NVIDIA AVO Agent Achieves 100% on ARC-AGI-3 Benchmark

A frontier language model represents only one component of an AI agent, requiring a surrounding harness to manage context, tool usage, and failure recovery. [1] The NVIDIA research project Agentic Variation Operators (AVO) targets the challenge of building a general-purpose agent architecture for reliable performance on multistep tasks. [1] AVO is a coding agent system that can inspect and edit code, run commands, consult documentation, and validate work through execution. [1]

AVO was initially demonstrated on software engineering and GPU-kernel optimization tasks. [1] In these optimization tasks, the system replaces predefined variation steps of traditional evolutionary search with an autonomous agent that decides what to inspect, change, test, and commit. [1] For the ARC-AGI-3 benchmark, researchers connected this same general-purpose agent to a different task interface. [1]

NVIDIA's achievement suggests that general-purpose agent architectures can be successfully adapted across diverse challenges like GPU-kernel optimization and the ARC-AGI-3 benchmark simply by altering the task interface. Unlike approaches relying solely on raw frontier language models, AVO emphasizes the importance of the surrounding agent harness for managing state and feedback during long-running autonomous work. What the sources don't address: How AVO's computational cost or execution time on ARC-AGI-3 compares to human baselines or other state-of-the-art agent frameworks.