Nvidia's AVO harness boosts Opus 5 to 100% on ARC-AGI 3 (developer.nvidia.com)

Nvidia has announced the successful application of its novel Agentic Variation Operators (AVO) architecture, achieving a remarkable 100% Relative Human Action Efficiency (RHAE) score across all environments in the challenging ARC-AGI-3 benchmark. This marks a significant advancement for the AI/ML community as it demonstrates AVO's versatility beyond its initial focus on GPU-kernel optimization, showcasing its capability in tackling diverse multi-step reasoning tasks where agents interact with complex environments without predefined instructions or goals. The AVO architecture emphasizes sustained autonomous operation through mechanisms like persistent memory and supervisory oversight, allowing agents to build upon prior knowledge and strategize effectively over long periods. In the benchmarks, AVO not only excelled in solving intricate GPU optimization challenges but also adapted seamlessly to the interactive reasoning demands of ARC-AGI-3. The results highlight that successful agent design encompasses more than just advanced models; it hinges on the entire system's capacity for managing context and feedback effectively. This research is poised to influence the development of general-purpose AI agents capable of tackling a wide range of real-world tasks with improved reliability and efficiency.