Nvidia AVO achieves 100% on ARC-AGI-3 reasoning benchmark
August 21, 2026
Nvidia AVO, a general-purpose coding agent, successfully completed all 183 levels across 25 public environments in the ARC-AGI-3 benchmark. The agent demonstrated zero-shot interactive reasoning without explicit instructions or stated goals.
HOW THIS AFFECTS YOU
●
builderThis signals progress in agents that can operate autonomously in novel environments without manual prompting.
●
researcherThis provides a new baseline for evaluating zero-shot reasoning and agentic problem-solving capabilities.