MKEvolve Cuts LLM Token Usage 35% on Accelerator Kernel Generation via Modular Multi-Agent Decomposition

arXiv 2607.20501·low signal

Writing correct and performant kernels for hardware accelerators remains a scaling bottleneck for ML workloads even as general LLM code generation improves. MKEvolve decomposes the problem across a modular multi-agent framework rather than a monolithic generate-and-check loop, improving both correctness and achieved speedup while reducing LLM token usage by up to 35%. Read alongside JAXBench and CANN Bench announced the same day, it points at kernel authoring as the current concentration point for agentic coding research — and token cost, not just pass rate, as a reported axis.