Systems at the Crossroad of Agents & Infrastructure (MLSys ’26 Digest Talk)

Abstract

As Large Language Models transition from isolated text generators to autonomous, multi-step agentic systems, traditional serving frameworks face unprecedented performance bottlenecks. In this talk, I will share my first-hand takeaways from MLSys 2026, presenting a structured digest of the most exciting research and ideas I encountered on the ground. I will organize the content into three main areas: Agentic AI & Systems, Systems for LLMs, and Compilers & Kernels. Finally, I will offer my personal thoughts on the future trends and directions where agent-native systems infrastructure is heading, followed by an open discussion with the audience to share perspectives and insights.

Bio

Yiyu Liu is a first-year PhD student in the Harvard Systems Group, advised by Professors Minlan Yu and Juncheng Yang. His research explores systems for machine learning, with a focus on improving the management of KV cache. Prior to joining Harvard, he was a student intern at the University of Washington under the guidance of Professor Baris Kasikci.