LLM agents have become capable kernel engineers on mainstream GPU stacks, yet they struggle with emerging accelerators such as Ascend NPUs because their programming models are scarcely represented in pre‑training data. Existing NPU agents mainly rely on experience: they store their own generation and optimization attempts in memory and reuse them for later tasks. Experience alone, however, is insufficient for kernel generation. An agent can only remember what it has tried and rarely explores unknown mechanisms; when a new operator requires a different implementation, the memory offers little assistance, and even familiar operators can present unseen situations during optimization.
To overcome this limitation we introduce ASCENDEVOLVE. Every step of its evolutionary kernel search is grounded in a large external knowledge base built from platform documentation, toolkit headers, and operator repositories. Because this corpus far exceeds any context window, ASCENDEVOLVE first indexes it and then routes each task to a small, relevant slice of evidence instead of loading the whole corpus. The retrieved knowledge seeds diverse implementation strategies, guides repair when a candidate fails, and constrains how strong candidates are recombined.
On 53 Ascend C operators from CANN‑Bench spanning four difficulty levels, ASCENDEVOLVE achieves an overall score of 71.39, improving over Reflexion and OpenEvolve by 11.6% and 17.2% respectively.
Review: By converting platform‑level documentation into a searchable knowledge source, ASCENDEVOLVE blends experience with explicit knowledge, enabling rapid adaptation to new operator mechanisms and demonstrating the promise of knowledge‑guided evolution in hardware‑specific compilation.