Nvidia AVO Completes ARC-AGI-3’s Public Game Set
Nvidia says its general-purpose coding agent solved all 183 public ARC-AGI-3 levels, highlighting the growing influence of agent harness design.
A perfect result on the public environments
Nvidia says its AVO coding agent completed all 183 levels across the 25 public ARC-AGI-3 environments, producing a 100% score. The interactive benchmark presents agents with unfamiliar visual games whose objectives must be inferred through exploration, feedback and repeated action rather than supplied as explicit instructions.
AVO is described as a general-purpose coding agent, not a model trained solely to play the benchmark. Its loop continuously inspects the environment, forms a plan, writes or modifies solutions, evaluates the result and stores useful information in memory. Nvidia argues that this surrounding system — often called an agent harness — allowed the agent to build on execution feedback over long trajectories.
The result should be interpreted carefully. ARC-AGI-3’s 25 public environments are visible development examples, and the benchmark’s maintainers distinguish them from its unseen private evaluation set. Other research systems have also reached 100% on the public games. Completing the public set therefore demonstrates that a system can master those released environments; it does not establish a perfect score on the full benchmark or prove general intelligence.
The public games nevertheless offer a useful controlled setting for studying exploration, memory and recovery from mistakes. Nvidia reported completing the set with fewer environment actions than a comparison system, suggesting that orchestration can affect efficiency as well as eventual success. Independent reproduction and performance on private environments will be needed to determine how well the approach generalizes.
Why it matters
Frontier-model comparisons increasingly understate how much application performance comes from the software around the model. Memory policy, tool selection, verification, state tracking and retry logic can turn the same underlying model into markedly different agents.
Nvidia’s result adds evidence for investment shifting toward harnesses and runtime infrastructure, especially for long-duration tasks. It also illustrates why benchmark disclosures must specify whether results concern public development material or hidden tests. For buyers and researchers, the practical question is no longer simply which model scored highest, but which complete system can reproduce reliable performance on genuinely new tasks at an acceptable cost.