manuscript

A parameter-efficient latent multi-token prediction objective that bootstraps future-token representations from the model’s own hidden states, adding at most 1% extra parameters while matching the gains of conventional MTP in planning, coding, and reasoning.