That surrounding architecture, often described as an agent harness, determines how a model receives context, interacts with external tools, retains information and responds when an approach fails. NVIDIA's research focuses on improving those system-level components rather than treating the underlying language model as the complete agent.
The distinction was particularly visible in NVIDIA's ARC-AGI-3 testing. ARC Prize separately reports a score of about 30% for Claude Opus 5 at High reasoning effort on the public benchmark. NVIDIA's AVO system, also using Claude Opus 5 but under a different reasoning setting and evaluation setup, reached 100.00.
NVIDIA cautioned against treating those figures as a direct measurement of AVO's contribution. The configurations differ beyond the agent architecture, including reasoning settings and evaluation setup. Instead, the company presented the results as evidence that evaluating a model by itself does not necessarily predict the performance of a complete agent built around it.
ARC-AGI-3 tests agents in unfamiliar interactive environments without providing instructions, explicit rules or stated objectives. The agent must experiment with available actions, determine how the environment works and identify what it needs to accomplish. Performance is measured using Relative Human Action Efficiency, or RHAE, which accounts for both completing levels and the number of actions used compared with first-time human performance.
AVO completed the public benchmark's 183 levels using 6,624 environment actions. NVIDIA compared that result with VISTA, which reports 7,542 actions when using Claude Opus 5 to finish the same public levels. That gives AVO an action count roughly 12% lower in the cross-system comparison.
The company emphasized that the comparison is not a controlled test. AVO and VISTA differ in their agent backends, memory systems, context handling and observation formats, among other implementation choices.
For its ARC-AGI-3 run, AVO also operated entirely through text. Each observation was represented as an exact 64-by-64 text grid, with no images or image tokens provided to the language model. The system received the available actions but had to determine their meaning and the objectives of each environment through interaction.
Persistent memory is one of the central pieces of AVO. It allows information from earlier attempts, evaluations and reasoning to remain available as the agent continues working, reducing the need to reconstruct previous discoveries after a model context ends.
A supervisor provides another layer. It watches the broader trajectory of the task and can redirect the main agent when it becomes stuck or repeatedly follows an unproductive strategy. The primary agent remains responsible for deciding what actions to take and how to respond to results.
NVIDIA initially developed AVO around a substantially different type of work: autonomous GPU-kernel optimization. In an attention-kernel experiment, the agent ran for seven days, investigated more than 500 possible optimization directions and committed 40 kernel versions.
On NVIDIA DGX B200 systems, the resulting multihead attention kernels exceeded cuDNN performance by as much as 3.5% and FlashAttention-4 by as much as 10.5% across the configurations NVIDIA evaluated. AVO then adapted the evolved kernel for grouped-query attention in about 30 additional minutes of autonomous work.
Moving the same architecture from GPU optimization to ARC-AGI-3 allowed NVIDIA to test whether its approach could extend beyond software engineering. Although the two tasks use different tools and forms of feedback, both require an agent to develop hypotheses, act on incomplete information, learn from results, preserve useful discoveries and recover when its assumptions prove incorrect.
NVIDIA also conducted limited ARC-AGI-3 experiments pairing AVO with GPT-5.6 Sol. On a challenging subset of games, GPT-5.6 Sol reached matched levels more quickly by wall-clock time in several cases, while Claude Opus 5 required fewer environment actions in matched-level comparisons. NVIDIA described those findings as preliminary and did not conduct a broader systematic comparison.
The published 100.00 result applies specifically to ARC-AGI-3's 25-environment public set. NVIDIA said it does not represent performance on either the semi-private or fully private competition sets.
The research extends NVIDIA's earlier work with AVO by showing that the architecture can operate across two very different problem types. Its ARC-AGI-3 result also reinforces the distinction between model benchmarks and complete agent systems: for long-running autonomous work, performance depends not only on the underlying model but also on how memory, tools, feedback, context and recovery are organized around it.
This analysis is based on reporting from NVIDIA.
Image courtesy of NVIDIA.
This article was generated with AI assistance and reviewed for accuracy and quality.
