Multi-Token Prediction Needs Registers

TL;DR

MuToR is a simple, parameter-efficient method that enhances multi-token prediction by interleaving learnable register tokens, improving performance across language and vision tasks without architectural changes.

Contribution

The paper introduces MuToR, a novel approach that integrates learnable register tokens into input sequences, compatible with existing models and effective across various tasks.

Findings

Effective in supervised fine-tuning and PEFT

Supports scalable prediction horizons

Demonstrates versatility in language and vision tasks

Abstract

Multi-token prediction has emerged as a promising objective for improving language model pretraining, but its benefits have not consistently generalized to other settings such as fine-tuning. In this paper, we propose MuToR, a simple and effective approach to multi-token prediction that interleaves learnable register tokens into the input sequence, each tasked with predicting future targets. Compared to existing methods, MuToR offers several key advantages: it introduces only a negligible number of additional parameters, requires no architectural changes--ensuring compatibility with off-the-shelf pretrained language models--and remains aligned with the next-token pretraining objective, making it especially well-suited for supervised fine-tuning. Moreover, it naturally supports scalable prediction horizons. We demonstrate the effectiveness and versatility of MuToR across a range of use cases, including supervised fine-tuning, parameter-efficient fine-tuning (PEFT), and pretraining, on challenging generative tasks in both language and vision domains. Our code will be available at: https://github.com/nasosger/MuToR.