Published at ICML 2026, EvoKernel addresses cold-start code generation in data-scarce accelerator ecosystems. It casts NPU kernel synthesis as a memory-based reinforcement-learning process, learns stage-specific experience values for initial drafting and later latency refinement, and shares useful experience across tasks. The resulting agent accumulates practical optimization knowledge online instead of relying on expensive domain-specific fine-tuning.
Value-guided experience reuse turns sparse NPU feedback into a continual drafting-and-refinement loop.