NVIDIA AVO Achieves Perfect Public ARC-AGI-3 Score
NVIDIA's AVO agent achieved a 100.00 RHAE score on the public ARC-AGI-3 benchmark on August 21, completing all 183 levels across 25 environments. NVIDIA reports that AVO uses Claude Opus 5 as its primary model and that the full agent system outperformed the model's reported 30% baseline. The result applies to the public set, not a hidden evaluation set.
NVIDIA's AVO agent achieved a 100.00 RHAE score on the public ARC-AGI-3 benchmark, completing all 183 levels across 25 environments, according to an NVIDIA developer blog published August 21. NVIDIA reports that the run used 6,624 environment actions and required about 12% fewer actions than VISTA, a comparison system cited in the post.
AVO, short for Agentic Variation Operators, was originally developed for autonomous GPU-kernel optimization. NVIDIA describes the system as an agent architecture combining persistent memory, supervision, and tool use for long-horizon tasks. For ARC-AGI-3, NVIDIA applied the core agent loop to an interactive reasoning benchmark rather than CUDA kernel engineering.
A system-level result
According to NVIDIA, Claude Opus 5 is AVO's primary underlying model. The company's post reports a 30% ARC-AGI-3 baseline for Opus 5 alone, compared with AVO's 100% result as a complete agent system. Wccftech similarly characterizes AVO as a software harness around Opus 5, with tools, feedback loops, memory, and prompt structure surrounding the model.
That distinction is central to interpreting the reported result. NVIDIA explicitly frames the experiment as evidence that agent-system design, rather than base-model capability alone, can materially affect long-horizon task performance. However, Explainx notes that NVIDIA did not present the result as a controlled ablation that isolates the contribution of each architectural component. The public reporting therefore supports a comparison between the full AVO system and a stated model baseline, but not a precise causal allocation among memory, tool interfaces, supervision, prompting, and the underlying model.
Explainx also reports that the benchmark run covered only ARC-AGI-3's public environments, not a private or hidden test set. A perfect score on a public set is notable, but it is not equivalent to proving that the benchmark is solved under undisclosed evaluation conditions.
From kernels to interactive reasoning
NVIDIA's developer post reports that AVO previously explored more than 500 directions and committed 40 kernel versions in GPU-optimization work. The company reports up to 10.5% better performance than FlashAttention-4 on NVIDIA DGX B200 systems in that application.
For the ARC-AGI-3 experiment, Explainx reports that NVIDIA changed the task interface and tools while retaining the core agent loop. Wccftech reports that the agent received no explicit rules or stated goals for the benchmark games, instead learning through interactions with the environments.
For ML engineers, the result reinforces a familiar systems lesson
model evaluations increasingly measure a compound stack, not merely a model checkpoint. Tool access, state management, retry policies, evaluators, memory selection, and action budgets can all alter an agent's measured capability. Companies building comparable autonomous systems commonly need to track those components separately in experiment logs and ablations, particularly when comparing an agentic workflow with a direct model prompt.
The reported action count also matters operationally. An agent that solves a benchmark through iterative environment interaction has a different latency, cost, and reliability profile from a single-pass inference system. Public benchmark scores should therefore be read alongside action budgets, tool permissions, evaluation-set exposure, and reproducibility details.
NVIDIA's reported ARC-AGI-3 outcome is a strong public-set demonstration of AVO's architecture transfer. The available reporting leaves open how consistently the same design generalizes across hidden benchmarks, production software tasks, and different foundation models.
Key Points
- 1NVIDIA reports AVO completed all 183 public ARC-AGI-3 levels, demonstrating an unusually strong result for a long-horizon agent system.
- 2The reported 30% Claude Opus 5 baseline versus AVO's 100% score emphasizes that harness design can materially change agent evaluations.
- 3Public-set scores require careful interpretation because action budgets, tools, memory, and hidden-set performance affect real-world generalization claims.
Scoring Rationale
A perfect public ARC-AGI-3 result from an NVIDIA agent architecture is a notable benchmark and systems-engineering development for ML practitioners. Its practical significance is tempered by the public-only evaluation and the absence, in the available reporting, of controlled ablations separating model and harness effects.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
