NVIDIA AVO scores 100% on ARC-AGI-3's public set — a self-reported harness win, not a model breakthrough

NVIDIA reports its AVO agent harness scored 100.00 RHAE on the ARC-AGI-3 public set — all 25 environments and 183 levels — using Claude Opus 5 as the base model, in 6,624 environment actions. The result is real but narrower than the headline: it is self-reported, it covers the public demonstration set rather than the held-out evaluation sets, and the arXiv paper NVIDIA links for technical details is about GPU-kernel optimization and never mentions ARC-AGI-3.

What shipped

AVO (Agentic Variation Operators) is a general-purpose coding-agent system from NVIDIA. Its core idea is to replace the fixed mutation, crossover, and hand-designed heuristics of classical evolutionary search with an autonomous agent that decides what to inspect, what to change, what to test, and what to commit at each step. The paper describes variation as "a self-directed agent loop that can consult the current lineage, a domain-specific knowledge base, and execution feedback to propose, repair, critique, and verify implementation edits."

Two mechanisms carry the long-horizon load:

• Persistent memory carries prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning forward, so the agent resumes from its current state instead of reconstructing the search.
• A supervisor watches the broader trajectory for stagnation or repeated unproductive cycles and redirects the main agent when progress plateaus.

The paper covers the first test, GPU-kernel optimization. Over a seven-day autonomous run on NVIDIA DGX B200 systems, NVIDIA reports AVO explored more than 500 optimization directions and committed 40 kernel versions; the resulting multihead-attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations. The agent then adapted the evolved kernel to grouped-query attention in about 30 minutes of additional autonomous work, with gains up to 7.0% over cuDNN and 9.3% over FlashAttention-4.

The ARC-AGI-3 result is blog-only. NVIDIA connected the same agent to a different task interface — text-only 64×64 grids, with no images or image tokens sent to the model — and reports 100.00 RHAE across all 25 public-set environments, completing all 183 levels in 6,624 environment actions. For comparison it cites VISTA, a direct-interaction harness, as reporting 7,542 actions with the same model on the same levels — roughly 12% fewer for AVO. On ARC-AGI-3, a 100% score means an agent beats every game as efficiently as humans.

Why it matters

The substantive claim is the system-over-model one. ARC Prize's verified result page reports Claude Opus 5 at 30.2% on ARC-AGI-3 at High reasoning effort — a standardized, no-harness score. NVIDIA's argument is that the same model inside a harness with memory, tools, feedback, and recovery reaches 100% on the public set. Whether or not the number survives scrutiny, the direction matches everything else that has happened on this benchmark.

Three independent harnesses have now reported 100 RHAE on the same 25 public games: Tycho (executable programmatic world models), VISTA (lossless visual memory over a rendered 512×512 PNG), and AVO (text grids plus persistent memory and a supervisor). The efficiency numbers are close — Tycho's Opus 5 run used 6,641 scored actions, AVO's 6,624, VISTA's 7,542 — despite very different designs. The common thread: none of them throw state away; all preserve observations, reasoning, and notes across the run. OpenAI's own report found the same lever, with GPT-5.6 Sol going from 13.3% to 38.3% on the public set after its runner retained reasoning state and replaced rolling deletion with context compaction.

The practical lesson is more nuanced. Harness design — memory and recovery in particular — moves these scores substantially, but the base model still matters: genaipod's hands-on testing found that upgrading the model or giving it more reasoning time worked more reliably than any single harness trick. NVIDIA's framing: the same frontier model is a 30% model or a 100% model depending on what surrounds it.

Caveats

• Self-reported, on the demonstration set. The 100% is NVIDIA's own report on the public set, which ARC Prize's François Chollet has stressed is a "demonstration set," not an eval set; public-set scores, he says, do not indicate private-set scores. The semi-private and private evaluation sets remain unsolved by any system, and the $700K grand prize requires 100% on the evaluation set — which AVO has not reported attempting.
• The 30% → 100% framing mixes two evals. ARC Prize's 30.2% is a verified, standardized-runner score; the 100% is a self-reported harness run on the public games. NVIDIA itself says the comparison "should not be interpreted as a direct measurement of the performance contribution of AVO."
• The paper doesn't cover the headline. The linked arXiv paper, submitted March 25, 2026, is entirely about GPU-kernel evolution. The ARC-AGI-3 run and the 6,624-action count appear only in the blog post.
• Not a controlled ablation. NVIDIA notes the VISTA comparison differs in agent backend, observation representation, memory, and context management, and that no experiment isolates the memory system's contribution.
• Saturation. With three harnesses at 100 RHAE on the same 25 games, the public set has stopped discriminating between architectures. Laura Martel's analysis adds that agents get lossless rewind, zoom, and persistent notes while the human baseline is a single first-time pass, and that training-data exposure for models released after the games shipped "cannot be ruled out."

Sources

• NVIDIA: AVO Reaches 100% on ARC-AGI-3
• AVO: Agentic Variation Operators for Autonomous Evolutionary Search (arXiv)
• ARC-AGI-3 benchmark
• ARC Prize: Claude Opus 5 verified result
• ARC Prize 2026 ARC-AGI-3 competition
• VISTA project page
• Tycho: Active Abstraction with Programmatic World Models (arXiv)
• genaipod: ARC-AGI-3, a game with no instructions
• Laura Martel: VISTA and the harness race
• NVIDIA DGX B200