LLamjinLab
AI NewsWords 1524Read time≈ 4 min

NVIDIA Says AVO Completed All 183 ARC-AGI-3 Public Levels

NVIDIA says AVO completed all 183 ARC-AGI-3 public levels, highlighting agent-harness gains while leaving verification questions open.

AIAgentReasoning Model

1. NVIDIA’s AVO cleared the benchmark’s public environments

NVIDIA says its AVO coding agent achieved a 100% score on the public portion of ARC-AGI-3, completing all 183 levels across 25 interactive environments.

The company announced the result through its verified NVIDIA AI account on August 21. According to the post, AVO entered each environment without instructions, explicit rules, or a stated goal. It had to discover the available controls, infer what constituted progress, and determine how to complete each game through interaction.

This is materially different from answering a fixed collection of questions. ARC-AGI-3 presents an agent with a sequence of visual states and allows it to take actions. The agent must use the resulting changes to infer hidden rules, remember discoveries across a level, revise incorrect beliefs, and avoid wasting actions.

The announcement does not identify the foundation model used in the run. It also does not provide the inference cost, number of attempts, total action count, compute budget, complete trajectories, or a reproducible scorecard. The confirmed claim is therefore narrower than “a new model solved ARC-AGI-3”: NVIDIA says an AVO agent system completed the entire public set.

2. What a 100% ARC-AGI-3 score measures

ARC-AGI-3 was introduced in March 2026 as an interactive successor to the static grid-transformation tasks in ARC-AGI-1 and ARC-AGI-2. Its environments are designed to test four connected capabilities: exploration, goal acquisition, world-model construction, and planning.

A test taker receives no natural-language explanation of a game. It must determine which visual elements matter, test possible actions, infer causal relationships, and use those discoveries to act efficiently. The benchmark deliberately avoids tasks that depend on linguistic knowledge or facts retrieved from the internet.

Completion alone is insufficient for a perfect score. ARC-AGI-3 uses Relative Human Action Efficiency, or RHAE, to compare an agent’s action count with human performance. Under the current scoring system, a 100% result means the agent completed every level while meeting the benchmark’s aggregate human-relative efficiency requirement.

ARC Prize revised that system on April 14 after studying 458 human participants. The per-level baseline moved from the second-best human performance to the median human performance, reducing sensitivity to unusually lucky runs. The maximum contribution from an individual level also increased from 100% to 115%, allowing unusually efficient play on one level to offset weaker efficiency elsewhere. ARC Prize said the combined change raised human and AI scores by approximately 0.5 percentage points.

The difficulty gap at launch was substantial. ARC Prize’s technical report recorded humans as capable of solving every environment, while the listed frontier systems scored below 1% under the benchmark’s standardized, no-harness evaluation. By July 24, Claude Opus 5 at high reasoning effort had reached a verified 30.16%, showing rapid model-level progress but still leaving most of the benchmark unsolved without specialized orchestration.

3. AVO is an agent system, not a newly announced foundation model

AVO stands for Agentic Variation Operators. NVIDIA researchers introduced the approach in a March paper on autonomous evolutionary search.

Conventional LLM-assisted evolutionary systems place the model inside a fixed pipeline. The framework selects parent solutions, asks the model to generate a candidate, evaluates it, and updates the population using rules written in advance.

AVO expands the model’s role. Its general-purpose coding agent can inspect the full lineage of earlier solutions and scores, consult a domain-specific knowledge base, run evaluation tools, diagnose failures, edit an implementation, and decide when another test is useful. Persistent memory allows evidence from earlier attempts to influence later decisions.

In the published implementation, AVO accepts a new version only when it passes correctness checks and matches or improves the best committed score. Failed experiments remain available as part of the search history, while successful versions are preserved as commits. This turns variation from a one-shot generation call into an extended plan–implement–test–repair loop.

NVIDIA originally evaluated AVO on attention-kernel optimization for Blackwell B200 GPUs. During seven days of autonomous operation, the agent explored more than 500 optimization directions and committed 40 kernel versions. Its best multi-head attention implementation reached 1,668 TFLOPS at BF16 precision and exceeded NVIDIA cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the tested configurations.

The system then adapted the resulting work to grouped-query attention in approximately 30 minutes, producing gains of up to 7.0% over cuDNN and 9.3% over FlashAttention-4. Those experiments established AVO as a long-running software-optimization method. Applying the agent to unfamiliar interactive games is evidence that its loop may transfer beyond the kernel-search setting for which it was first documented.

NVIDIA has not yet published enough detail to determine exactly how the ARC-AGI-3 configuration maps onto the architecture described in the kernel paper. In particular, the announcement does not disclose what tools, memory structure, environment adapter, or benchmark-specific instructions were present.

4. The public-set result is not the benchmark’s first perfect score

AVO is not the first agent system reported to have completed all 183 public levels.

The Tycho research team reported a 100.00 RHAE score in July using both GPT-5.6 Sol and Claude Opus 5. Its Opus 5 run completed the public set in 6,641 scored actions, while its GPT-5.6 Sol run used 7,766. Tycho constructs executable, game-specific world models that can be tested, repaired, used for planning, or bypassed when direct reasoning is more appropriate.

ARC Prize’s Community Leaderboard lists Tycho at 100% with a July 29 date. It also lists several other heavily scaffolded systems near saturation, including Retrodict at 99.9% and the executable-world-model baseline1 system at 99.0%.

NVIDIA’s result therefore does not establish that ARC-AGI-3 has been solved for the first time. Its narrower research value is that another agent architecture—one previously demonstrated on autonomous software optimization—reportedly reached the same public-set ceiling.

The distinction also matters because ARC Prize does not independently verify Community Leaderboard submissions by default. Results produced on public tasks are generally self-reported, and the public environments, tools, scoring code, and human baselines are available to system developers.

ARC Prize’s technical report is unusually direct about this limitation. It describes the public set as a demonstration and says performance on it is “emphatically not a valid measure of progress towards AGI.” Official capability comparisons prioritize controlled evaluations on undisclosed tasks and exclude task-specific external harnesses.

A public-set score can still measure whether a documented agent design successfully explores and completes the available environments. It cannot, by itself, establish performance on unseen ARC-AGI-3 games, rule out benchmark-specific engineering, or demonstrate general intelligence.

5. What the result changes for agent developers

The AVO result adds evidence that the system surrounding a language model can be as consequential as the model selected.

ARC-AGI-3 rewards agents that preserve observations, turn tentative rules into testable representations, detect contradictions, and choose actions that both gather information and advance toward a goal. These requirements closely resemble long-running software work, where an agent must retain repository state, interpret test failures, revise an implementation, and avoid repeating unsuccessful approaches.

For developers, the useful comparison is therefore not simply 100% versus a lower model score. It is the difference between a stateless or lightly scaffolded model call and an agent equipped with durable memory, executable hypotheses, evaluation tools, and an iterative control loop.

The unresolved question is generalization. AVO’s earlier kernel results and the new game result cover two distinct domains, which is more informative than a benchmark-only demonstration. However, NVIDIA has not released an ARC evaluation package that would let independent researchers reproduce the run, inspect its trajectories, or test the same configuration on undisclosed environments.

Until those artifacts or a verified evaluation appear, the result should be treated as a strong company-reported demonstration of agent orchestration—not as a new foundation-model score or proof that ARC-AGI-3’s private generalization problem has been solved.

Frequently Asked Questions

What is NVIDIA AVO?

AVO, or Agentic Variation Operators, is an autonomous coding-agent approach that uses planning, tools, persistent memory, execution feedback, and repeated testing to improve candidate solutions.

Did NVIDIA release a new model?

No new foundation model was announced with the ARC-AGI-3 result. NVIDIA described AVO as a general-purpose coding agent, and the post did not identify its underlying model.

What does 100% on ARC-AGI-3 mean?

It means the reported system completed the benchmark’s 25 public environments and 183 levels while achieving the maximum aggregate human-relative action-efficiency score.

Is AVO the first system to reach 100%?

No. Tycho reported 100.00 RHAE on the same public set in July 2026 using GPT-5.6 Sol and Claude Opus 5.

Was NVIDIA’s result independently verified?

No independent ARC Prize verification or reproducible scorecard was included in NVIDIA’s announcement. Public-set and Community Leaderboard results are self-reported unless explicitly marked otherwise.

Sources

Share

Share this article