存量由 73.5C 压回至 40C，接近 mHC 的水平而性能基本不损失。最后我们对两条轴做逐维度对照，并讨论其正交叠加的潜力。

报告人简介：上海交通大学人工智能台湾av女优教授，基金委A/B青年项目及重大研究计划、科技部重大专项、教育部U40及学科突破先导单方向项目负责人，国家级规划教材主编、教育部深度学习虚拟教研室负责人。曾任IBM研究院（首席）研究员和亚马逊高级顾问10年。任ACM-Multimedia 2026大会程序共同主席、ICML理事、CVPR/ICML/NeurIPS/ICLR/AAAI等CCF-A类会议(高级)领域主席、JMLR/TMLR/PRJ/IEEE TPAMI/TNNLS/TEVC/CIM期刊(高级)编委、ACM TOPML、ML4CSE创始高级编委。获省自然科学一等奖、AAAI21/IJCAI23最具影响力论文榜首、CVPR24/IROS25最佳论文候选、ACL25杰出论文、爱思唯尔高被引学者、学术引用近5万次。入选IEEE Xplore封面作者，国际基础科学大会ICBS一小时大会报告人。任国际科学智能联盟副秘书长、IAPR/IET Fellow(兼评审官)，IEEE-CS AI'10 to Watch、IEEE CIS Outstanding Early Career Award获得者。

报告三（4:30-5:30）

报告人：张宇昊，上海交通大学

题目：Constant-Competitive Online Selfish Load Balancing

报告摘要：We study online load balancing on selfish related machines, where the goal is to design online mechanisms with good competitive ratios under the constraint of machine-side truthfulness: each machine has a private speed and should be incentivized to report its true speed. The offline version of this problem on related machines was first studied by Archer and Tardos [FOCS~2001], who gave a constant-approximation randomized truthful-in-expectation mechanism. This result was later improved to a PTAS by Christodoulou and Kovács [SICOMP~2013] and Epstein, Levin, and Stee. [MOR~2016].

In contrast, whether a constant competitive ratio can be achieved in the online setting has remained open for decades. Such a guarantee was known only for the special case of two machines, due to Auletta, Prisco, Penna, and Persiano [TCS~2009]. For the general case, the best-known competitive ratio was still $O(\log m)$, due to Wang, Xu, and Zhang [EC~2026]. Our main contribution is to achieve a constant competitive ratio for this problem via a randomized truthful-in-expectation online mechanism.

报告人简介：张宇昊，上海交通大学长聘教轨副教授，国家海外高层次引进青年人才。博士毕业于香港大学，本科毕业于浙江大学。研究方向为理论计算机，包括在线算法设计，近似算法设计等。论文发表于理论计算机顶级会议期刊，如STOC，FOCS，SODA，JACM，SICOMP。曾主持国自然青年基金一项，面上基金一项，获得中国运筹学会青年科技奖提名奖。