NVIDIA AVO achieves 100% efficiency on ARC-AGI-3 benchmark

NVIDIA’s Agentic Variation Operators (AVO) research project has achieved a 100% Relative Human Action Efficiency (RHAE) score on the ARC-AGI-3 benchmark. The system successfully completed all 183 levels across 25 different environments.

Unlike traditional AI models that rely solely on internal language capabilities, the AVO architecture utilizes a multi-layered approach including persistent memory, supervision, and feedback-driven mechanisms to handle long-horizon, multi-step tasks. While the system used Claude Opus 5 as its underlying engine, NVIDIA noted that the success was driven by the agentic structure—specifically its ability to learn from errors and adjust strategies autonomously in unfamiliar environments without predefined rules.

The ARC-AGI-3 benchmark is designed to test interactive reasoning, requiring agents to infer objectives and environmental dynamics. AVO’s performance represents a significant advancement over previous AI attempts, which struggled to achieve even a 0.51% RHAE score.