Diffusion Language Models Can Perform Many Tasks with Scaling and Instruction-Finetuning

TL;DR

This paper shows that scaling diffusion language models through data, size, and instruction finetuning enables them to perform a wide range of language tasks effectively, rivaling autoregressive models.

Contribution

It introduces a method to adapt pretrained masked language models into diffusion models and demonstrates their scalability and versatility in general language tasks.

Findings

Scaling improves downstream task performance

Instruction finetuning enables zero-shot and few-shot learning

Diffusion models show promising reasoning abilities

Abstract

The recent surge of generative AI has been fueled by the generative power of diffusion probabilistic models and the scalable capabilities of large language models. Despite their potential, it remains elusive whether diffusion language models can solve general language tasks comparable to their autoregressive counterparts. This paper demonstrates that scaling diffusion models w.r.t. data, sizes, and tasks can effectively make them strong language learners. We build competent diffusion language models at scale by first acquiring knowledge from massive data via masked language modeling pretraining thanks to their intrinsic connections. We then reprogram pretrained masked language models into diffusion language models via diffusive adaptation, wherein task-specific finetuning and instruction finetuning are explored to unlock their versatility in solving general language tasks. Experiments show that scaling diffusion language models consistently improves performance across downstream language tasks. We further discover that instruction finetuning can elicit zero-shot and few-shot in-context learning abilities that help tackle many unseen tasks by following natural language instructions, and show promise in advanced and challenging abilities such as reasoning.