NVIDIA AVO

NVIDIA AVO, short for Agentic Variation Operators, is an experimental AI agent architecture developed by NVIDIA for sustained, feedback-driven search. NVIDIA introduced the method in a March 2026 preprint on autonomous GPU-kernel optimization. In August, the company reported applying the same general agent architecture to the ARC-AGI-3 interactive reasoning benchmark.^[1][2]

AVO is a coding agent system, not a foundation model. Its main agent inspects context, plans, edits or acts, evaluates results, and retains state across a long run; a supervisor can redirect the search after stagnation. The underlying large language model is one component of that system. The kernel paper did not name its model backend, while NVIDIA's later ARC-AGI-3 run used Claude Opus 5.^[1][2]

NVIDIA reported that AVO reached 100.00 Relative Human Action Efficiency (RHAE) on all 25 environments and 183 levels in the public ARC-AGI-3 demonstration set. This was not a semi-private or fully private competition result. ARC Prize defines its private environments as the benchmark's primary evaluation basis and says the public set is a demonstration interface rather than a measure of progress toward AGI.^[2][4]

Key facts

Scroll sideways for more →

Agentic variation operators

Classical LLM-assisted evolutionary search usually keeps selection and evaluation in a fixed outer algorithm. An LLM receives sampled parent solutions and generates a candidate. The AVO paper expresses that pattern as Vary(P_t) = Generate(Sample(P_t)). AVO instead makes the full variation step Agent(P_t, K, f): the agent can inspect the complete lineage P_t, consult a domain knowledge base K, and call a scoring function f as it decides what to try next.^[1]

For the kernel experiment, each lineage item was source code with inline CUDA PTX plus its scores. The knowledge base included CUDA programming guides, the PTX instruction-set reference, Blackwell specifications, and existing kernel implementations. The evaluator checked numerical correctness and measured throughput across the benchmark configurations. A correctness failure received a score of zero.^[1]

One variation step could contain many edits, compilations, profiler runs, failed tests, and revisions. A candidate became a committed version only after it passed correctness and matched or improved the best committed score. Failed or regressing attempts remained part of the private agent trajectory rather than the visible lineage. This makes the paper's 40 committed versions a filtered history, not a count of all generated programs or all model calls.^[1]

The reported implementation used one continuous lineage. The paper presents population branching, bounded archives, and other evolutionary structures as possible extensions, not evaluated features. Persistent conversation history carried forward code changes, compiler output, profiles, scores, and reasoning. A supervisor watched for plateaus or repeated unproductive cycles and could suggest alternative directions, while the main agent retained control of inspection, implementation, testing, and evaluation.^[1][2]

GPU-kernel optimization study

The initial study targeted forward-pass attention kernels on NVIDIA Blackwell B200 GPUs. NVIDIA reported a seven-day continuous multi-head attention run without human intervention, more than 500 explored optimization directions, and 40 committed versions. The agent later adapted the resulting kernel to grouped-query attention in about 30 minutes.^[1]

The software setup used CUDA 13.1 and PyTorch 2.10.0. The comparisons used cuDNN 9.19.1 and the official FlashAttention-4 implementation at commit 71bf77c. The experiment reused FlashAttention-4's timing script and warmup/repeat convention, then ran the experiment ten times to calculate averages and standard deviations.^[1][11][12]

The benchmark scope was narrow and hardware-specific: BF16 forward prefilling, head dimension 128, sequence lengths from 4,096 through 32,768, and a fixed total of 32,768 tokens maintained by changing batch size. Multi-head attention used 16 heads under causal and noncausal masking. The grouped-query attention tests used 32 query heads with either four or eight key-value heads, based on Qwen3 configurations.^[1]

For causal multi-head attention, NVIDIA reported gains from 0.4% to 3.5% over cuDNN and from 5.0% to 10.5% over FlashAttention-4, with a peak of 1,668 TFLOPS. For noncausal attention, the reported cuDNN gain was 1.8% to 2.4% only at sequence lengths above 16K; the shorter-sequence results were within measurement noise. The adapted grouped-query kernel reached reported maxima of 7.0% over cuDNN and 9.3% over FlashAttention-4 for causal configurations, and 6.0% and 4.5% respectively for noncausal configurations.^[1]

Three version-to-version ablations illustrate what the authors attributed to the agent's changes. Replacing branched accumulator rescaling with a speculative branchless path, together with a lighter fence, produced an 8.1% noncausal and 1.6% causal geometric-mean gain from version 19 to 20. Overlapping correction work with the second matrix-multiply stage added 1.1% and 0.4% from version 29 to 30. Reallocating registers among warp groups added 2.1% for noncausal attention and approximately zero for causal attention from version 32 to 33.^[1]

These results were reported by NVIDIA authors in a preprint and were not independently reproduced in the sources reviewed. The percentage maxima apply only to the evaluated forward BF16 B200 configurations. They do not establish a general speedup for all CUDA kernels, model training, inference workloads, GPU generations, or precision formats.^[1]

ARC-AGI-3 application

NVIDIA later connected the same general AVO architecture to ARC-AGI 3. According to the company, the agent loop remained in place while the environment-specific tools and evaluator changed. The task interface was independently reimplemented and informed by the direct-interaction design of VISTA, rather than centered on Tycho's explicit executable world-model approach.^[2][7][8]

The full public-set run used Claude Opus 5. Each observation was serialized as an exact 64 by 64 text grid, so the model received neither rendered images nor image tokens. The agent was given the currently available actions but no description of the game rules, goals, or action effects. It had to infer mechanics and objectives through interaction.^[2]

NVIDIA reported a 100.00 RHAE score after completing all 183 levels across the 25 public environments in 6,624 environment actions. Its Aug. 21 announcement accurately named those environments as public, and the later technical post added an editor's note explaining that the wording had been revised to distinguish the public set from the semi-private and fully private competition sets.^[2][3]

Interpreting the public-set score

RHAE combines completion with action efficiency relative to first-time human baselines. ARC Prize establishes the per-level baseline from the upper median among first-time human players, gives later levels more weight, normalizes scores within games, and averages across games. A score of 100 on this set means completing the public levels at human-comparable or better measured action efficiency under that aggregation.^[5]

The metric counts only discrete commands that change the environment. Internal reasoning, read-only inspection, tool calls, and retries do not count as actions. The figure of 6,624 therefore describes the scored interaction budget, not the number of model calls, total tool uses, tokens, retries, wall-clock time, compute, or cost. NVIDIA did not disclose those broader resource totals or repeated-run variance for the ARC run.^[2][5]

The public/private split is equally important. ARC Prize's technical report calls the public set a demonstration interface and says the broader, out-of-distribution private sets are the primary evaluation basis. It defines success on ARC-AGI-3 in terms of first-time performance averaged across private environments. AVO's 100.00 public score cannot be converted into a private score and did not constitute an ARC Prize competition win.^[4]

Public-set saturation was not unique to AVO. ARC Prize's community leaderboard, where public scores are generally self-reported, already listed Tycho at 100.0. VISTA also reported a full 183-level sweep. These systems use different observation formats, memory strategies, prompts, model interfaces, and planning mechanisms, so their public results show that multiple systems reached the demonstration-set ceiling, not that their architectures are equivalent.^[6][7][8]

NVIDIA compared AVO's 6,624 environment actions with VISTA's 7,542 actions using Claude Opus 5, a difference of about 12%. The company explicitly said this was not a controlled ablation because the agent backend, observation representation, memory, context management, and other implementation details differed. The comparison does not isolate the value of AVO's memory, supervisor, or any other component.^[2][7]

The separate ARC Prize result of 30.16% for Claude Opus 5 High also used a different reasoning setting, agent system, and evaluation setup. NVIDIA states that the 30.16-to-100 contrast is not a direct measurement of AVO's contribution. It demonstrates the importance of evaluating a complete agent configuration, but it cannot assign the score difference to the harness alone.^[2][9]

NVIDIA also paired AVO with GPT-5.6 Sol on a challenging subset. It reported that Sol was faster in wall-clock time in some matched cases while Opus used fewer environment actions, but called the observations preliminary and left a systematic comparison for future work. No full public-set AVO result with Sol was reported.^[2]

Reproducibility, licensing, and security

The kernel paper provides enough configuration detail to define its measured workload and baselines, but not enough artifacts to reproduce AVO itself. It does not identify the kernel-run model, publish the generated kernels or unsuccessful trajectories, disclose total model usage, or link agent and supervisor source code. The ARC post likewise does not release its interface code, prompts, scorecard bundle, replays, per-level logs, token totals, or cost.^[1][2]

No official AVO source repository or software license was located by Aug. 22, 2026. A public repository named coder-2011/avo was created after the paper by an account not listed among its authors; it describes itself as an implementation effort and has no detected license. It is an unaffiliated third-party reproduction attempt, not evidence that NVIDIA released AVO. The arXiv paper's distribution terms are also not a software license for the agent or kernels.^[1][13]

AVO's documented tools create a security-relevant boundary: the kernel agent can edit files, run shell commands, retrieve documentation, compile code, and execute generated kernels. Numerical-correctness and throughput gates protect the experiment's result quality, but they are not a published security model. The reviewed materials do not describe an AVO-specific sandbox, permission scheme, credential isolation, network policy, dependency controls, audit log, red-team report, or incident process. NVIDIA's blog says trusted agents require system-wide security, but it does not provide an AVO security assessment.^[1][2]

Independent reporting described AVO as research rather than a new NVIDIA product and quoted NVIDIA distinguishing the model from the harness, runtime, skills, and tools around it.^[10] As of the cutoff, there was no public AVO API, download, product pricing, or release schedule. Later code publication, private-set evaluation, or a controlled component study would materially change the reproducibility and evidence picture.

References

1. ^Terry Chen et al. "AVO: Agentic Variation Operators for Autonomous Evolutionary Search." arXiv:2603.24517v1, March 25, 2026. arxiv.org/...2603.24517
2. ^Terry Chen, Yeyin Zhu, Zhifan Ye, Jean-Francois Puget, and Humphrey Shi. "NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents." NVIDIA Developer Blog, Aug. 21, 2026. developer.nvidia.com/...-horizon-autonomous-agents
3. ^NVIDIA AI. "Our general-purpose coding agent just scored 100% on the ARC-AGI-3 interactive reasoning benchmark." X, Aug. 21, 2026. x.com/...2090786258981466231
4. ^ARC Prize Foundation. "ARC-AGI-3 Technical Report." 2026. arcprize.org/...ARC_AGI_3_Technical_Report.pdf
5. ^ARC Prize Foundation. "ARC-AGI-3 Scoring Methodology." Accessed Aug. 22, 2026. docs.arcprize.org/methodology
6. ^ARC Prize Foundation. "ARC-AGI Community Leaderboard." Accessed Aug. 22, 2026. arcprize.org/...community
7. ^MIT VISTA team. "VISTA: A Visual Harness for Reasoning in an Interactive World." Accessed Aug. 22, 2026. vista-research.github.io
8. ^Jens Lehmann, Andrei Aioanei, and Sahar Vahdati. "Tycho: ARC-AGI-3 Solver using Rendered Deterministic Moore Machines." GitHub repository, accessed Aug. 22, 2026. github.com/...Tycho
9. ^ARC Prize Foundation. "Claude Opus 5 - ARC-AGI Results." Accessed Aug. 22, 2026. arcprize.org/...anthropic-claude-opus-5
10. ^Julie Bort. "Nvidia just showed that the harness, not the AI model, is now the real hero." TechCrunch, Aug. 21, 2026. techcrunch.com/...he-ai-model-is-now-the-real-hero
11. ^Ted Zadouri et al. "FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling." arXiv:2603.05451v1, March 5, 2026. arxiv.org/...2603.05451
12. ^Dao-AILab. "benchmark_attn.py at commit 71bf77c." FlashAttention GitHub repository. github.com/...benchmark_attn.py
13. ^coder-2011. "avo: Implementation of the AVO paper." GitHub repository, accessed Aug. 22, 2026. github.com/...avo