NVIDIA's AVO Achieves 100% on ARC-AGI-3, Outperforms FlashAttention-4

daniel_mac8 · x · 2026-08-22

NVIDIA published AVO (Agentic Variation Operators), a method that replaces classical evolutionary search heuristics with autonomous coding agents. AVO uses a self-directed loop to propose, repair, critique, and verify edits based on lineage, knowledge bases, and feedback. In optimizing multi-head attention on Blackwell GPUs, AVO discovered kernels outperforming cuDNN by 3.5% and FlashAttention-4 by 10.5%. Additionally, AVO combined with Opus 5 boosted ARC-AGI-3 benchmark scores from 30% to 100%, demonstrating how advanced harnesses expose latent LLM capabilities.

Related event: NVIDIA's AVO Coding Agent Scores 100% on ARC-AGI-3(12 posts)→

Original post →

More from coding & agent

coding & agent channel →