Agentic Variation Operators (AVO) replace fixed evolutionary heuristics with coding agents to discover GPU kernels that outperform FlashAttention-4 by 10.5%.
March 26, 2026
Original Paper
AVO: Agentic Variation Operators for Autonomous Evolutionary Search
arXiv · 2603.24517
The Takeaway
By elevating LLMs from simple generators to self-directed variation operators that critique and verify code, this method discovered optimizations on NVIDIA B200 GPUs that surpass expert-engineered SOTA. It effectively automates the hardest part of systems programming: micro-architectural optimization.
From the abstract
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, a