How NVIDIA’s AVO Agent Reached 100% on ARC-AGI-3

NVIDIA says its Agentic Variation Operators (AVO) system achieved a perfect 100.00 RHAE score on the public set of the interactive ARC-AGI-3 benchmark. Running with Claude Opus 5, AVO completed all 183 levels across 25 environments in 6,624 actions.

The striking part is not that a new foundation model suddenly solved the benchmark. It is that a long-horizon agent architecture built for software engineering and GPU-kernel optimization transferred to a very different interactive reasoning task.

The short version: AVO improved the system around the model

Claude Opus 5 alone was listed at 30.2% on the published ARC-AGI-3 leaderboard snapshot, while GPT-5.6 Sol scored 7.8% under the standard harness.

NVIDIA’s AVO result used the same general underlying model together with an execution system designed to sustain work over many steps. That system can inspect and modify artifacts, run tools and commands, consult documentation, evaluate results and revise its approach based on external feedback.

In practical terms, AVO replaces a one-shot question-and-answer pattern with a loop:

1. Form a hypothesis about the task.
2. Take an action through an available tool.
3. Observe the environment’s response.
4. Store useful discoveries and state.
5. Check whether the result works.
6. Revise the strategy and continue.

That loop is especially important in ARC-AGI-3, where agents must explore unfamiliar interactive environments and infer rules and objectives through interaction rather than simply respond to a fully specified prompt.

What AVO’s architecture adds

The reported architecture combines several capabilities that are easy to describe individually but powerful when coordinated over a long task horizon.

Persistent memory

Persistent memory allows the agent to retain discoveries, failed approaches and useful state instead of treating every move as an isolated exchange. In a benchmark built around exploration, that can reduce repeated mistakes and help the system build a working model of an unfamiliar environment.

Execution tools and external feedback

AVO was developed as a coding-agent system that can inspect work, make changes, execute commands and validate outcomes. Its reasoning is therefore connected to observable results: a proposed change can be tested, measured and revised rather than accepted because it sounds plausible.

Supervision and recovery

A supervisory layer helps coordinate the longer process. Rather than allowing one model call to control every decision indefinitely, the system can maintain oversight of progress, identify unproductive directions and support recovery when an assumption fails. NVIDIA and related descriptions characterize AVO as combining memory, execution tools, external feedback and supervision for sustained autonomous work.

Together, these elements give the model a structured way to explore, act, observe and correct. The model still supplies much of the reasoning, but the harness determines how that reasoning is carried through the environment.

How the result compares with Claude Opus 5 and GPT-5.6 Sol

The comparison is useful, but the scores must be read alongside their evaluation setups.

System or configuration	Reported ARC-AGI-3 result	Evaluation context
AVO with Claude Opus 5	100.00 RHAE	Public set; 183 of 183 levels completed
Claude Opus 5 alone	30.2%	Published leaderboard snapshot and standard comparison setup
GPT-5.6 Sol	7.8%	Standard or verified semi-private comparison
GPT-5.6 Sol with retained reasoning and compaction	38.3%	OpenAI-reported custom run on public tasks

The AVO and standalone Opus figures are not identical kinds of measurement: the first is a complete agent-system result on the public set, while the second is a model score under a benchmark harness. That distinction is central to the story.

GPT-5.6 Sol makes the same point from another direction. ARC Prize’s published results list Sol at 7.78% at maximum reasoning effort, while OpenAI separately reported 38.3% on public tasks after preserving reasoning across turns and using compaction. Those results are not a like-for-like replacement for the official leaderboard score, but they show how memory and state handling can materially affect an agent benchmark.

The safest conclusion is not that one model is universally better. It is that autonomous performance depends substantially on the model, the memory policy, the tool interface, the state-management strategy and the evaluation harness working together.

Why AVO came from GPU-kernel optimization

AVO’s original environment was difficult software engineering and GPU-kernel optimization. In that setting, an agent must inspect an implementation, propose a change, run hardware-grounded tests, interpret performance feedback and decide what to try next. Success requires repeated experimentation rather than a single correct generation.

NVIDIA reports that AVO explored more than 500 directions, committed 40 kernel versions and achieved up to 10.5% better performance than FlashAttention-4 on DGX B200 systems.

The transferable capability is not necessarily knowledge of GPU kernels. It is the experimental process: generate a candidate, execute it, measure the result, preserve what was learned and change direction when the evidence disagrees with the hypothesis. ARC-AGI-3 presents a different task interface, but it also rewards agents that can discover structure through repeated interaction.

AVO also emphasized action efficiency

AVO reportedly solved the full public set in 6,624 environment actions, compared with 7,542 reported for VISTA—a reduction of about 12% while completing the same 183 levels.

That matters because ARC-AGI-3’s RHAE metric incorporates action efficiency relative to human performance, not just whether an agent eventually reaches a successful state. A long-horizon system must therefore manage exploration carefully: excessive trial and error can make an otherwise capable agent inefficient, expensive or impractical.

What the result means for enterprise AI

For enterprise applications, the most useful lesson is architectural. A reliable autonomous system is unlikely to be just an LLM connected to a handful of tools. It needs a governed execution layer around the model.

That layer can include:

• persistent task memory and explicit state management;
• narrow, testable tool interfaces;
• validation gates before important actions are accepted;
• observable intermediate steps and audit logs;
• recovery, rollback and escalation paths; and
• clear permissions for actions that affect customers, systems or data.

AVO also strengthens the case for modular agent stacks. If much of the performance gain comes from the harness, organizations may benefit from keeping memory, tool adapters, evaluation suites, policy controls and observability separate from the foundation model. That makes it easier to compare models, replace a provider or tune the system for a particular workflow without rebuilding the entire stack.

But the benchmark result should not be treated as proof of enterprise reliability. NVIDIA’s 100% figure is a reported result on ARC-AGI-3’s public set. It does not establish equivalent performance on private benchmark tasks, production data or high-risk business processes. Independent replication, hidden-set testing, cost and latency analysis, security review and safety evaluation are still necessary.

The broader takeaway

AVO’s result shifts attention from the question “Which model is smartest?” to a more operational question: Which system can preserve context, use tools, learn from feedback and recover reliably over a long workflow?

The model remains important, but ARC-AGI-3 offers a vivid demonstration that the surrounding harness can determine whether that capability survives contact with an unfamiliar environment. For organizations building autonomous AI, the competitive advantage may increasingly lie in the quality of the execution system—not in the model call alone.