NVIDIA AVO Achieves 100% on ARC-AGI-3 Public Set, Demonstrating Long-Horizon Autonomous Architecture

NVIDIA AVO Achieves 100% on ARC-AGI-3 Public Set, Demonstrating Long-Horizon Autonomous Architecture NVIDIA demonstrated its general-purpose coding agent system, Agentic Variation Operators (AVO), on the ARC-AGI-3 public benchmark, achieving a 100.00 Relative Human Action Efficiency (RHAE) score across all 25 environments and 183 levels. The architecture, previously used for autonomous GPU-kernel engineering, relies on persistent memory and supervision to sustain progress over extended tasks. AVO sustained a continuous seven-day run in GPU-kernel optimization, producing attention kernels that outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 systems. Using Claude Opus 5, AVO solved all 183 levels of the ARC-AGI-3 public benchmark in 6,624 environment actions, roughly 12% fewer actions than reported by VISTA. The system operates with a text-only interface of exact 64 x 64 text grids rather than image rendering or dedicated programmatic world models. Persistent memory and an active supervisory monitor enable AVO to preserve state, recover from incorrect assumptions, and avoid stagnation across long execution horizons. AVO Architecture and GPU-Kernel Optimization AVO replaces predefined evolutionary variation steps with an autonomous agent loop that forms hypotheses, edits code, runs execution tests, and commits candidates. In an attention-kernel experiment on NVIDIA DGX B200 systems, AVO ran autonomously for seven days across more than 500 optimization paths, yielding 40 committed kernel versions that beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%. It subsequently adapted the kernel to grouped-query attention in roughly 30 minutes. To maintain sustained work, AVO pairs persistent memory—which retains compiler output, profiler metrics, and prior implementations—with a supervisor that tracks search trajectories and intervenes when exploration stalls. ARC-AGI-3 Benchmark Performance NVIDIA ported the same underlying architecture to the ARC-AGI-3 interactive benchmark, which tests exploration and rule discovery in unfamiliar environments. Running Claude Opus 5 with a direct-interaction approach and pure 64 x 64 text grids, AVO scored 100.00 RHAE on the 25-environment public set, completing all 183 levels in 6,624 actions compared to 7,542 actions reported by VISTA. In limited tests with GPT-5.6 Sol, Sol demonstrated faster wall-clock completion on matched levels, while Opus required fewer environment actions. These evaluations reflect the public test set rather than semi-private or fully private competition sets. System-Level Generalization The transition from GPU engineering to ARC-AGI-3 illustrates that long-horizon agent performance depends on the surrounding system harness rather than model capability alone. Despite differences in feedback—compilers and profilers versus environment state transitions—both settings rely on forming hypotheses, acting through interfaces, observing outcomes, and updating persistent state. Generality stems from compounding feedback loops, tool grounding, and state retention rather than domain-specific models.

NVIDIA推出通用长时程智能体架构AVO，在ARC-AGI-3公开集斩获满分 NVIDIA介绍了通用编码智能体系统Agentic Variation Operators（AVO），该架构依托持久化记忆与监督机制应对多步骤长时程任务。在完成高性能GPU算子优化实验后，AVO在ARC-AGI-3公开集测试中以100.00 RHAE得分通过全部183个关卡，展现了智能体系统层架构对模型能力的转化作用。 AVO将传统进化搜索的预设变异替换为自主决策智能体，通过持久化记忆和轨迹监督机制保障长期运行稳定性。 在NVIDIA DGX B200上的注意内核研究中，AVO连续运行7天探索500多种优化方向，多头注意力算子性能超越cuDNN达3.5%、超越FlashAttention-4达10.5%。 在ARC-AGI-3公开集所有25个环境中，AVO搭载Claude Opus 5实现100.00 RHAE满分并完成全部183个关卡，总计使用6,624次环境动作，比VISTA系统减少约12%。 与ARC Prize此前公布的Claude Opus 5约30%得分相比，AVO的成绩印证了完整的智能体外围系统（Harness）对最终效能具有决定性作用。 AVO架构与长时程维持机制 AVO是NVIDIA开发的通用编码智能体系统。针对前沿语言模型无法单次解决复杂扩展任务的局限，AVO建立了完整的系统级外围框架（Harness）。其核心包含两项关键机制：持久化记忆负责传递过往代码、编译器和分析器输出及推理积累，避免重复重建搜索路径；监督模块则负责监控全局轨迹，识别停滞或无效循环并在必要时介入重定向，从而维系长达数日的自主工作闭环。 从GPU算子优化验证工程闭环 AVO早期在GPU内核优化中接受了高难度检验。在NVIDIA DGX B200系统上针对多头注意内核的测试中，AVO连续运行7天，自主测试了逾500个方向并最终提交40个内核版本，所获内核性能较cuDNN高出最高3.5%，较FlashAttention-4高出最高10.5%。随后，该系统仅用约30分钟便将成果适配至分组查询注意力（GQA），证明了其持续自主迭代工程方案的能力。 跨领域适配：ARC-AGI-3基准测试与对比 随后NVIDIA将相同的AVO架构迁移至互动推理基准ARC-AGI-3中。智能体需在无规则、无目标指示的未知环境中通过交互推断动态机制。AVO采用纯文本网格交互，搭载Claude Opus 5以6,624次环境动作完成了公开集全部25个环境的183关，获得100.00 RHAE满分。在非控制变量横向对比中，其动作消耗较VISTA报告的7,542次减少约12%；而在部分子集测试中，搭配GPT-5.6 Sol则在部分关卡展现出更快的耗时表现。 智能体系统设计的方法论启示 从硬件编译优化跨越至无指令游戏推理，AVO证明了通用性不仅源于领域知识，更源于能够形成复合推理与反馈循环的系统机制。ARC Prize此前记录Claude Opus 5在该基准上得分约为30%，而AVO通过系统层设计取得满分，突显出基准测试反映的是包含记忆、工具、反馈与恢复机制的完整智能体堆栈效能，而非单纯的模型参数能力。

A frontier language model is only one component of an AI agent . The surrounding agent system—often called a harness—determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks. The challenge is how to build the agent architecture that makes frontier language models work reliably on extended, multistep tasks. The NVIDIA research project Agentic Variation Operators (AVO) addresses this challenge of building a general-purpose agent architecture. AVO was first demonstrated on difficult software engineering and GPU-kernel optimization tasks, where success requires far more than generating code in a single response. An agent must inspect existing implementations, form hypotheses, make changes, execute hardware-grounded tests, interpret feedback, and repeatedly revise its approach. This post introduces the AVO architecture and the system-level mechanisms that enable it to sustain long-running autonomous work. We first summarize its application to GPU-kernel optimization, then describe how the same architecture was adapted to the ARC-AGI-3 benchmark and present the resulting public-set performance, efficiency, and key lessons. What is AVO? AVO is a general-purpose coding agent system developed by NVIDIA. Like modern coding agents, AVO can inspect and edit code, run commands, consult documentation, and validate its work through execution. Its distinguishing focus is sustained autonomous operation across long horizons. In our GPU-kernel optimization work, AVO replaces the predefined variation step of conventional evolutionary-search systems with an autonomous agent that decides how to generate the next candidate—what to inspect, what to change, what to test, and what to commit. For ARC-AGI-3, we connected the same general-purpose agent to a different task interface. The underlying agent remains the same; only the environment-specific tools and evaluation change. Figure 1. AVO architecture for long-horizon autonomous agent work. The main agent iteratively inspects context, plans, implements changes, and evaluates results using persistent memory and tools, while a supervisor monitors the broader search trajectory and can intervene when progress stalls Improving performance with autonomous GPU-kernel optimization GPU-kernel optimization provided an early demanding test of this architecture. The search space is large, the performance landscape is difficult to reason about directly, and small implementation changes can affect correctness, memory behavior, scheduling, and throughput in ways that are hard to predict without execution. In our attention-kernel study, AVO operated continuously for seven days, 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 evaluated configurations. The agent subsequently adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work. Beyond the final kernel performance, this experiment demonstrates that AVO can sustain a productive engineering loop over many iterations without requiring each step to be manually prescribed. More broadly, AVO reinforces a system-level view of agent design: building a trusted agent stack requires performance, reliability, and security to be designed across the full system—not treated as properties of the model alone. This same principle also shapes how we think about securing the AI agent stack. Sustaining long-running agentic work AVO is designed to preserve progress beyond a single model context. Two mechanisms are particularly important: persistent memory and supervision. Persistent memory carries forward prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning, allowing the agent to resume from the current state rather than repeatedly reconstructing the search. The supervisor monitors the broader trajectory for stagnation or repeated unproductive cycles and can redirect the main agent toward alternative strategies when needed. During the seven-day attention-kernel run, the main agent remained responsible for deciding what to inspect, change, test, and evaluate, while the supervisor helped maintain forward progress when the search plateaued. Evolving from high-performance engineering to general-purpose reasoning The NVIDIA team recently applied the same underlying AVO architecture to a very different challenge: the interactive reasoning benchmark ARC-AGI-3 , in which agents enter unfamiliar environments without instructions, stated rules, or stated goals. AVO achieved a 100.00 RHAE score across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels. The result illustrates a broader point: evaluating a model is not the same as evaluating an agent. Model capability matters enormously, but the surrounding system determines how effectively that capability can be converted into sustained autonomous progress. GPU-kernel optimization and the ARC-AGI-3 benchmark appear very different on the surface. One involves source code, compilers, profilers, and throughput. The other involves unfamiliar interactive environments in which an agent must infer the effects of available actions, discover objectives, and act efficiently enough to make progress. But the underlying computational pattern is similar. In both settings, the agent must: Build hypotheses from incomplete evidence Take actions through an external interface Observe the consequences Preserve useful state Revise its model of the problem Recover from incorrect assumptions Continue making progress over a long horizon The domain changes. The feedback channel changes. The core agent loop does not. What transfers is not domain knowledge, but the machinery for sustained autonomous progress. For this reason, ARC-AGI-3 provided a useful test of whether the AVO architecture was fundamentally tied to software engineering or captured something more general. Figure 2. AVO achieved a 100.00 RHAE score across all 25 game environments in the ARC-AGI-3 public set, completing all 183 levels Evaluating AVO on ARC-AGI-3 ARC-AGI-3 is an interactive reasoning benchmark. An agent enters unfamiliar game-like environments without instructions, explicit rules, or a stated goal. It must explore through interaction, infer the environment’s dynamics and objectives, and plan actions efficiently across progressively harder levels. The benchmark uses Relative Human Action Efficiency (RHAE) , a metric that combines task completion with per-level action efficiency relative to first-time human baselines. Performance is aggregated across levels and environments. This metric makes ARC-AGI-3 a demanding long-horizon agent task. Success requires more than solving an isolated state. The agent must preserve useful knowledge, learn from previous interactions, recover from mistakes, and spend environment actions efficiently. Rather than centering our ARC-AGI-3 system on explicit programmatic world-model construction, as explored by Tycho , we adopted the direct-interaction design principles described by VISTA and reimplemented the task interface independently. This was better aligned with our goal of evaluating AVO as a general-purpose agent rather than introducing an ARC-specific world-model layer. Several elements of the task interface were informed by VISTA, but the agent backend was fundamentally different. VISTA instantiates the harness with Claude Opus 5 through Claude Code or GPT-5.6 Sol through Codex, whereas our system uses AVO, the NVIDIA long-horizon agent architecture with persistent memory, supervision, and its own execution loop. The observation interface also differs. VISTA’s primary configuration uses a rendered 512 x 512 PNG, while also exploring textual-grid representations. In the AVO configuration, the LLM operated in a text-only modality: each observation was supplied as an exact 64 x 64 text grid, with no images or image tokens sent to the model. Consistent with the direct-interaction setup, the agent received the available actions without descriptions of the game’s rules or goals and had to infer their effects through interaction. AVO performance results on the ARC-AGI-3 benchmark Recent ARC-AGI-3 systems have explored substantially different agent architectures, from explicit executable world models such as Tycho to direct-interaction harnesses such as VISTA. Together, these results highlight that benchmark performance reflects the complete agent system, not only the underlying model. Using Claude Opus 5, AVO completed the full 25-environment public set with a 100.00 RHAE score, solving all 183 levels in 6,624 environment actions. For reference, VISTA reports 7,542 environment actions with Claude Opus 5 while completing the same 183 public-set levels. AVO therefore used approximately 12% fewer actions in this cross-system comparison. This should not be interpreted as a controlled ablation: the two systems differ in agent backend, observation representation, memory, context management, and other implementation details. One architectural difference that may matter over long horizons is the AVO memory system, which is designed to carry useful understanding forward and reduce repeated exploration, although this experiment does not isolate its individual contribution. ARC Prize separately reports approximately 30% for Claude Opus 5 at High reasoning effort. Our run used the same model family under a different reasoning setting and a substantially different agent system and evaluation setup. These numbers therefore should not be interpreted as a direct measurement of the performance contribution of AVO; rather, they illustrate that model-level evaluation alone does not characterize the performance of a complete agent. AVO is also designed to operate across frontier models. While our full public-set result used Claude Opus 5, we additionally paired AVO with GPT-5.6 Sol on a challenging subset of games. In these limited experiments, Sol reached matched levels faster in wall-clock time in several cases, while Opus used fewer environment actions in matched-level comparisons. These preliminary results suggest complementary operating profiles across models, and we leave a broader systematic comparison to future work. These results cover the 25-environment ARC-AGI-3 public set using the official scorecard and RHAE metric. They are not results on the semi-private or fully private competition sets. What we learned from benchmarking AVO on ARC-AGI-3 The most important result was not simply the 100.00 score, but that the same agent architecture transferred from highly specialized GPU-kernel optimization to a very different interactive reasoning task. In GPU optimization, feedback comes from compilers, tests, profilers, and performance benchmarks. In ARC-AGI-3, feedback comes from environment transitions and action outcomes. The interfaces differ, but the loop is the same: form a hypothesis, act, observe evidence, update state, and continue. This suggests that generality can come not only from domain knowledge, but from the machinery that allows reasoning and feedback to compound over time. More broadly, long-horizon capability is a property of the full system. Memory determines what survives, tools determine what actions are possible, feedback grounds progress, and recovery allows work to continue beyond a single model invocation. Looking ahead NVIDIA AVO research began with autonomous software engineering and high-performance GPU-kernel optimization. ARC-AGI-3 shows that the same underlying architecture can transfer to a very different reasoning environment. The larger opportunity is to build general-purpose agent systems around persistent state, tool use, grounded feedback, recovery, and long-horizon context management—systems that can accumulate evidence and sustain progress across increasingly diverse tasks. The model matters, but the model is not the entire agent. To learn more, check out these resources: Read the paper, AVO: Agentic Variation Operators for Autonomous Evolutionary Search Explore the ARC-AGI-3 benchmark Review the ARC-AGI-3 scoring methodology Browse related work VISTA: A Visual Harness For Reasoning in an Interactive World Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3 Editor’s note: We updated the wording to more precisely distinguish the ARC-AGI-3 public set from the semi-private and private competition sets.