Simple and Effective Masked Diffusion Language Models
This work addresses the problem of inefficient diffusion models for language generation, offering a more competitive alternative to autoregressive methods, though it appears incremental as it builds on existing diffusion techniques.
The authors tackled the performance gap between diffusion and autoregressive models in language modeling by developing a simple masked discrete diffusion approach with an improved training recipe and a simplified Rao-Blackwellized objective, achieving a new state-of-the-art among diffusion models and approaching autoregressive perplexity on benchmarks.
While diffusion models excel at generating high-quality images, prior work reports a significant performance gap between diffusion and autoregressive (AR) methods in language modeling. In this work, we show that simple masked discrete diffusion is more performant than previously thought. We apply an effective training recipe that improves the performance of masked diffusion models and derive a simplified, Rao-Blackwellized objective that results in additional improvements. Our objective has a simple form -- it is a mixture of classical masked language modeling losses -- and can be used to train encoder-only language models that admit efficient samplers, including ones that can generate arbitrary lengths of text semi-autoregressively like a traditional language model. On language modeling benchmarks, a range of masked diffusion models trained with modern engineering practices achieves a new state-of-the-art among diffusion models, and approaches AR perplexity. We provide the code, along with a blog post and video tutorial on the project page: https://s-sahoo.com/mdlm