NVIDIA AVO Smashes ARC-AGI-3 Benchmark with Perfect Score
NVIDIA's Agentic Variation Operators (AVO) research project has achieved a major breakthrough in long-horizon autonomous agent systems. AVO reached a perfect Relative Human Action Efficiency (RHAE) score of 100% on the ARC-AGI-3 benchmark, completing all 183 levels across 25 environments.
This milestone underscores the potential of advanced agent systems to tackle complex tasks that traditional AI models struggle with. Unlike standalone models, AVO integrates persistent memory, supervision, and a feedback-driven architecture to sustain ongoing progress over extended tasks.
NVIDIA adapted its AVO system from GPU-kernel optimization to this diverse and complex reasoning benchmark, further proving its versatility. The company's success demonstrates a significant leap in the development of autonomous AI agents, which could be applied to various use cases, including software engineering and robotics.
The achievement also signals a shift in how AI systems are evaluated. Traditional benchmarks often focus on isolated model performance, but ARC-AGI-3 highlights the importance of the full agent system, the “harness” around the model. Persistent memory, recovery mechanisms, and real-world adaptability are becoming central to advancing AI toward general-purpose intelligence.