NVIDIA AVO scores 100% on ARC-AGI-3 public set (announcement)

Summary

NVIDIA announced that its general-purpose coding agent AVO (Agentic Variation Operators) scored 100% on the ARC-AGI-3 interactive reasoning benchmark's public set: 100.00 RHAE across all 183 levels in 25 environments, using 6,624 environment actions, about 12% fewer than prior leader VISTA's 7,542. The backbone model is Claude Opus 5, which scores roughly 30% standalone — the 70-point lift comes from the agent scaffolding (persistent memory, stagnation-supervision loop, inspect-plan-implement-evaluate agent loop), not a model change.

Keywords

agents · ARC-AGI-3 · reasoning · NVIDIA · scaffolding · benchmark

Topics

agents, ARC-AGI-3, reasoning, NVIDIA, scaffolding

Research notes

• Discovery: Announcement X post from NVIDIA AI; underlying work is NVIDIA's AVO technical blog post of 2026-08-21.
• Method: Agentic harness around Claude Opus 5 with three components: persistent memory carrying implementations and reasoning between attempts, a supervision loop flagging stagnation, and an agent loop that inspects, plans, implements, and evaluates. AVO was originally built for GPU-kernel optimization (500+ directions explored over 7 days, 40 committed versions, beating FlashAttention-4 by up to 10.5% on DGX B200), then pointed at ARC-AGI-3 with no game-specific teaching.
• Key findings: 100.00 RHAE on all 183 ARC-AGI-3 public levels vs ~30% for the same Claude Opus 5 backbone alone; 6,624 actions vs 7,542 for VISTA (~12% more action-efficient); Same agent architecture transferred from CUDA kernel optimization to interactive reasoning puzzles without retraining for games
• Limitations: Company-reported result; covers only the 25-environment public set, not ARC Prize's semi-private/private held-out sets, and has not appeared as a verified entry on the official ARC Prize leaderboard. ARC-AGI-3's platform does not currently allow external harnesses on the hidden private set.
• X post itself is login-walled; substance verified via multiple secondary reports (aiweekly, nextaipress, courionai, wccftech, cryptobriefing) of the 2026-08-21 NVIDIA technical blog post.