Xiaomi MiMo Team and Peking University Develop ARL-Tangram System, Cutting Computing Costs by 71.2%

Led by Luo Fuli, head of Xiaomi's MiMo large model team, the team collaborated with Peking University to develop ARL-Tangram, a resource management system for AI agents. The system uses unified action-level formulas and elastic scheduling algorithms to meet heterogeneous resource constraints and significantly improve task efficiency.

In real-world reinforcement learning task tests, ARL-Tangram increased average task completion time by 4.3x, reduced training step duration by 1.5x, and cut external resource costs by up to 71.2%.

This is Luo Fuli's second technical breakthrough at Xiaomi. Last October, she co-published a paper on MoE and reinforcement learning with Peking University, marking her first achievement at Xiaomi.