NVIDIA researchers have demonstrated that a general-purpose agent architecture called Agentic Variation Operators (AVO) can achieve a perfect 100.00 RHAE score on the ARC-AGI-3 public benchmark, completing all 183 levels across 25 interactive environments. The system, built around Anthropic's Claude Opus 5, solved the full public set using 6,624 environment actions — about 12% fewer than the VISTA system's 7,542 actions on the same tasks.
The result reframes a benchmark that had previously stumped every frontier model. As of March 2026, ARC Prize Foundation testing showed GPT-5.4, Claude Opus 4.6, and Gemini 3.1 all scoring below 1% on ARC-AGI-3 when run as straightforward agents. NVIDIA's AVO architecture lifted the same model family from that near-zero baseline to full completion, illustrating that system-level design — persistent memory, supervisory intervention, and autonomous tool-use loops — can unlock capabilities that model scaling alone has not delivered.
What's new
- Architecture: AVO is a general-purpose coding agent system with persistent memory, a supervisor that monitors for stagnation, and an autonomous execution loop that decides what to inspect, change, test, and commit.
- ARC-AGI-3 result: 100.00 RHAE score on the public set (25 environments, 183 levels) using Claude Opus 5; 6,624 environment actions versus VISTA's 7,542 on the same levels.
- GPU-kernel proof point: Before ARC-AGI-3, AVO ran autonomously for seven days on attention-kernel optimization, exploring over 500 directions and committing 40 kernel versions. The best kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 systems, then adapted to grouped-query attention in ~30 minutes of additional autonomous work.
- Model-agnostic design: Limited experiments pairing AVO with GPT-5.6 Sol on a challenging subset showed Sol reaching matched levels faster in wall-clock time, while Opus used fewer environment actions — suggesting complementary operating profiles across models.
- Text-only observation: AVO received each ARC-AGI-3 observation as an exact 64×64 text grid, with no images or image tokens sent to the model.
Why the benchmark matters
ARC-AGI-3 departs from static grid puzzles (ARC-AGI-1/2) by dropping agents into unfamiliar turn-based game environments with no instructions, rules, or stated goals. Agents must explore, infer dynamics, discover objectives, and plan efficiently across progressively harder levels. The Relative Human Action Efficiency (RHAE) metric combines task completion with per-level action efficiency relative to first-time human baselines, making it a long-horizon test of sustained autonomous progress rather than single-shot reasoning.
Humans solve 100% of ARC-AGI-3 environments. Until this result, frontier models scored below 1% — a gap the ARC Prize Foundation explicitly framed as the measurable distance between current AI and human-like fluid intelligence. NVIDIA's result does not close that gap for models alone; it shows the gap can be narrowed by a different agent architecture.
Our take
The 100% score is real on the public set, but the comparison to VISTA is not a controlled ablation — the two systems differ in backend, observation format, memory, and context management. What transfers from GPU-kernel optimization to ARC-AGI-3 is not domain knowledge but the machinery for long-horizon autonomous progress: hypothesis formation, action, observation, state preservation, and recovery. That transfer is the stronger claim than the benchmark number itself.
