Start with a counterintuitive pair of numbers: the same Claude Opus 5 model scores roughly 30% in ARC Prize's official runs, yet after NVIDIA's research team wrapped it in their AVO agent architecture, it solved all 183 levels of the ARC-AGI-3 public set with a 100.00 RHAE score. The model didn't change at all — only the shell around it did.
What Happened: AVO Clears the ARC-AGI-3 Public Set
On August 21, NVIDIA announced on its technical blog that its long-horizon autonomous agent architecture, Agentic Variation Operators (AVO), passed all 25 environments and 183 levels of the ARC-AGI-3 benchmark with a 100.00 RHAE score, using 6,624 environment actions in total. For comparison, VISTA — an earlier system that completed the same 183 public-set levels with the same Claude Opus 5 model — used 7,542 actions, meaning AVO used roughly 12% fewer.
Two caveats matter: this covers the public set only, not the semi-private or private competition sets; and NVIDIA itself notes the AVO-vs-VISTA comparison is not a controlled ablation — the two systems differ in observation representation, memory management, context management, and other implementation details.
What ARC-AGI-3 Actually Tests
ARC-AGI-3 is an interactive reasoning benchmark: an agent enters completely unfamiliar game-like environments with no instructions, no stated rules, and no stated goal. It must explore through interaction, infer the environment's dynamics and objectives, and plan actions efficiently. The RHAE metric combines task completion with per-level action efficiency relative to first-time human baselines.
AVO's observation channel is particularly interesting: the LLM operated entirely in text — each observation was supplied as an exact 64×64 text grid, with no images or image tokens sent to the model. VISTA's primary configuration, by contrast, used a rendered 512×512 PNG. The same set of tasks, two very different perception channels, and both eventually passed.
The Earlier Track Record: Seven Days of Autonomous GPU-Kernel Optimization
AVO was first validated on software engineering work. In an attention-kernel optimization study, it ran continuously and autonomously for seven days, explored more than 500 optimization directions, and committed 40 kernel versions. On NVIDIA DGX B200 systems, the evolved multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%. The agent then adapted the evolved kernel to grouped-query attention in roughly 30 minutes of additional autonomous work.
Two mechanisms sustain this kind of long-horizon run: persistent memory (carrying forward prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning, so the agent resumes from its current state instead of rebuilding the search) and a supervisor (monitoring the broader trajectory for stagnation or unproductive loops and redirecting the main agent when needed).
The Real Message: Capability Is a Property of the System
NVIDIA's conclusion is blunt: evaluating a model is not the same as evaluating an agent. Model capability matters enormously, but the surrounding system determines whether that capability converts into sustained autonomous progress — memory decides what survives, tools determine which actions are possible, feedback grounds progress in reality, and recovery mechanisms let work continue beyond a single model invocation.
The team also paired AVO with GPT-5.6 Sol on a challenging subset of games: in those limited experiments, Sol reached matched levels faster in wall-clock time in several cases, while Opus used fewer environment actions — complementary operating profiles across models, with a systematic comparison left to future work.
The industry implication: over the next year, the engineering value of the agent-framework layer may get repriced. Everyone is competing on model parameters and raw benchmark scores, but what turned 30% into 100% was memory, supervision, and feedback loops — system design that looks unglamorous. The paper is on arXiv (2603.24517); the original report is on the NVIDIA Technical Blog.
So the next time a model posts an unimpressive score, hold your judgment — it may simply not have met a harness that knows how to use it.