Memento 3 Achieves Perfect Score on ARC-AGI-3 Interactive Reasoning Benchmark with Frozen Language Model

Researchers at University College London and Huawei's Noah's Ark Lab have announced that their Memento 3 system achieved a maximum score of 100.0 RHAE (Relative Human Action Efficiency) on all 25 public games of the ARC-AGI-3 benchmark, matching human-level performance in an interactive reasoning challenge that has confounded state-of-the-art AI systems. The breakthrough represents a significant milestone in agentic intelligence, a shift in how artificial intelligence capability is measured beyond static pattern recognition.