Saved in:
Bibliographic Details
Main Authors: Chen, Terry, Ye, Zhifan, Xu, Bing, Ye, Zihao, Liu, Timmy, Hassani, Ali, Chen, Tianqi, Kerr, Andrew, Wu, Haicheng, Xu, Yang, Chen, Yu-Jung, Chen, Hanfeng, Kane, Aditya, Krashinsky, Ronny, Liu, Ming-Yu, Grover, Vinod, Ceze, Luis, Bringmann, Roger, Tran, John, Liu, Wei, Xie, Fung, Lightstone, Michael, Shi, Humphrey
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.24517
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912982299049984
author Chen, Terry
Ye, Zhifan
Xu, Bing
Ye, Zihao
Liu, Timmy
Hassani, Ali
Chen, Tianqi
Kerr, Andrew
Wu, Haicheng
Xu, Yang
Chen, Yu-Jung
Chen, Hanfeng
Kane, Aditya
Krashinsky, Ronny
Liu, Ming-Yu
Grover, Vinod
Ceze, Luis
Bringmann, Roger
Tran, John
Liu, Wei
Xie, Fung
Lightstone, Michael
Shi, Humphrey
author_facet Chen, Terry
Ye, Zhifan
Xu, Bing
Ye, Zihao
Liu, Timmy
Hassani, Ali
Chen, Tianqi
Kerr, Andrew
Wu, Haicheng
Xu, Yang
Chen, Yu-Jung
Chen, Hanfeng
Kane, Aditya
Krashinsky, Ronny
Liu, Ming-Yu
Grover, Vinod
Ceze, Luis
Bringmann, Roger
Tran, John
Liu, Wei
Xie, Fung
Lightstone, Michael
Shi, Humphrey
contents Agentic Variation Operators (AVO) are a new family of evolutionary variation operators that replace the fixed mutation, crossover, and hand-designed heuristics of classical evolutionary search with autonomous coding agents. Rather than confining a language model to candidate generation within a prescribed pipeline, AVO instantiates variation as a self-directed agent loop that can consult the current lineage, a domain-specific knowledge base, and execution feedback to propose, repair, critique, and verify implementation edits. We evaluate AVO on attention, among the most aggressively optimized kernel targets in AI, on NVIDIA Blackwell (B200) GPUs. Over 7 days of continuous autonomous evolution on multi-head attention, AVO discovers kernels that outperform cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations. The discovered optimizations transfer readily to grouped-query attention, requiring only 30 minutes of additional autonomous adaptation and yielding gains of up to 7.0% over cuDNN and 9.3% over FlashAttention-4. Together, these results show that agentic variation operators move beyond prior LLM-in-the-loop evolutionary pipelines by elevating the agent from candidate generator to variation operator, and can discover performance-critical micro-architectural optimizations that produce kernels surpassing state-of-the-art expert-engineered attention implementations on today's most advanced GPU hardware.
format Preprint
id arxiv_https___arxiv_org_abs_2603_24517
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle AVO: Agentic Variation Operators for Autonomous Evolutionary Search
Chen, Terry
Ye, Zhifan
Xu, Bing
Ye, Zihao
Liu, Timmy
Hassani, Ali
Chen, Tianqi
Kerr, Andrew
Wu, Haicheng
Xu, Yang
Chen, Yu-Jung
Chen, Hanfeng
Kane, Aditya
Krashinsky, Ronny
Liu, Ming-Yu
Grover, Vinod
Ceze, Luis
Bringmann, Roger
Tran, John
Liu, Wei
Xie, Fung
Lightstone, Michael
Shi, Humphrey
Machine Learning
Agentic Variation Operators (AVO) are a new family of evolutionary variation operators that replace the fixed mutation, crossover, and hand-designed heuristics of classical evolutionary search with autonomous coding agents. Rather than confining a language model to candidate generation within a prescribed pipeline, AVO instantiates variation as a self-directed agent loop that can consult the current lineage, a domain-specific knowledge base, and execution feedback to propose, repair, critique, and verify implementation edits. We evaluate AVO on attention, among the most aggressively optimized kernel targets in AI, on NVIDIA Blackwell (B200) GPUs. Over 7 days of continuous autonomous evolution on multi-head attention, AVO discovers kernels that outperform cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations. The discovered optimizations transfer readily to grouped-query attention, requiring only 30 minutes of additional autonomous adaptation and yielding gains of up to 7.0% over cuDNN and 9.3% over FlashAttention-4. Together, these results show that agentic variation operators move beyond prior LLM-in-the-loop evolutionary pipelines by elevating the agent from candidate generator to variation operator, and can discover performance-critical micro-architectural optimizations that produce kernels surpassing state-of-the-art expert-engineered attention implementations on today's most advanced GPU hardware.
title AVO: Agentic Variation Operators for Autonomous Evolutionary Search
topic Machine Learning
url https://arxiv.org/abs/2603.24517