NVIDIA AVO Architecture Hits 100% on ARC-AGI-3, Showing Agent Design Beats Model Capability Alone

NVIDIA's research team has published results showing its Agentic Variation Operators (AVO) architecture achieved a perfect 100.00 RHAE score on the ARC-AGI-3 benchmark, completing all 183 levels across 25 environments. The ARC-AGI-3 test challenges AI agents to navigate unfamiliar interactive environments without instructions, stated rules, or defined goals.

AVO is a general-purpose coding agent built by NVIDIA. Its core design separates system architecture from model capability. The system uses Claude Opus 5 as its underlying language model, which on its own scores about 30% on ARC-AGI-3. With the full AVO agent harness, that same model reaches 100%.

The architecture centers on two mechanisms that allow agents to sustain work over long timeframes: persistent memory, which carries forward prior implementations, evaluation results, and accumulated reasoning so the agent can resume rather than restart; and a supervisor component, which monitors the agent's search trajectory and redirects it when progress stalls or repeats.

NVIDIA first applied AVO to GPU-kernel optimization, where the agent ran continuously for seven days, explored more than 500 optimization directions, and committed 40 kernel versions. The resulting attention kernel outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%.

The team then applied the same architecture to ARC-AGI-3 without redesigning the agent, changing only the environment-specific tools and evaluation interface. The underlying agent remained identical.

The central finding is that evaluating a model and evaluating an agent are not the same thing. The surrounding system, including how the model receives context, uses tools, maintains state, recovers from failure, and sustains progress, determines how effectively model capability converts into autonomous results.

Key takeaways:

• NVIDIA AVO scored 100% on ARC-AGI-3 using Claude Opus 5 as its base model (30% baseline without the harness)
• Two mechanisms drive long-horizon performance: persistent memory and a supervisor module
• Same architecture applied to GPU-kernel optimization achieved a 7-day autonomous run, outperforming cuDNN by 3.5%
• Core finding: agent system design determines autonomous performance, not model capability alone
• AVO architecture is general-purpose and transfers across task types without redesign

Read the full article on NVIDIA Developer Blog