daVinci-kernel: Co-Evolving Skill Selection, Summarization, and Utilization via RL for GPU Kernel Optimization

A new reinforcement learning framework called daVinci-kernel jointly trains three agents sharing a single large language model backbone to optimize GPU kernel execution efficiency, according to a paper posted to arXiv on June 15, 2026 [1][2]. The system co-evolves skill selection, summarization, and utilization through a dynamically updated skill library [2]. The framework consists of a Skill Selection Agent that retrieves relevant optimization techniques using BM25 and LLM reranking, a Policy Agent that generates multi-turn CUDA or Triton kernels conditioned on those skills, and a Skill Summary Agent that distills successful rollouts into reusable skills [2]. New skills are added to the library only after execution-based verification confirms reproducible speedups [2]. All three agents are initialized through a structured supervised fine-tuning cold start on diversity-filtered data, then optimized end-to-end with multi-turn REINFORCE and per-agent advantage estimation [2]. On the KernelBench benchmark, daVinci-kernel-14B achieved 37.2 percent on Level 1, 70.6 percent on Level 2, and 32.2 percent on Level 3 under the Fast₁ threshold [1][2]. The paper states the model outperforms the strongest prior RL-trained model, Dr.Kernel-14B [2]. Large language models of this scale contain billions of parameters and are trained with self-supervised learning on vast text corpora [8]. The paper appeared on arXiv, an open-access repository of electronic preprints that has hosted scientific papers since August 1991 and now receives approximately 24,000 submissions per month [6]. Submissions are moderated but not peer reviewed [6]. The repository passed the two-million-article milestone by the end of 2021 [6]. The daVinci-kernel work targets a setting where functional correctness of GPU kernels is assumed and the objective is purely execution efficiency [2]. The joint training approach couples skill discovery with skill exploitation, allowing the system to build and refine a library of optimization techniques over successive rollouts [2].