Caiwen Ding of University of Minnesota

Modern computing is often optimized in isolated layers, with separate tools for interfaces, software, and hardware. In this talk, we will start with expertise-driven design acceleration and expand to automated optimization. We show that large language model (LLM) agents can iteratively generate, test, and improve GPU kernels by reading hardware feedback, from single kernels to full end-to-end GPU programs. We then broaden the picture by showing how agents can more accurately ground interface elements and carry out computer interaction step by step,
giving them the ability to operate real software, hardware environments, and applications.

Biography:

Caiwen Ding is an Associate Professor in the Department of Computer Science & Engineering at the University of Minnesota Twin Cities, which he joined in August 2024. He had been an Assistant Professor in the School of Computing at the University of Connecticut from 2019-2024. He received his Ph.D. from Northeastern University, Boston in 2019. His research interests span algorithm–system co-design for machine learning and artificial intelligence, privacy-preserving machine learning, machine learning for electronic design automation (EDA). Dr. Ding’s research  has been published in leading venues across systems, architecture, and AI/ML—including DAC, ICCAD, ASPLOS, ISCA, MICRO, HPCA, SC, FPGA, MLSys, NeurIPS, ICML, CVPR, ACL, EMNLP, and IJCAI. He is a recipient of the NSF CAREER Award, Amazon Research Award, and Cisco Research Award. He has received multiple paper distinctions, including the Best Paper Award at the 2025 IEEE International Conference on LLM-Aided Design (ICLAD), the Best Paper Award at the 2023 AAAI DCAA workshop, Outstanding Student Paper at HPEC 2023, Best Paper Award
Nominations at DATE 2018 and DATE 2021, and a publicity paper at DAC 2022.

Email:

Address:Dept. Computer Science & Eng., University of Minnesota, Minneapolis, Minnesota, United States, 55455