Nvidia’s AVO Clears ARC-AGI-3’s Public Set. Withheld Tests Still Matter.

Nvidia says its AVO agent system completed every level in ARC-AGI-3’s public benchmark set, reaching a 100.00 Relative Human Action Efficiency score with Claude Opus 5 underneath it. The result is a vivid demonstration of how much an agent’s memory, tools and control loop can matter—but it stops before the benchmark’s semi-private and private tests.

AVO cleared 183 levels across 25 interactive environments in 6,624 environment actions, according to Nvidia. ARC-AGI-3 puts agents into unfamiliar game-like settings without instructions, stated rules or stated goals; they must experiment, infer the environment and use actions efficiently as levels become harder.

RHAE combines task completion with action efficiency relative to first-time human baselines, then aggregates performance across levels and environments. Clearing levels alone is not enough: the metric also rewards reaching them with fewer interactions.

The system keeps working after one model context ends

AVO is Nvidia’s general-purpose coding-agent system. Around the language model, it adds persistent memory, software tools, an execution loop and a supervisor that can redirect work when progress stalls. The arrangement is designed to carry prior results and reasoning across work that outlasts a single model context.

In its evolutionary-search work, AVO replaces a predefined variation step with an agent that chooses what to inspect, change, test and commit next. For ARC-AGI-3, Nvidia connected that same agent architecture to a different task interface rather than building an ARC-specific world-model layer.

What AVO adds around Claude Opus 5

• Persistent memory retains prior implementations, evaluation results and accumulated reasoning for later work.
• Tools and an execution loop let the system inspect code, make changes, run commands and validate outcomes through execution.
• Supervision monitors the broader search trajectory and can push the main agent toward another strategy after stagnation.

100.00RHAE score

Nvidia says AVO achieved a 100.00 RHAE score on the 25-environment ARC-AGI-3 public set.

183Levels completed

The run completed all 183 public-set levels.

6,624Environment actions

Nvidia says the system used 6,624 environment actions to finish the public set.

An efficiency gap with too many moving parts

Nvidia compares AVO’s 6,624 actions with 7,542 actions reported by VISTA, another Claude Opus 5 harness that also completed the 183 public levels. That is approximately 12% fewer actions for AVO, but Nvidia explicitly says it is not a controlled ablation: the systems differ in their agent backends, observation formats, memory and context management.

The input path is one concrete difference. AVO gave the model exact 64-by-64 text grids, with no images or image tokens. Nvidia says VISTA’s primary configuration uses a rendered 512-by-512 PNG, though VISTA has also explored text-grid representations.

A transfer from kernel work, not proof of broad generalization

Nvidia first used AVO for GPU-kernel optimization, where an agent must inspect implementations, change code, test it on hardware and revise its approach. In one seven-day attention-kernel experiment, Nvidia says AVO explored more than 500 optimization directions and committed 40 kernel versions.

On Nvidia DGX B200 systems, the company says those multihead-attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations. Nvidia says AVO then adapted the evolved kernel to grouped-query attention in about 30 minutes of further autonomous work.

The public-set sweep supports Nvidia’s narrower conclusion: its long-horizon architecture can be adapted from software engineering to an interactive reasoning task. It cannot yet establish how the system performs on ARC-AGI-3’s semi-private or fully private competition sets, which Nvidia did not test.