NVIDIA AVO Agent Scores 100% on ARC-AGI-3 Benchmark

NVIDIA’s AVO coding agent achieved a perfect score on a public AI benchmark, demonstrating that specialized wrappers can significantly outperform their underlying base models. This result matters because it suggests that agent architecture, rather than raw model intelligence, may be the primary driver for complex reasoning tasks. Developers and researchers are monitoring this trend as the industry shifts focus from training larger models to building more effective interaction layers.

Specialized wrapper beats base model on public AI test

AVO functions as a wrapper around Anthropic’s Claude Opus 5 model, designed to handle autonomous problem-solving without explicit instructions. The agent navigated 25 publicly available games comprising 183 distinct levels by learning through trial, error, and observation. It received no prior goals or rules, relying instead on its ability to remember past actions and correct mistakes in real time.

The underlying Claude Opus 5 model scored only 30 percent on the same ARC-AGI-3 public dataset. In contrast, AVO solved all 183 levels using a total of 6,624 actions. This performance represents a 12 percent increase in efficiency compared to other leading agent wrappers like VISTA, which required 7,542 actions to clear the public set.

AVO was originally built to optimize CUDA GPU kernels, where it previously beat FlashAttention-4 by up to 10.5 percent. The ARC-AGI-3 public test serves as a new benchmark for this broader capability. We do not know how AVO would have performed on the hidden private ARC-AGI-3 dataset, as the evaluation platform does not allow custom-built external agent harnesses to run against it.

The agent's performance illustrates a move toward autonomous coding tools capable of self-correction without human guidance. NVIDIA has positioned AVO as a tool that learns by observing results and adjusting its approach. This capability has the potential to streamline software development workflows by reducing the need for manual debugging and instruction tuning.