AVO: Autonomous Evolutionary Search Optimises GPU Kernels Using Coding Agents

ArticleTerry Chen, Zhifan Ye, Bing Xu, et al.Research

Researchers have introduced Agentic Variation Operators (AVO), a framework that replaces fixed mutation and crossover operators in evolutionary search with autonomous coding agents. Evaluated on NVIDIA Blackwell (B200) GPUs, the generated attention kernels outperformed cuDNN and FlashAttention-4.

Key points

  1. AVO replaces traditional, fixed mutation and crossover heuristics with autonomous coding agents within evolutionary search.
  2. The agent loop consults the current lineage, a domain knowledge base, and execution feedback to propose, repair, critique, and verify implementation changes.
  3. Over seven days of continuous autonomous evolution on NVIDIA B200 GPUs, AVO discovered multi-head attention kernels outperforming cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%.
  4. Optimisations transferred to grouped-query attention (GQA) within 30 minutes, delivering gains of up to 7.0% over cuDNN and 9.3% over FlashAttention-4.

Why it matters

Hand-tuning GPU kernels for advanced hardware demands deep micro-architectural expertise. AVO shows that autonomous agentic loops can uncover low-level performance optimisations that exceed state-of-the-art human-engineered implementations.

Analysis

This work demonstrates an evolution from simple LLM-in-the-loop candidate generators to self-directed variation operators. By integrating compiler and runtime feedback into the iterative mutation process, the system autonomously navigates complex search spaces. The 30-minute adaptation to GQA highlights strong transferability across related kernel variants.

Mentioned links

Found in

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Article:Terry Chen, Yeyin (Eva) Zhu, Zhifan Ye, Jean-Francois Puget and Humphrey Shi

    NVIDIA AVO Achieves 100% on ARC-AGI-3 Benchmark

    NVIDIA has introduced Agentic Variation Operators (AVO), an agent architecture that enables Claude Opus 5 to reach a 100% completion rate on the ARC-AGI-3 interactive reasoning benchmark.

    AI & LLMs· Research

  • Article:Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu

    WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

    This research paper introduces WikiSkill, a framework that co-evolves AI agent skills alongside a persistent knowledge base (wiki). By disentangling raw execution experience, accumulated knowledge, and executable skills, WikiSkill enables systematic reuse of insights across iterations.

    AI & LLMs· Research

  • Article:Shuai Shao, Kangning Zhang, Qingyao Li, Shijian Wang, Hao Wang, Wenxiang Jiao, Yuan Lu, Yi Guo, Weiwen Liu, Weinan Zhang

    Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories

    Researchers introduce Harness-R1, a framework that trains a dedicated agent to adapt and patch the runtime harness of LLM agents based on failure logs. Using online reinforcement learning, a separate 9B model learns to generate executable patches for context handling, tool mediation, and error recovery to directly improve task completion.

    AI & LLMs· Research

  • Article:Devansh

    Stateful Swarms: How Persistent Memory Outperforms Traditional Agent Architectures

    Devansh explores 'Stateful Swarms', an architecture tackling volatile contexts and repetitive document processing through a persistent blackboard state. Instead of repeatedly processing large inputs or losing signal across conversational handoffs, specialized agents populate an append-only, structured memory. Evaluated on the Harvey Legal Agent Benchmark, this setup yielded higher task pass rates at significantly lower operational costs.

    AI & LLMs· Opinion

  • Article:Chris Alexiuk and Chintan Patel

    NVIDIA Nemotron 3.5 Lightning: Compact MoE Model for Agentic Execution

    NVIDIA has announced Nemotron 3.5 Lightning, an open 30B mixture-of-experts (MoE) model featuring 3B active parameters. It is tailored for the high-volume execution layer of long-running AI agents, targeting routine tasks such as tool calling and result validation with low latency.

    AI & LLMs· Announcement

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.