Tag
This paper presents CAKE, a compiler-agent co-design framework that lets AI agents author a hardware-explicit IR for GPU kernels, achieving significant speedups over baselines across various workload families.