2026年6月25日（周四）下午两点半，深圳河套学院荣幸地邀请到穆罕默德·本·扎耶德人工智能大学（Mohamed bin Zayed University of Artificial Intelligence，MBZUAI）VILA实验室的沈志强助理教授，在B401教室展开“智能体系列之第七讲”，沈教授将聚焦“智能体AI时代：从以模型为中心的学习到以数据为中心的学习”做报告分享。诚挚欢迎学院师生及对该领域感兴趣的产学研界同仁参与讲座。

Title

智能体AI时代：

从以模型为中心的学习到以数据为中心的学习

From Model-Centric to Data-Centric Learning in the Agentic AI Era

About the Speaker

沈志强 助理教授

Assistant Prof. Zhiqiang Shen

沈志强现任穆罕默德·本·扎耶德人工智能大学（MBZUAI）长聘助理教授，主持该校VILA实验室。加入MBZUAI前，他曾于卡内基梅隆大学从事博士后研究，并在伊利诺伊大学厄巴纳-香槟分校与复旦大学联合完成博士阶段培养。他的研究方向涵盖机器学习、高效深度学习、智能体人工智能、计算机视觉与自然语言处理领域的基础大模型，以及以数据为中心的学习范式。他曾担任ICML、ICLR、NeurIPS等顶级机器学习会议领域主席，同时出任AACL高级领域主席，并获评NeurIPS 2025年度杰出审稿人。

Swipe up for EN version

Zhiqiang Shen is a tenure-track Assistant Professor leading the VILA Lab at Mohamed bin Zayed University of Artificial Intelligence. Before joining MBZUAI, he was a postdoctoral researcher at Carnegie Mellon University and a joint PhD training at the University of Illinois Urbana-Champaign and Fudan University. His research interests span machine learning, efficient deep learning, agentic AI, foundation models in computer vision and natural language processing, and data-centric learning. He has served as an Area Chair for major ML conferences including ICML, ICLR, and NeurIPS, as well as a Senior Area Chair for AACL. He was also selected as a NeurIPS 2025 Top Reviewer.

Abstract

过去十五年间，人工智能的发展主要由模型架构的革新所驱动，从AlexNet、ResNet到 Transformer架构，再到如今各类现代基础大模型，皆是这一发展路径的缩影。但随着模型架构日趋标准化，可规模化的训练流程也逐步成熟，未来十年人工智能的发展，将更关键地取决于数据：包括数据的采集、压缩、蒸馏、生成、筛选、安全防护与对抗攻防全流程。这一范式转变在新兴的智能体人工智能时代显得尤为重要——智能系统需要与动态变化的环境、工具、人类用户以及其他智能体持续交互，这使得数据的质量、多样性、可靠性与安全性，成为决定系统整体能力的核心要素。

本次报告将探讨人工智能研究从以模型为中心向以数据为中心的范式转型，并介绍课题组在该方向下近期完成的多项研究成果。报告涵盖多项研究内容：数据采集与生成；数据压缩与数据集蒸馏技术（在不损失模型性能的前提下降低训练成本）；知识与证据蒸馏（将大模型的能力迁移至体积更小、更便于实际部署的轻量化模型）；大语言模型混合训练数据的诊断分析；AI生成数据检测（用于解析、识别纯合成内容或经多轮AI迭代修改的素材）；以及视觉语言模型数据对抗攻击，该方向可揭示多模态基础大模型在对抗样本场景下存在的安全缺陷。依托上述系列研究，沈志强助理教授将阐明以数据为中心的技术思路如何提升人工智能系统的运行效率、鲁棒性、可信度与落地适配能力，同时也会指出该研究范式在模型评测、数据安全、负责任人工智能研发领域带来的全新挑战。

Swipe up for EN version

Over the past fifteen years, the progress of artificial intelligence has been largely driven by advances in model architectures, from AlexNet and ResNet to Transformers and modern foundation models. However, as model architectures become increasingly standardized and scalable training pipelines mature, the next decade of AI progress may depend more critically on data: how it is collected, compressed, distilled, synthesized, filtered, protected, and attacked. This shift is particularly important in the emerging era of agentic AI, where intelligent systems must interact with dynamic environments, tools, users, and other agents, making data quality, diversity, reliability, and safety central to their capabilities.

In this talk, Assistant Prof. Zhiqiang Shen will discuss the transition from model-centric to data-centric learning and present several recent works from our group along this direction. The talk will cover data collection/generation, data compression and dataset distillation for reducing training cost while preserving model performance; knowledge and evidence distillation for transferring capabilities from large models to smaller and more deployable models; diagnosing data mixture for large language models; AI-generated data detection for understanding and identifying synthetic or progressively AI-edited content; and visual-language model data attacks, which reveal vulnerabilities in multimodal foundation models under adversarial data settings. Through these works, Assistant Prof. Zhiqiang Shen will highlight how data-centric methods can improve efficiency, robustness, trustworthiness, and deployability of AI systems, while also raising new challenges in evaluation, security, and responsible AI development.

Host

姚天峻 助理教授

Assistant Prof. Tianjun Yao

Date & Time

2026年6月25日（星期四）

下午14:30-16:00

June 25, 2026, Thursday,

14:30pm-16:00pm

Venue

【公众号定位插件】

深圳河套学院B401教室

（深圳市福田区福保街道红棉路6号，地图导航“深圳河套学院-南门”）

Room B401, Shenzhen Loop Area Institute

(6 Hongmian Rd, Fubao Sub-Street, Futian District, Shenzhen, navigate to "Shenzhen Loop Area Institute (South Gate)" on the map.)

Online link

扫码加入会议
Join the Meeting Online

腾讯会议号：948-1135-4043
Tencent Meeting：948-1135-4043