NVIDIA's AVO agent aces ARC-AGI-3: perfect score, zero instructions
NVIDIA's AVO agent scored a perfect 100.00 RHAE on ARC-AGI-3, clearing all 183 levels across 25 public environments with no instructions, rules, or stated goals. The agent figures out each unfamiliar environment by inspecting, acting, and learning from feedback, sustained by memory and long-horizon state rather than fresh context each step. The telling part: the underlying model scored roughly 30% bare, which puts the win on the harness, not the model. It is the argument that evaluating a model is not the same as evaluating an agent, made with a number.