Study tests when looped language models continue improving beyond their trained recurrence depth

The paper asks whether additional inference loops help language models, which task types benefit, and whether gains continue beyond training depth.