Yonsei University Team Wins Top Prize at MLSys 2026 Google Graph Scheduling Contest

A team from Yonsei University's School of Electrical and Electronic Engineering has achieved first place in the prestigious MLSys 2026 Google Graph Scheduling Contest. The winning team, comprising graduate students Yeom Ho-yoon and Lee Chan, under the guidance of Professor Kim Han-jun, secured the top spot at the international competition held in the United States in May 2026.

The contest, organized by Google as part of the Conference on Machine Learning and Systems (MLSys), one of the leading academic conferences in machine learning systems, challenges participants to optimize the scheduling of computation graphs to minimize execution costs. The specific track the Yonsei team excelled in, 'Agent Reasoning (Track B),' evaluated the ability of Large Language Model (LLM)-based agents to solve complex scheduling problems through reasoning.

Their winning submission, titled 'Pre-detecting Structural Opportunities for Gemini-Driven DAG Scheduling,' introduced a novel technique. This method enables LLM-based agents to proactively identify and utilize structural optimization opportunities within Directed Acyclic Graphs (DAGs). Computation graph scheduling is a notoriously difficult problem, requiring simultaneous consideration of operator fusion and memory input/output costs. The research team achieved superior performance by implementing a distinct division of labor, clearly separating the reasoning domain handled by LLMs from the search domain managed by a deterministic engine.

This achievement is particularly significant as it effectively integrates the reasoning capabilities of LLMs into system optimization problems. It demonstrates the expanding potential role of LLMs in next-generation compilers and machine learning systems, validating the practical application of AI-driven software optimization technologies. The success underscores Yonsei University's School of Electrical and Electronic Engineering's international competitiveness in AI-based system software and next-generation compiler technologies. The research team plans to continue expanding their work on fusing LLMs with compiler and system technologies for next-generation software optimization.