A Transformer layer is executed once. Then again. And again. No new parameters — just deeper computation. Looped Transformers show promise on reasoning and math, but one question has never been answered fairly: is the gain from the looping architecture itself, or simply from

→ View original post on X — @jiqizhixin