NVIDIA’s AI coding agent AVO has achieved a 100% score on the ARC-AGI-3...

NVIDIA’s AI coding agent AVO has achieved a 100% score on the ARC-AGI-3 interactive reasoning benchmark, solving all 183 levels across 25 public environments. AVO operates without predefined rules or explicit goals. It learns by experimenting, observing outcomes, and adjusting its strategy based on past experiences. Unlike previous models, AVO maintains and builds on its learning through context resets rather than starting from scratch. The system is powered by Claude Opus 5, which has previously been used to autonomously optimize GPU code over several days. This result demonstrates AVO’s ability to master environment-specific tasks through repeated interaction and continuous memory updates. 📰 @aipost