Nvidia's AVO: Revolutionizing AI with 100% Benchmark Score
Nvidia on Friday published research showing its general-purpose agent system, Agentic Variation Operators (AVO), achieved a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark — a result that underscores how agent architecture, not just model capability, determines performance on complex autonomous tasks.
The Architecture That Made the Difference
Using Claude Opus 5 as its underlying model, AVO completed all 183 levels across 25 environments in the ARC-AGI-3 public set with a 100.00 Relative Human Action Efficiency (RHAE) score, accomplishing the task in 6,624 environment actions. Without the AVO harness, Opus 5 scored just 30% on the same benchmark — the top result among all models tested at high reasoning effort.
ARC-AGI-3 is a benchmark consisting of 2D game-like environments with no instructions, rules, or stated goals. An agent must explore, infer dynamics, and act efficiently across progressively harder levels. The benchmark has proven difficult for frontier models, with OpenAI's models scoring less than 10% before the company conducted its own harness-optimization research last month.
AVO's distinguishing features are persistent memory, which carries forward prior implementations and reasoning across sessions, and a supervisor component that monitors the agent's trajectory and redirects it when it stagnates or revisits unproductive paths. "The more interesting part was introducing a supervising agent in addition to your main agent that's doing the work," Adel El Hallack, vice president of product in Nvidia's AI unit, told TechCrunch. It "almost acts like a CEO to nudge the agent when it goes off direction."
From GPU Kernels to General Reasoning
AVO was originally developed for GPU-kernel optimization, where it ran continuously for seven days, explored more than 500 optimization directions, and produced attention kernels that outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 systems. The same architecture transferred to ARC-AGI-3 without modification to its core loop — only the environment-specific tools and evaluation changed.
The system also showed versatility across models. In limited experiments pairing AVO with GPT-5.6 Sol, the model reached matched levels faster in wall-clock time in several cases, while Opus used fewer environment actions.
Implications for Enterprise AI
Nvidia's research adds to a growing body of evidence that the harness — not model selection alone — is the critical variable in agentic performance. Databricks CEO Ali Ghodsi noted in July that harness choice alone can double AI costs for the same model. "You think, oh, this is an expensive model. This is a cheap model. But wait, which harness are you using?"
El Hallack framed the results as an argument for open agent stacks. "We believe in having an open agent stack — where you have control across the harness, across the infrastructure, across the runtime — is what's required for us to usher the ecosystem forward and securely," he said.
