NVIDIA AVO reportedly reaches 100% on ARC-AGI-3 demo tasks

• AVO completed all 183 levels across ARC-AGI-3's 25 public environments for a 100.00 RHAE score, according to NVIDIA's technical post; testingcatalog's report described the result as a full task sweep.
• The widely repeated 30% → 100% comparison joins a different agent system and evaluation setup to the same Claude Opus 5 family. NVIDIA's report explicitly says it is not a controlled AVO ablation, while AILeaksAndNews' post supplied the 30% baseline framing.
• The 100% claim covers the public set, not ARC-AGI-3's semi-private or private competition sets, as fchollet's correction stressed.
• AVO adds persistent memory, an inspect-to-repair loop, execution feedback, and a supervisor around the model, mechanics daniel_mac8's AVO outline broke down from NVIDIA's release.

AVO's original research paper describes a seven-day attention-kernel run that explored more than 500 directions and committed 40 kernel versions. NVIDIA reports the resulting kernels beat FlashAttention-4 by up to 10.5% on tested DGX B200 configurations in its technical account.

Relative Human Action Efficiency, or RHAE, combines task completion with action efficiency relative to first-time human baselines. NVIDIA's 100.00 score aggregates all 183 levels in the 25 public environments.

François Chollet, ARC-AGI's creator, compared calling that result a full benchmark score to claiming a video game win after clearing its tutorial. NVIDIA's post now carries an editor's note saying its wording was revised to distinguish the public set from semi-private and private competition sets.

AVO's research paper replaces a conventional evolutionary search system's fixed mutation, crossover, and hand-designed heuristics with a coding agent that chooses its next candidate.

The system combines:

• Persistent memory for prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning.
• A five-step loop: inspect context, plan, implement, evaluate, then diagnose and repair.
• A solution lineage that stores candidates and scores, plus a knowledge base of code and documentation.
• An evaluator for correctness and performance.
• A supervisor that intervenes when the search trajectory stalls; failed candidates are repaired and retried, while successful ones update the lineage.

For ARC-AGI-3, AVO supplied Claude Opus 5 with exact 64 × 64 text grids, not rendered images. The agent received available actions without descriptions of rules or goals, and inferred both by interacting with the environment.

NVIDIA says AVO took 6,624 environment actions, versus 7,542 for VISTA, a 12% reduction in this cross-system comparison. The VISTA project and AVO differ in agent backend, observation representation, memory, and context management, so NVIDIA says the action gap does not isolate any individual AVO component.

NVIDIA's technical post reports environment actions but no model-call count, token total, wall-clock duration, or API cost for the ARC run. François Chollet asked for the cost per run in fchollet's follow-up.