NVIDIA AVO Hits 100.00 RHAE on ARC-AGI-3 Public Set; Same Harness Transfers from GPU-Kernel Optimization
NVIDIA reports that its AVO agent architecture achieved 100.00 RHAE on the ARC-AGI-3 public set, solving all 183 levels across 25 environments using Claude Opus 5 and 6,624 environment actions, about 12% fewer than VISTA's reported 7,542. The same AVO system was previously applied to GPU-kernel optimization, and both sources frame the result as evidence that the agent harness—persistent memory, supervision, and an inspect-plan-implement-evaluate loop—rather than model capability alone drives long-horizon performance.
The binding constraint moves from model capability to the agent harness: the same model family reportedly scores about 30% when evaluated as a model and 100.00 RHAE when wrapped in AVO, with persistent memory and supervision presented as the mechanisms that let evidence and feedback compound over long horizons. This reframes long-horizon ability as a systems property, with model selection as one component rather than the ceiling.
The core factual claim—a 100.00 RHAE score on the ARC-AGI-3 public set, all 183 levels in 25 environments—is asserted by NVIDIA and repeated by an independent newsletter; both rely on NVIDIA's own reporting. Evidence is strongest for the existence of the public-set result, moderate for action efficiency relative to VISTA, and weak for the 'general-purpose' transfer claim because the comparison is non-controlled, the model-only baseline comes from a different evaluation setting, and no private-set score or third-party replication is provided.