EGG: An Expert-Guided Agent Framework for Kernel Generation

A new framework called EGG aims to automate the creation of high-performance GPU kernels for large language models, a task that currently depends on manual work by domain experts, according to a preprint submitted on 25 Jun 2026 [1]. The framework, detailed in a paper on the arXiv preprint server, uses expert optimization principles to guide the decisions of large language models (LLMs) during kernel generation [1]. The work addresses a bottleneck in artificial intelligence, where the computational cost of training and running neural networks has grown exponentially. Training these networks is a compute-intensive process accelerated by graphics processing units (GPUs) and large datasets [3]. Transformer architectures, which introduced attention mechanisms, now form the basis of large language models [3]. EGG decomposes the kernel generation process into two hierarchical stages, mirroring the workflow of human experts [1]. The first stage, algorithmic structure design, establishes a high-quality computational foundation. The second stage, hardware-specific tuning, performs targeted adjustments through parallel mapping, tensor tiling, and memory optimization [1]. A stage-aware multi-agent collaboration mechanism manages context between and within these stages to maintain stable optimization trajectories [1]. In experiments on the KernelBench benchmark and real-world workloads, EGG achieved a 2.13x average speedup over PyTorch [1]. The paper states the framework outperformed existing agent-based and reinforcement learning-based approaches [1]. The preprint was posted on arXiv, an open-access repository for electronic preprints that are moderated but not peer reviewed [7]. As of November 2024, the submission rate to arXiv was about 24,000 articles per month [7].