NVIDIA researchers say an agent architecture called AVO has completed every level of the public ARC-AGI-3 reasoning benchmark, a rare perfect score on one of the toughest tests built for autonomous AI agents. The result, published in a technical blog post on August 21, 2026, credits the system’s design — not raw model power alone — for the jump.
From chip design to abstract reasoning
AVO, short for Agentic Variation Operators, wasn’t built for ARC-AGI-3. NVIDIA originally developed it as an autonomous evolutionary-search system that writes and refines its own GPU kernels: over seven days of continuous operation, it explored more than 500 optimization directions and produced 40 committed kernel versions for Blackwell GPU attention operations, beating NVIDIA’s own cuDNN library by up to 3.5% and the FlashAttention-4 kernel by up to 10.5%, according to the underlying research paper.
To test whether that architecture generalizes beyond code, the team wrapped it around Claude Opus 5 and pointed it at ARC-AGI-3, a benchmark that drops agents into unfamiliar game-like environments with no instructions and expects them to infer the rules through trial and error. Running alone at high reasoning effort, Claude Opus 5 solves only about 30% of the public set’s 183 levels across 25 environments. Wrapped in AVO’s loop of planning, execution, evaluation and persistent memory, the same model cleared all of them, for a perfect 100.00 Relative Human Action Efficiency score.
Efficient, but not the full picture
AVO also used about 12% fewer environment actions than VISTA, the system that previously topped the public leaderboard — 6,624 actions versus 7,542 — though NVIDIA cautions the two systems differ enough in design that this isn’t a clean, controlled comparison.
The bigger caveat is scope: the perfect score covers only ARC-AGI-3’s public set of 25 environments. The benchmark’s harder semi-private and fully private test sets, which the ARC Prize competition uses for ranking and are much harder to game, remain untested by AVO. “A frontier language model is only one component of an AI agent,” the researchers wrote. “The surrounding agent system…determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks.”