我们坚信,好的研究不该只停留在 PDF 和 GitHub Repo 里,而应该在面对面的深入交流中口口相传,找到自己的 target audience,才是对辛苦工作和开源贡献的确认。
CUA 领域的下一段新新影响力,将从这场直播开始。
📅 完整议程
上午场
The Illusion of Self-Improving Agents and What It Actually Takes to Build One
分享嘉宾:谷雨,co-founder of NeoCognition
CUA-Suite: A Unified Ecosystem for Training and Evaluating Computer-Use Agents From Pixels to Actions
Computer-use agents (CUAs) that autonomously interact with desktop interfaces hold great promise for automating complex real-world workflows. Yet progress has been bottlenecked by the lack of large-scale, high-quality training and evaluation resources.
In this talk, I present CUA-Suite, a unified ecosystem of human-annotated data and benchmarks designed to support the full development pipeline of desktop CUAs. CUA-Suite integrates three complementary resources: UI-Vision, a comprehensive desktop GUI benchmark spanning diverse professional applications;
GroundCUA, a large-scale grounding dataset with 3.5M densely annotated UI screenshots covering interactable elements as well as GroundNext models trained upon it; and VideoCUA, the largest open expert video corpus for desktop computer use, comprising 55 hours of continuous 30 fps demonstrations across 87 professional applications.
Together, these resources have accumulated over downloads ~500k on Hugging Face, reflecting broad community adoption. I will discuss key design decisions, annotation pipelines, findings in training ane evaluation, and my understanding on open research directions toward more capable and generalizable computer-use agents.
分享嘉宾:Xiangru (Edward) Jian is a PhD candidate in the Data Systems Group at the University of Waterloo, advised by Prof. M. Tamer Özsu, with research focused on LLM applications for data management spanning both structured data (relational tables, knowledge graphs) and unstructured data (images, videos).
He is also an incoming research intern at Salesforce AI Research and previously a visiting researcher at ServiceNow AI Research, where he co-led several projects on multimodal learning and computer-use agents. His work has been published at top venues including ICLR, ICML, NeurIPS, ACL, EMNLP, NAACL and ICDE.
分享嘉宾:张熙,阿里通义实验室高级算法工程师,博士毕业于中科院自动化研究所,负责多模态智能体 Mobile-Agent 等工作。在国际顶级期刊和会议 TPAMI/TIP/CVPR/ACM MM/TMM/TCSVT 等发表论文,曾获得 CCL BEST Demo、China MM BEST Paper、VCIP BEST Student Paper、中科院院长奖等,担任多个顶级会议与期刊 Reviewer,并在开源社区 Github 获得 236k star。
构建全场景计算机智能体:跨领域探索、安全护栏与开源生态
计算机智能体(Computer-Using Agents, CUA)在近年来获得了长足的发展,拥有了主动操控真实软件和操作系统的能力。无论是在包含复杂专业软件的科学计算界面,还是在移动端手机系统,CUA 有望成为不可或缺的 AI 助手与合作者。
面对截然不同的数字环境,我们系统性地探索了智能体落地的三个命题:如何在专业领域激发其探索未知的能力?如何在开放环境中约束其高危操作?以及如何在动态 Mobile 环境中合成高质量智能体轨迹数据?为此,我们本次分享围绕如下三个内容:
1、 ScienceBoard (ICLR’26) 计算机智能体助力全自动化 AI 科学家之路:为探索 CUA 成为全自动化的科学合作者的愿景,我们开创性地构建了可靠的自主科学探索环境与多领域评测基准,用以精确衡量智能体在真实科研软件中完成自动化任务的可行性与边界。