What happened
NVIDIA reports that its Agentic Variation Operators system, or AVO, achieved a 100.00 RHAE score on ARC-AGI-3’s 25-environment public set. Using Claude Opus 5 and text-only 64-by-64 grid observations, AVO completed all 183 levels in 6,624 environment actions. NVIDIA also describes a separate seven-day GPU-kernel optimization run in which AVO explored more than 500 directions and produced 40 committed versions.
NVIDIA’s Aug. 21, 2026 technical blog presents Agentic Variation Operators, or AVO, as a general-purpose coding-agent architecture designed for work that extends beyond a single model response. The system can inspect and edit code, run commands, consult documentation and validate changes through execution. Its central design includes persistent memory for prior implementations, test results, compiler and profiler outputs, and accumulated reasoning, along with a supervisor that watches for stagnation or repeated unproductive cycles.
The company first describes AVO in a seven-day GPU-kernel optimization experiment. According to NVIDIA, the system explored more than 500 optimization directions and produced 40 committed kernel versions. On NVIDIA DGX B200 systems, the resulting multihead-attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the configurations evaluated by the team. NVIDIA says AVO then adapted the evolved kernel to grouped-query attention in about 30 minutes of additional autonomous work. The source does not provide independent measurements, full configuration details or a reproducible account of every test.
NVIDIA then connected the same underlying AVO architecture to ARC-AGI-3, an interactive reasoning benchmark made up of unfamiliar game-like environments. The company says the agent received no stated rules or goals and had to infer how actions worked through interaction. In this configuration, the model received exact 64-by-64 text grids rather than images or image tokens, and the system used environment-specific tools and evaluation while retaining the broader AVO execution loop.
The company reports a 100.00 Relative Human Action Efficiency score across all 25 public environments, completing all 183 levels in 6,624 environment actions. NVIDIA compares that with 7,542 actions reported by VISTA for Claude Opus 5 on the same public levels, a difference of approximately 12%. NVIDIA explicitly says this is not a controlled ablation: the systems differ in backend, observation format, memory, context management and other implementation details. The source also says limited tests paired AVO with GPT-5.6 Sol, but these preliminary comparisons were not a full systematic evaluation.
Read the primary source: developer.nvidia.com ↗
Why it matters
The result supports NVIDIA’s argument that persistent memory, supervision, tool use and recovery can materially affect long-horizon agent performance. It is notable because the same underlying architecture was applied to GPU-kernel optimization and an unfamiliar interactive benchmark. However, the source does not establish that AVO caused the full performance gain, nor that the system generalizes to real-world work.
The important technical claim is about system design. AVO’s performance is presented as the product of a loop that forms hypotheses, takes actions, observes results, preserves useful state, revises assumptions and recovers from errors. That is different from evaluating a model on isolated questions. If the approach holds up, developers may increasingly treat memory, supervision, tool interfaces and recovery mechanisms as core capability components rather than optional wrappers around a language model.
The transfer between tasks is the most consequential part of NVIDIA’s report. GPU-kernel optimization depends on source code, compilers, profilers and hardware-grounded measurements. ARC-AGI-3 instead supplies unfamiliar interactive environments and action outcomes. NVIDIA argues that the transferable element is not domain knowledge but the machinery for sustaining progress when feedback arrives over many steps. The result is therefore relevant to research on agents that must work through changing states rather than produce one-off outputs.
The report also illustrates why headline benchmark scores can obscure what is being measured. The 100.00 RHAE score belongs to the complete AVO-plus-model-plus-interface system. It does not show that Claude Opus 5 alone improved from the source’s cited approximately 30% public-set result, because the runs used different reasoning settings and substantially different evaluation systems. Nor does the action comparison prove that persistent memory, the supervisor or any other individual component caused the difference.
Practical implications remain uncertain. The source does not establish how often AVO fails, how much human oversight is required in ordinary use, what the seven-day optimization run costs, how much compute it consumes, or whether its behavior is dependable in safety-critical settings. It also does not demonstrate that success in unfamiliar benchmark environments translates to software engineering, scientific research or business operations outside the tested setups. NVIDIA’s evidence is significant as a company-reported research result, but it is not by itself independent confirmation of broad agent reliability.
What to watch next
The key follow-up is independent replication and controlled testing. Comparisons with VISTA are not ablations because the systems differ in model backend, observations, memory and context management. NVIDIA’s result covers only the public benchmark set; performance on semi-private or private sets, operating cost, failure modes, safety controls and transfer to practical deployments remain unknown.
Replication should be the first test. Researchers would need to rerun AVO or an equivalent implementation under documented conditions and verify the 6,624-action result on the public ARC-AGI-3 set. Useful reporting would include failure traces, action counts by environment, repeat-run variability, inference and tool costs, and whether results depend on particular prompts, memory contents or execution settings.
Controlled component studies would clarify the source of the gain. Comparisons should separately test persistent memory, the supervisor, context management, observation format and model choice while holding other variables constant. The NVIDIA post says memory may reduce repeated exploration, but it does not isolate that contribution. The same caution applies to the comparison with VISTA, whose architecture and inputs differ in several ways.
The benchmark’s unseen portions are another unresolved measure. NVIDIA states that its reported results cover the 25-environment public set and are not results on the semi-private or fully private competition sets. Future results on held-out environments would provide stronger evidence about generalization, especially if the evaluation prevents the agent or its developers from tailoring the system to public tasks.
Finally, readers should watch for evidence about deployment rather than benchmark completion alone. A long-running agent that can act through tools needs safeguards around permissions, state persistence, recovery and human intervention. The source mentions system-level security as an important design principle but does not report a security evaluation of AVO. Its real-world significance will depend on whether it can maintain accuracy, control costs, expose its reasoning and stop safely when feedback is ambiguous or its assumptions are wrong.
