NVIDIA AVO Solves All 183 Levels in the ARC-AGI-3 Public Set, With Key Gains Coming From the Agent Framework Rather Than a Model Change

NVIDIA connected Agentic Variation Operators (AVO), originally developed for GPU kernel optimization, to ARC-AGI-3 and achieved 100.00 RHAE across 25 public environments comprising 183 levels. This interactive benchmark provides neither game rules nor explicit objectives. The agent must observe state transitions through actions, then independently explore, build a world model, infer goals, and plan.

AVO did not add visual input to Claude Opus 5. Instead, it converted each screen into an exact 64×64 text grid. The primary agent proposes hypotheses, takes actions, and revises its strategy, while persistent memory stores prior observations and experimental results to prevent repeated exploration after the model context is reset. A separate supervisor agent detects stalls or unproductive loops and changes the search direction when necessary. The full evaluation used 6,624 environment actions, about 12% fewer than the 7,542 reported by VISTA for likewise completing the entire public set.

The result extends AVO’s design for performance engineering. Its original version replaced the fixed mutation/crossover operations of conventional evolutionary search with autonomous coding agents capable of consulting candidate lineages, documentation, compiler output, and profiler feedback. In the paper’s seven-day B200 experiment, the system explored more than 500 directions and produced 40 submitted versions, with some attention configurations running up to 10.5% faster than FlashAttention-4.

However, a score of 100 should not be interpreted as Claude Opus 5 itself jumping from roughly 30% performance to artificial general intelligence. NVIDIA explicitly states that this was not a controlled ablation: AVO and the comparison systems differed in memory, reasoning settings, observation formats, and context management. The evaluation also did not cover the semi-private or private environments. The official ARC Prize competition requires solutions to operate without internet access, under compute constraints, and winning entries to be open sourced. The next things to watch are whether NVIDIA releases the complete AVO implementation, its cost and run-to-run variance, and whether it can maintain comparable efficiency in unseen environments.