夜雨聆风
SEMINAR INFORMATION
Topic
Towards Reliable and Scalable Autonomous Agents
Time
10:00-11:00 AM
Date
Apr. 21st, 2026
Venue
TA 415
Zoom meeting
https://cuhk-edu-cn.zoom.us/j/6682399580?omn=96659800162
Meeting ID: 6682399580
Speaker
Dr. ZENG Xingshan
(Huawei Hong Kong Research Center)
Host
Prof. HU Junjie
(The Chinese University of Hong Kong, Shenzhen)
ABOUT THE SPEAKER
Dr. ZENG Xingshan
Huawei Hong Kong Research Center
Dr. ZENG Xingshan is a Senior Researcher at the Huawei Hong Kong Research Center. His research focuses on Natural Language Processing, Large Language Models, and AI agents, with the long-term goal of developing principled frameworks for agent systems that are reliable, scalable, and well-aligned for real-world deployment. He received his B.S. from USTC and his Ph.D. from CUHK.
He has published over 40 papers in top-tier conferences and journals in natural language processing and machine learning, and his work has received the Outstanding Paper Award at ACL 2024. He led the development of the open-source ToolACE series, which has accumulated over 600K downloads on Hugging Face. Related methods have also been adopted in Huawei Device and Cloud Services. He has served as an Area Chair for major conferences including ACL, EMNLP, and NAACL.
ABSTRACT
Large Language Models (LLMs) are emerging as a promising foundation for autonomous agents. However, a substantial gap remains between basic tool use and robust, long-horizon reasoning in real-world environments.
In this talk, the speaker presents a systematic approach to advancing LLM tool-use capabilities, with a focus on scalable data generation and reliable inference-time decision-making. The speaker begins by identifying the key characteristics of high-quality agentic data, which motivate the design of the ToolACE framework: a series of methods for the scalable, controllable, and robust synthesis of agentic data, together with corresponding training strategies that improve data utilization. Building on this foundation, the speaker then explores techniques for enhancing agent reliability through inference-time scaling, enabling more robust behavior in complex, real-world settings. Taken together, this line of work suggests a data-centric path toward reliable and scalable LLM-based agents for real-world applications.
All of you are warmly welcome!
