NVIDIA AVO achieves 100% on ARC-AGI-3, proving system design over model capability

NVIDIAAI · x · 2026-08-21

NVIDIA released the Agentic Variation Operators (AVO) architecture, achieving a 100% score on the ARC-AGI-3 interactive reasoning benchmark by completing all 183 levels. By integrating persistent memory, supervision, and tool-use, AVO elevated Claude Opus 5's baseline performance from 30% to 100%. In GPU kernel optimization tasks, AVO autonomously explored over 500 directions, achieving up to 10.5% better performance than FlashAttention-4. The study demonstrates that high performance in long-horizon tasks derives more from system-level architecture than from model capability alone.

Related event: NVIDIA's AVO Agent Scores 100% on ARC-AGI-3(2 posts)→

Original post →

More from coding & agent

coding & agent channel →