NVIDIA AVO Agent Achieves 100% Success Rate on ARC-AGI-3 Benchmark
NVIDIA's Agentic Variation Operators (AVO) agent has achieved a 100% success rate on the ARC-AGI-3 benchmark. The system successfully completed all 183 levels across 25 public environments without any prior instructions, explicit rules, or stated goals. Achieving a perfect score on ARC-AGI-3, a benchmark designed to measure general intelligence and adaptability, represents a major milestone in AI agent reasoning. It demonstrates that autonomous agents can solve complex, long-horizon tasks and adapt to novel environments without human intervention. The AVO architecture enables sustained autonomous operation by integrating persistent memory, supervision, and tool-use. It acts as an evolutionary variation operator, allowing the agent to autonomously explore domain knowledge, edit code, and validate results through iterative interaction.
## BACKGROUND
The Abstraction and Reasoning Corpus (ARC-AGI) is a benchmark designed to measure an AI's ability to acquire new skills and solve novel reasoning tasks that are easy for humans but difficult for AI. Traditional AI models often struggle with ARC because it requires abstract reasoning rather than pattern matching from training data. NVIDIA's AVO framework addresses this by using evolutionary search and agentic workflows to autonomously iterate on solutions.