Simple and Effective Masked Diffusion Language Models

Subham Sekhar Sahoo, Marianne Arriola, Aaron Gokaslan, Yair Schiff, Edgar Mariano Marroquin, Justin T Chiu, Alexander M Rush, Volodymyr Kuleshov

Keywords: discrete diffusion, masked diffusion, bert, dna modeling, language modeling

TL;DR: A masked diffusion framework featuring a novel objective that enhances log-likelihood estimation and achieves state-of-the-art results, including applications to DNA sequence modeling.

Abstract: While diffusion models excel at generating high-quality images, prior work reports a significant performance gap between diffusion and autoregressive (AR) methods on 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.

Is Neurips Submission: Yes

Submission Number: 35