Training_Large_Language_Models_to_Reason_in_a_Continuous_Latent_Space