NVIDIA's AVO agent scores 100% on ARC-AGI-3, beats FlashAttention-4 in kernel optimization

MickeySteamboat · x · 2026-08-22

NVIDIA's new AVO agent architecture achieved 100% on the ARC-AGI-3 public set, completing all 183 levels across 25 environments. While the Claude Opus 5 model baseline scores around 30%, integrating it into AVO with persistent memory, supervision, and tools yields perfect performance. Previously, the same architecture ran autonomously for 7 days to optimize GPU kernels, exploring 500+ directions and producing kernels up to 10.5% faster than FlashAttention-4. This demonstrates that the AI frontier is shifting from the model itself to the entire agent system built around it.

Related event: NVIDIA's AVO Agent Scores 100% on ARC-AGI-3 Public Set(18 posts)→

Original post →

More from coding & agent

coding & agent channel →