Nvidia agent AVO solves ARC-AGI-3 outright — 100 percent
Nvidia's agent system AVO cleared all 183 levels of the ARC-AGI-3 reasoning benchmark — with about 12 percent fewer actions than the prior best system.
Illustration · AI-generated (AI IN LIFE)
At a glance
- AVO solves all 183 levels across 25 public ARC-AGI-3 environments (NVIDIA, August 21, 2026)
- 6,624 environment actions — about 12 percent fewer than VISTA's 7,542
- The base model alone scores only about 30 percent in the ARC Prize framework
- Kernel marathon: 7 days autonomous, 500+ directions explored, 40 versions committed
- Up to 10.5 percent faster than FlashAttention-4, 3.5 percent better than cuDNN
What happened? Nvidia's agent system AVO (Agentic Variation Operators) is the first to fully solve the interactive reasoning benchmark ARC-AGI-3: all 183 levels across all 25 public environments, with no instructions, explicit rules or stated goals. The result was published on the NVIDIA Developer Blog on August 21, 2026.
How efficient was it? AVO needed 6,624 environment actions and scored 100.00 on RHAE (Relative Human Action Efficiency) — roughly 12 percent fewer actions than previous leader VISTA, which used 7,542 actions with the same base model. Strikingly, the underlying frontier model on its own reaches only about 30 percent in the ARC Prize framework.
Why does it matter? The authors, including Terry Chen and Humphrey Shi, put it plainly: "A frontier language model is only one component of an AI agent." What decides performance is the harness — the architecture that feeds context, orchestrates tools and maintains state over long horizons. Persistent memory management and a supervisor that detects stagnation and redirects the search made the difference.
Can AVO do more than benchmarks? Yes. In a second run, the system optimised GPU kernels for seven continuous days, explored more than 500 optimisation directions and committed 40 kernel versions. The outcome: up to 10.5 percent better performance than FlashAttention-4 on DGX B200 systems and a 3.5 percent improvement over cuDNN.
What does this mean for businesses? The case shows that competitive advantage increasingly sits in agent architecture, not in the model alone. Anyone deploying AI agents in production — including Europe's SMBs — should watch individual model releases less and invest more in orchestration, memory and control mechanisms.
FAQ
What is ARC-AGI-3?
An interactive reasoning benchmark with 183 levels across 25 environments where agents must figure out rules and goals on their own — with no instructions.
What makes AVO special?
Not the model but the harness: persistent memory, supervisor oversight and long autonomous work cycles lift performance from about 30 to 100 percent.
Is the record unqualified?
It covers ARC-AGI-3's public level set; results on hidden test sets are still outstanding.
