AVO agent architecture achieves perfect ARC-AGI-3 score and produces optimized GPU kernels

A general-purpose agent architecture called AVO (Agentic Variation Operators) has been detailed in a new NVIDIA publication; it is designed to sustain long-running autonomous work. The system was first used for GPU-kernel optimization and has now been applied to the interactive reasoning benchmark ARC-AGI-3, where it achieved a perfect 100.00 RHAE score across all 25 public-set environments, completing all 183 levels.

The core claim is that the agent harness—the surrounding system that handles context, tool use, memory, and feedback—matters as much as the model itself. AVO uses two mechanisms to keep work moving beyond a single model context: persistent memory that carries forward implementations, evaluation results, and accumulated reasoning, and a supervisor that monitors the broader search trajectory and can redirect the main agent when progress stalls.

In the GPU-kernel work, AVO ran continuously for seven days on NVIDIA DGX B200 systems, exploring more than 500 optimization directions and producing 40 committed kernel versions. The resulting multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across evaluated configurations. The agent then adapted the evolved kernel to grouped-query attention in roughly 30 minutes of additional autonomous work.

The ARC-AGI-3 run involved no instructions, stated rules, or goals. The agent entered unfamiliar environments and operated in a text-only modality: each observation was a 64×64 text grid, with no images sent to the model. The agent received available actions without descriptions of their effects and had to infer the environment's dynamics through interaction.

The comparison with VISTA, a similar direct-interaction harness, comes with qualifications. Both systems completed the same 183 public-set levels using Claude Opus 5, with AVO using 6,624 environment actions versus VISTA's 7,542—roughly 12% fewer. NVIDIA notes this is not a controlled ablation: the systems differ in agent backend, observation representation, memory, and context management. ARC Prize separately reports approximately 30% for Claude Opus 5 at High reasoning effort, but that run used a different reasoning setting and evaluation setup, so the numbers are not directly comparable.

The transferability claim—that the same architecture works for both GPU-kernel optimization and interactive reasoning—rests on the observation that both tasks share a common loop: form a hypothesis, act, observe consequences, update state, and continue. NVIDIA frames this as generality emerging from the machinery of sustained progress rather than from domain knowledge.

Two limitations apply to these results. The results cover only the ARC-AGI-3 public set, not the semi-private or private competition sets. Preliminary experiments pairing AVO with GPT-5.6 Sol on a subset of games showed Sol reaching matched levels faster in wall-clock time in several cases while Opus used fewer environment actions—suggesting complementary operating profiles across models, but leaving systematic cross-model comparison to future work.