What's this about?
TaH2 reuses transformer layers but learns which tokens benefit from extra internal passes. Easy tokens can move on while difficult ones receive more latent computation.
What exactly is a looped transformer?
Instead of owning many distinct layers, it sends representations through some of the same layers repeatedly—like rereading a hard sentence without buying a larger book.
Why does this matter?
14 HF upvotes. On AIME, the authors report a 53% improvement in the accuracy-compute slope and continued gains at deeper loops where fixed-loop baselines plateau.