Nvidia agent system reaches perfect score on interactive reasoning test

Nvidia agent system reaches perfect score on interactive reasoning test

The same architecture first shown on graphics processor code optimization also finished every public level of a demanding benchmark that requires discovering rules through trial and error.

GP
Giulio Prisco
Aug 24, 2026
2 min read

Nvidia researchers have introduced a general-purpose agent system called Agentic Variation Operators, or AVO. An agent system surrounds a large language model (LLM) so the LLM can use tools, keep track of past steps, recover from mistakes, and keep working on long tasks. AVO focuses on sustained autonomous work over many steps (long-horizon tasks). The system was first shown on graphics processor kernel optimization, where it ran for seven days, tried more than five hundred directions, and produced kernels that ran up to 10.5 percent faster than leading methods on Nvidia hardware.

The same AVO design was then connected to a different test called ARC-AGI-3. ARC-AGI-3 is an interactive reasoning benchmark in which an agent enters unfamiliar game-like environments with no instructions, no listed rules, and no stated goal. The agent must explore by taking actions, learn how the environment works, discover objectives, and complete levels while using as few actions as possible. Success is measured by Relative Human Action Efficiency, or RHAE, a score that rewards both finishing the levels and doing so efficiently compared with human baselines.

How the agent keeps working over time

Two parts of AVO help it continue for a long time. Persistent memory stores earlier code versions, test results, and reasoning so the agent can pick up where it left off instead of starting over. A supervisor watches the overall progress and can steer the main agent if it begins repeating unproductive steps. On the ARC-AGI-3 public set of twenty-five environments, AVO using the Claude Opus 5 language model completed all one hundred eighty-three levels and earned a perfect 100.00 RHAE score. A standalone evaluation of Claude Opus 5 scores far lower, around thirty percent, showing that the surrounding agent system matters greatly.

The main lesson is that the same core loop of forming a hypothesis, acting, observing results, updating memory, and continuing can transfer from specialized engineering work to open-ended interactive reasoning. The language model is important, but the full agent system determines how effectively its ability turns into lasting progress.

The researchers had previously described AVO in a paper published in arXiv.

About the Writer

More from Mindplex

Keep reading

Three more ideas worth your time.

Browse News

Discussion

Join the discussion

Sign in to share a response with the community.

Type @ to mention someone Type / or use + to add a block Highlight text, then choose Link
Loading editor

Comments cannot be edited after posting because they become part of the reputation record. Give yours a quick review first.