Inlay

We evaluate models on two axes: whether models can achieve given tasks with high efficiency (i.e., fewer steps) and whether they can explicitly recover the rule. We find that overall humans still lead by ~13% on success and ~30% on stating the rule.