NVIDIA AVO Hits 100% on ARC-AGI-3: System Design Beats Raw Model Power
A coding agent built for GPU kernel optimization just scored perfect on an interactive reasoning benchmark it was never designed for — revealing that agent architecture is the real frontier multiplier
Listen to this article
NVIDIA's AVO agent scored a perfect 100.00 on the ARC-AGI-3 benchmark, completing all 183 levels across all 25 public environments — and it did so with 12% fewer actions than VISTA, the previous best agent on the benchmark. What makes this result significant: AVO was not designed for reasoning benchmarks. It was built to autonomously evolve GPU kernels, and the ARC-AGI-3 score came from pointing the same architecture at an entirely different problem domain.
The blog post was published on August 21, 2026 on the NVIDIA Developer Blog [Link]. The underlying research paper, "AVO: Agentic Variation Operators for Autonomous Evolutionary Search," was published on arXiv in March 2026 [Link].
The underlying model powering AVO is Claude Opus 5. Running that same model directly on ARC-AGI-3 without the AVO agent harness yields a score of 30.2%. The agent system delivers a 3.3x multiplier over the bare model — a stark illustration that evaluating a model is not the same as evaluating an agent.
What Is ARC-AGI-3?
ARC-AGI-3 is an interactive reasoning benchmark where an agent enters unfamiliar, game-like environments with no instructions, no explicit rules, and no stated goals. The agent must infer dynamics and objectives purely through interaction — observing the environment, trying actions, learning from outcomes, and building an internal model of how the world works.
The benchmark measures Relative Human Action Efficiency (RHAE) — how efficiently the agent solves tasks compared to human baselines. A score of 100 means the agent matched or exceeded human-level efficiency across the entire test set. The public set contains 183 levels spread across 25 distinct environments, each with its own hidden rules and mechanics.
This is fundamentally different from standard LLM benchmarks. There's no multiple-choice format, no fixed context window, and no single correct answer. The agent must sustain reasoning over long horizons, recover from dead ends, and adapt its strategy as it learns more about each environment.
Public vs. Private Set
AVO's 100% score is on the public set — not the private evaluation set. As Hacker News commenters noted, this carries the inherent risk of overfitting. The private set results, which would confirm genuine generalization, have not been published. This is an important caveat: the result is impressive but not yet independently verified against unseen data.
The AVO Architecture: Four Steps, Two Layers
AVO stands for Agentic Variation Operators. The architecture wraps a frontier LLM in an iterative loop with two critical additions: persistent memory and a supervisory layer. The core cycle is simple — inspect, plan, implement, evaluate — but the infrastructure around it is what enables long-horizon autonomy.
The Main Agent Loop
The main agent operates in four phases that repeat continuously:
- Inspect — The agent examines the current state of the environment, reviewing available context, prior results, and accumulated knowledge from memory.
- Plan — Based on the inspection, the agent formulates a strategy for the next step. This is where reasoning, hypothesis formation, and decision-making happen.
- Implement — The agent executes its plan using available tools, making changes to the environment or running experiments.
- Evaluate — The agent assesses the results of its implementation, determining whether progress was made and what was learned.
After evaluation, the loop restarts. The agent doesn't start fresh each cycle — it builds on what it learned, using persistent memory to carry forward prior implementations, evaluation results, and accumulated knowledge.
Persistent Memory
This is the component that distinguishes AVO from standard agentic frameworks. Most agent systems lose context between calls — each LLM invocation starts with a prompt, generates a response, and forgets everything. AVO's persistent memory ensures that prior attempts, failed experiments, successful patterns, and execution feedback all survive across iterations.
The agent can consult its current lineage (the chain of prior attempts), a domain-specific knowledge base, and execution feedback. This means it doesn't repeat the same mistakes, doesn't re-explore dead ends, and can build increasingly sophisticated strategies as it accumulates evidence.
The Supervisor Layer
The second critical component is the supervisor — a separate process that monitors the main agent's broader search trajectory. When the main agent makes progress, the supervisor stays out of the way. When progress stalls, the supervisor intervenes, potentially redirecting the agent toward new strategies or helping it course-correct.
This two-layer design — a focused executor and a strategic overseer — mirrors how human teams work. The supervisor handles meta-reasoning about whether the current approach is productive, while the main agent focuses on the tactical work of solving the immediate problem.
Why This Matters
NVIDIA's key claim: "System design — not model capability alone — can unlock frontier-level long-horizon performance." The same Claude Opus 5 model goes from 30.2% (bare) to 100% (wrapped in AVO). That's not a model improvement; it's an architecture improvement.
Originally Built for GPU Kernel Evolution
ARC-AGI-3 was not AVO's original target. AVO was designed for autonomous evolutionary search of GPU kernels — specifically, optimizing attention implementations for NVIDIA hardware.
In the attention-kernel study described in the original paper, AVO operated continuously for seven days on NVIDIA DGX B200 systems. Over that period, it explored more than 500 optimization directions and produced 40 committed kernel versions for multihead attention. The resulting kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across evaluated configurations.
AVO also adapted its evolved multihead attention kernel to grouped-query attention in just 30 minutes of additional autonomous work, yielding gains of up to 7.0% over cuDNN and 9.3% over FlashAttention-4 for that variant.
The paper replaces manually designed mutation, crossover, and hand-crafted heuristics from classical evolutionary search with an autonomous coding agent that proposes, repairs, critiques, and verifies implementation edits. Rather than confining the LLM to candidate generation within an evolutionary pipeline, AVO elevates the agent to the role of variation operator itself.
From Kernels to Reasoning: The Generalization Test
Here's what makes the ARC-AGI-3 result interesting: NVIDIA took the same underlying AVO architecture — persistent memory intact, supervisor mechanism intact — and pointed it at an interactive reasoning benchmark without redesigning the system. The agent figured out what to do in each environment with no instructions, no explicit rules, and no stated goals.
This is a strong signal of genuine generality. Most agent systems are purpose-built for specific domains. AVO's ability to transfer from GPU kernel optimization to interactive reasoning suggests the architecture captures something fundamental about autonomous problem-solving that transcends domain-specific knowledge.
The comparison to VISTA is particularly telling. VISTA was designed specifically for ARC-AGI-3-style reasoning tasks. AVO, built for kernel evolution, matched its performance while using fewer actions. This suggests that a well-designed general-purpose agent loop can outperform specialized architectures on tasks they weren't optimized for.
The Public Set Caveat
The ARC-AGI-3 public set has been available for some time. As with any benchmark, there's a risk that agents can overfit to its specific environments. NVIDIA has not yet published results on the private evaluation set. The true test of generalization will come when AVO is evaluated against unseen environments. Until then, the 100% score should be viewed as a strong indicator rather than definitive proof of frontier-level generality.
Broader Implications: The Agent Architecture Arms Race
AVO's result points to a shift in where competitive advantage is being created in AI. The model layer — raw LLM capability — is becoming increasingly commoditized. What differentiates systems is the architecture that wraps the model: how context is managed, how tools are orchestrated, how memory persists, and how supervision guides exploration.
Three implications stand out:
1. The Model Is No Longer the System
Evaluating a model in isolation tells you nothing about what it can do when embedded in a capable agent system. Claude Opus 5 at 30.2% on ARC-AGI-3 is not the same system as Claude Opus 5 inside AVO at 100%. Future benchmark comparisons that don't account for the agent harness will be misleading.
2. Persistent Memory Is the Differentiator
The ability to carry forward knowledge across iterations — avoiding repeated mistakes, building on partial successes, and accumulating domain expertise — is what enables long-horizon performance. Agent systems without persistent memory are fundamentally capped in how far they can go.
3. Supervision Enables Robustness
The supervisor layer solves a critical problem: agents get stuck. Without meta-reasoning about whether the current approach is productive, agents waste compute on unproductive strategies. The supervisor's ability to detect stagnation and redirect is what turns a competent agent into a resilient one.
NVIDIA's message is clear: the frontier in autonomous agents is no longer about bigger models. It's about better architectures — systems that give models the memory, tools, feedback loops, and supervision they need to sustain progress across long-running, open-ended tasks.
Key Sources and References
- NVIDIA Developer Blog — "NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents" [Link]
- arXiv Paper — "AVO: Agentic Variation Operators for Autonomous Evolutionary Search" (2603.24517) [Link]
- NVIDIA AI on X — Official announcement thread [Link]
- Hacker News Discussion — Community analysis including the public vs. private set caveat [Link]
- The New Stack — "Claude Opus 5 scored 30% on ARC-AGI-3. Wrapped in Nvidia's AVO, it hit 100%." [Link]
- explainx.ai — "NVIDIA AVO: 100% ARC-AGI-3 Score, Public Set Only" — detailed breakdown [Link]
- NVIDIA AI on X — Architecture explanation: continuous inspect-plan-implement-evaluate loop [Link]