NVIDIA AVO Achieves a 100.00 RHAE Score on ARC-AGI-3

A frontier language model is only one part of an AI agent. The surrounding system, often called a harness, determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and continues working on long-running tasks. Building this architecture is essential for making frontier models reliable on extended, multistep problems.

NVIDIA's research project, Agentic Variation Operators (AVO), addresses this challenge with a general-purpose agent architecture. AVO was initially demonstrated on software engineering and GPU-kernel optimization tasks, where success requires more than producing code in a single response. The agent must inspect implementations, form hypotheses, modify code, run hardware-grounded tests, interpret results, and revise its approach repeatedly.

The same architecture was later adapted to the interactive reasoning benchmark ARC-AGI-3. This article describes AVO's design, its use in GPU-kernel optimization, and its performance and efficiency on the ARC-AGI-3 public set.

What Is AVO?

AVO is a general-purpose coding agent system developed by NVIDIA. Like other modern coding agents, it can inspect and edit code, run commands, consult documentation, and validate its work through execution. Its distinguishing focus is sustained autonomous operation over long horizons.

In NVIDIA's GPU-kernel optimization work, AVO replaces the predefined variation step used by conventional evolutionary-search systems. An autonomous agent decides how to generate each candidate, including what to inspect, change, test, and commit. For ARC-AGI-3, the same general-purpose agent was connected to a different task interface. The underlying agent remained the same, while the tools and evaluation process changed for the environment.

The architecture uses an iterative loop in which the main agent inspects context, plans, implements changes, and evaluates results. Persistent memory and tools support the process, while a supervisor monitors the broader search trajectory and can intervene when progress stalls.

GPU-Kernel Optimization

GPU-kernel optimization provided an early demanding test of AVO. The search space is large, performance is difficult to predict directly, and small implementation changes can affect correctness, memory behavior, scheduling, and throughput. Execution is therefore necessary to determine whether a proposed optimization is effective.

In an 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 later adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work.

The experiment showed that AVO could maintain a productive engineering loop across many iterations without requiring every step to be manually specified. It also illustrated that agent performance, reliability, and security depend on the complete system rather than the model alone.

Supporting Long-Running 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. This allows the agent to resume from the current state instead of repeatedly reconstructing its search.

The supervisor monitors the broader trajectory for stagnation and repeated unproductive cycles. When necessary, it can redirect the main agent toward alternative strategies. 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 progress when the search plateaued.

From GPU Engineering to General-Purpose Reasoning

NVIDIA applied the same underlying AVO architecture to ARC-AGI-3, an interactive reasoning benchmark in which agents enter unfamiliar environments without instructions, stated rules, or explicit goals.

AVO achieved a 100.00 RHAE score across all 25 environments in the ARC-AGI-3 public set and completed all 183 levels. The result demonstrates that evaluating a model is different from evaluating an agent. Model capability is important, but the surrounding system determines how effectively that capability becomes sustained autonomous progress.

GPU-kernel optimization and ARC-AGI-3 appear very different. One involves source code, compilers, profilers, and throughput. The other involves unfamiliar interactive environments where the agent must infer action effects, discover objectives, and act efficiently.

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 understanding of the problem
• Recover from incorrect assumptions
• Continue making progress over a long horizon

The domain and feedback channel change, but the core agent loop remains the same. What transfers is not necessarily domain knowledge, but the machinery for sustained autonomous progress.

ARC-AGI-3 therefore provided a test of whether AVO was tied specifically to software engineering or represented a more general agent architecture.

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 increasingly difficult 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. Results are aggregated across levels and environments.

This makes ARC-AGI-3 a long-horizon task. Success requires more than solving one state. The agent must retain useful knowledge, learn from previous interactions, recover from mistakes, and use environment actions efficiently.

Instead of centering the system on explicit programmatic world-model construction, as explored by Tycho, NVIDIA adopted direct-interaction design principles described by VISTA and independently reimplemented the task interface. This approach was intended to evaluate AVO as a general-purpose agent without adding an ARC-specific world-model layer.

Some task-interface elements were informed by VISTA, but the agent backend was fundamentally different. VISTA uses Claude Opus 5 through Claude Code or GPT-5.6 Sol through Codex. NVIDIA's system uses AVO, with persistent memory, supervision, and its own execution loop.

The observation interfaces also differ. VISTA's primary configuration uses a rendered 512 x 512 PNG while also exploring textual-grid representations. In the AVO configuration, the language model operated in a text-only modality. Each observation was supplied as an exact 64 x 64 text grid, and no images or image tokens were sent to the model. The agent received the available actions without descriptions of the game's rules or goals and had to infer their effects through interaction.

ARC-AGI-3 Performance Results

Recent ARC-AGI-3 systems have used different agent architectures, including explicit executable world models such as Tycho and direct-interaction harnesses such as VISTA. These results demonstrate 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. VISTA reports 7,542 environment actions with Claude Opus 5 while completing the same 183 public-set levels. In this cross-system comparison, AVO used approximately 12% fewer actions.

This is not a controlled ablation. The systems differ in agent backend, observation representation, memory, context management, and other implementation details. AVO's memory system may affect long-horizon performance because it is designed to preserve useful understanding and reduce repeated exploration, but this experiment does not isolate its individual contribution.

ARC Prize separately reports approximately 30% for Claude Opus 5 at High reasoning effort. NVIDIA's run used the same model family under a different reasoning setting and a substantially different agent system and evaluation setup. The figures should therefore not be interpreted as a direct measurement of AVO's performance contribution. They instead show why model-level evaluation does not fully characterize a complete agent.

AVO is also designed to operate across frontier models. Although the full public-set result used Claude Opus 5, NVIDIA 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. The preliminary results indicate different operating profiles, while a broader systematic comparison remains separate work.

These results cover the 25-environment ARC-AGI-3 public set using the official scorecard and RHAE metric. They do not cover the semi-private or fully private competition sets.

Lessons From ARC-AGI-3

The central result was not only the 100.00 score. It was that the same agent architecture transferred from specialized GPU-kernel optimization to a substantially 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 remains the same: form a hypothesis, act, observe evidence, update state, and continue.

This suggests that generality can arise not only from domain knowledge, but also from mechanisms that allow reasoning and feedback to accumulate over time.

More broadly, long-horizon capability is a property of the full system. Memory determines what persists, tools determine which actions are possible, feedback grounds progress, and recovery allows work to continue beyond a single model invocation.

Looking Ahead

NVIDIA's AVO research began with autonomous software engineering and high-performance GPU-kernel optimization. ARC-AGI-3 indicates that the same underlying architecture can transfer to a different reasoning environment.

The broader direction is the development of general-purpose agent systems built around persistent state, tool use, grounded feedback, recovery, and long-horizon context management. Such systems can accumulate evidence and sustain progress across increasingly diverse tasks.

The model remains important, but it is not the entire agent.