NVIDIA AVO Agent Hits 100% on ARC-AGI-3 Public Set

NVIDIA's Agentic Variation Operators (AVO) research project achieved a 100.00 RHAE score on the ARC-AGI-3 public set, completing all 183 levels across 25 environments. The agent, built on Claude Opus 5, used 6,624 environment actions, roughly 12% fewer than the 7,542 actions reported by the VISTA system with the same model. NVIDIA states this is not a controlled ablation, as the systems differ in agent backend, observation representation, memory, and context management.

System design over model capability

The ARC-AGI-3 benchmark places agents in unfamiliar environments without instructions, rules, or stated goals. AVO operated in a text-only modality, receiving each observation as an exact 64 x 64 text grid with no images or image tokens. The agent received available actions without descriptions of rules or goals and inferred their effects through interaction.

NVIDIA notes that ARC Prize separately reports approximately 30% for Claude Opus 5 at High reasoning effort. The AVO run used the same model family under a different reasoning setting and a substantially different agent system, so the numbers should not be read as a direct measurement of AVO's contribution. The result illustrates that model-level evaluation alone does not characterize a complete agent.

Persistent memory and supervision

AVO's architecture combines persistent memory with a supervisor. Persistent memory carries forward prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning, allowing the agent to resume from the current state. The supervisor monitors the broader trajectory for stagnation or repeated unproductive cycles and can redirect the main agent toward alternative strategies.

In a seven-day attention-kernel optimization run on NVIDIA DGX B200 systems, AVO explored more than 500 optimization directions and produced 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 later adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work.

Transfer across domains

NVIDIA applied the same general-purpose agent to ARC-AGI-3 by connecting it to a different task interface, with only environment-specific tools and evaluation changed. The underlying agent remained the same. The team adopted direct-interaction design principles described by VISTA and reimplemented the task interface independently, rather than using an explicit programmatic world-model construction approach like Tycho.

The common computational pattern across GPU-kernel optimization and ARC-AGI-3 involves building hypotheses from incomplete evidence, taking actions through an external interface, observing consequences, preserving useful state, revising the problem model, recovering from incorrect assumptions, and continuing progress over a long horizon. NVIDIA says what transfers is not domain knowledge but the machinery for sustained autonomous progress.

For inference engineers, the result suggests that serving a frontier model is only part of the equation. The harness, memory system, and supervision loop determine how effectively model capability converts into long-horizon task completion. AVO is designed to operate across frontier models.