NVIDIA's AVO agent sweeps ARC-AGI-3 with a perfect score

NVIDIA just demonstrated that when it comes to autonomous AI agents, the outer harness might matter even more than the raw intelligence of the underlying model.

Using a general-purpose agent architecture dubbed AVO (Agentic Variation Operators), NVIDIA pushed Claude Opus 5 to a flawless 100% score across all 25 environments and 183 levels of the ARC-AGI-3 public set. For context, standalone Claude Opus 5 scored around 30% under high reasoning effort in standalone benchmarks.

From seven-day GPU tuning to blindfolded puzzle solving

AVO was originally built for a completely different and notoriously painful engineering task: autonomous GPU kernel optimization.

Running on NVIDIA DGX B200 hardware, AVO spent seven straight days experimenting with more than 500 optimization strategies and committing 40 kernel versions. The resulting multihead attention kernels beat FlashAttention-4 by up to 10.5% and cuDNN by 3.5%. It then adapted those findings to grouped-query attention in just 30 minutes.

Instead of treating AVO as a specialized code-writing script, NVIDIA plugged the exact same core architecture into ARC-AGI-3.

ARC-AGI-3 drops an AI agent into interactive environments with zero instructions, no stated rules, and no goals. The agent has to poke at controls, infer what happens, deduce the objective, and clear increasingly hard stages.

NVIDIA realized that optimizing raw compute and blind-solving interactive puzzles share the exact same engineering loop: form a hypothesis from partial data, execute an action, parse feedback, retain state, recover from bad assumptions, and keep grinding over long horizons.

Pure text grids and a vigilant supervisor

While previous systems like VISTA relied heavily on 512x512 rendered images, AVO interacted with ARC-AGI-3 purely in text, passing each observation as a 64x64 text grid with zero vision tokens.

AVO solved the entire 183-level benchmark in 6,624 environment actions. That is roughly 12% fewer actions than VISTA, which required 7,542 steps using the same Claude Opus 5 engine.

The real secret sauce boils down to persistent memory and an active supervisor module. The persistent memory system stops the agent from forgetting previous lessons, while the supervisor monitors overall progress and pulls the agent out of unproductive rabbit holes whenever it gets stuck.

NVIDIA also tested GPT-5.6 Sol on a challenging slice of the benchmark. Interestingly, Sol finished runs faster in terms of wall-clock time, while Opus 5 proved more conservative with its action budget.

Why it matters

This experiment offers concrete proof that simply scaling a frontier model is only half the battle for autonomous work.

For software engineers and agent builders, the takeaway is clear: the scaffolding around the LLM—state preservation, context hygiene, and stagnation detection—is where real capability unlocks happen right now. Turning a 30% benchmark baseline into a 100% completion rate without retraining the core model shows that proper agent infrastructure can dramatically outpace pure raw weights.

For the broader industry, NVIDIA is signaling that high-end compute hardware like the DGX B200 can increasingly be self-optimized. When an autonomous system can spend seven days inventing kernel optimizations that surpass human-crafted baselines, software engineering loops begin to fundamentally accelerate.

We think NVIDIA’s decision to avoid building task-specific world models was a brilliant design choice. By treating the agent harness as a domain-agnostic operational loop, they proved that the machinery of long-horizon reasoning generalizes surprisingly well, even if running these massive feedback loops remains cost-prohibitive for most outside frontier labs.