According to a new arXiv paper, LLM agents have become capable kernel engineers on mainstream GPU stacks but stumble on emerging accelerators. On Ascend NPUs, the programming model is barely represented in LLM pretraining data, leaving agents without relevant experience to draw on.
The paper, Beyond Experience: Knowledge-Guided Evolution for NPU Kernel Generation, answers with a knowledge-guided evolution approach to NPU kernel generation. The available abstract does not describe how that evolution works; the method is identified in the paper's title. The abstract also begins to situate the work among recent NPU agents, but the provided text stops mid-sentence before that comparison is spelled out.