We released Guidance-TTT, which trains a small guidance model at test time while keeping a stronger execution model frozen, for open-ended discovery. Paper · Code.