NVIDIA AVO Scores 100 on the ARC-AGI-3 Public Set

NVIDIA reports 100.00 RHAE on all 25 ARC-AGI-3 public-set environments

NVIDIA reported on August 21, 2026 that AVO, running Claude Opus 5, completed the full 25-environment ARC-AGI-3 public set with a 100.00 RHAE score. The run solved all 183 levels in 6,624 environment actions. This is NVIDIA's own report, and it is limited to the public set.

ARC-AGI-3 is an interactive reasoning benchmark. An agent enters unfamiliar game-like environments with no instructions, no explicit rules and no stated goal, and it has to work out through interaction how the environment behaves and what the goal is. The benchmark scores it with Relative Human Action Efficiency (RHAE), which combines task completion with per-level action efficiency against first-time human baselines.

Finishing a level is not enough on its own, because the number of actions spent counts too. NVIDIA also states what the score does not cover. The post says the results use the official scorecard and the RHAE metric on the public set, and that they are not results on the semi-private or fully private competition sets.

NVIDIA compares its action count with VISTA, another system that reports a result with Claude Opus 5. VISTA reports 7,542 actions for the same 183 levels, so AVO used about 12% fewer.

The two systems also show the game to the model differently. VISTA's primary setup renders a 512 x 512 PNG, while the AVO setup sends only an exact 64 x 64 text grid and no images.

What an agent harness is, and what AVO adds around the model

An agent harness is the system around the language model that decides how the model receives context, uses tools, keeps state and recovers when something fails. NVIDIA's definition adds one more job: sustaining progress over long-running tasks. A model on its own answers one request at a time, while the harness is what lets a chain of requests add up to finished work.

AVO, short for Agentic Variation Operators, is NVIDIA's research project for building that layer as a general-purpose agent. It is a coding agent that can inspect and edit code, run commands, consult documentation and check its work by running it. NVIDIA says its distinguishing focus is sustained autonomous operation across long horizons.

NVIDIA built AVO to keep its progress beyond a single model context, and it names two mechanisms for this. Persistent memory carries forward prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning, so the agent resumes from the current state instead of rebuilding its search. A supervisor watches the trajectory for stagnation and can redirect the main agent toward other strategies.

AVO started on GPU-kernel optimization, where the work moves through many candidate versions rather than one generated answer. NVIDIA's attention-kernel study shows what the design did over a long run.

The attention-kernel study, as published by NVIDIA

Item	Result
Continuous operation	7 days
Optimization directions explored	More than 500
Kernel versions committed	40
Versus cuDNN	Outperformed it by up to 3.5%
Versus FlashAttention-4	Outperformed it by up to 10.5%
Adapting to grouped-query attention	About 30 minutes, autonomous

The kernels were measured on NVIDIA DGX B200 systems across the evaluated configurations. The figures in the table come from this kernel study, not from the ARC-AGI-3 run.

The jump from 30 and the 12% action gap are not controlled comparisons

NVIDIA itself says neither comparison attached to the result is a controlled comparison or a measurement of what AVO contributed. The 12% action gap against VISTA is the first. The two systems differ in agent backend, observation representation, memory, context management and other implementation details, so the gap cannot be traced to any single one of them.

The 30% figure is the second. ARC Prize separately reports about 30% for Claude Opus 5 at High reasoning effort. NVIDIA's run used the same model family, but with another reasoning setting, a substantially different agent system and its own evaluation setup.

That means the jump from 30 to 100 is a comparison of two different evaluations. Subtracting one from the other does not give AVO's contribution, and NVIDIA says the numbers should not be read as a direct measurement of it. Describing the result as the same model going from 30% to 100% skips this step.

NVIDIA does point to one architectural difference that may matter over long horizons: AVO's memory system, which NVIDIA built to carry useful understanding forward and cut repeated exploration. The post adds that this experiment does not isolate that contribution, so it stays a hypothesis.

What the public-set result can and cannot show about agent harnesses

The result supports the claim NVIDIA makes itself: evaluating a model alone does not characterize the performance of a complete agent. A benchmark number belongs to a whole system, meaning the model, its reasoning setting, the harness around it and the way the environment is shown to it. There are three things the result does not show:

• The effect of the AVO harness by itself. Neither comparison separates it from the reasoning setting, the observation format or the rest of the agent system.
• Performance beyond the public set. NVIDIA says the results are not on the semi-private or fully private competition sets.
• Behavior across models. The full public-set result used Claude Opus 5. NVIDIA describes its pairing with GPT-5.6 Sol as a limited experiment on a challenging subset of games, with preliminary results.

When you compare agent scores, the practical check is short: the system that produced the number, its reasoning setting and the set it was run on. NVIDIA's post states the set and the system, and it says its reasoning setting differed from ARC Prize's. Those caveats are easy to quote and easy to lose in a headline.

A related case points the same way. OpenAI reported that two API options moved its ARC-AGI-3 score (related article: ARC AGI 3: Two Settings Tripled OpenAI's Score).