More about HKUST
A Survey of Diffusion Language Models: From Modeling to Scaling
PhD Qualifying Examination
Title: "A Survey of Diffusion Language Models: From Modeling to Scaling"
by
Mr. Fangyu DING
Abstract:
Autoregressive language models generate text by predicting one token at a time
from left to right. While this sequential formulation is simple and effective,
it introduces an intrinsic inference bottleneck: generating a long sequence
requires thousands of sequential forward passes, and tokens cannot be easily
revised once emitted. Diffusion language models offer a fundamentally
different perspective. Instead of generating tokens sequentially, they
initialize the entire sequence as noise and iteratively refine all positions
in parallel. This formulation provides natural advantages: parallel decoding,
full bidirectional context, and, depending on the noise process, iterative
revision or constraint satisfaction.
This survey provides an accessible synthesis of diffusion language models,
from their continuous-time mathematical foundations through four modeling
paradigms to foundation-scale, hardware-compatible serving. We begin with a
unified continuous-time stochastic framework that covers both continuous
diffusion and discrete Markov chains. Within this framework, we systematically
dissect four foundational modeling paradigms:
(1) Mask-Based Diffusion, which corrupts tokens toward an absorbing mask state
and simplifies score matching into weighted cross-entropy; (2) Insertion-Based
Diffusion, which models variable-length text via continuous-time deletion and
insertion; (3) Uniform-State Diffusion, which substitutes tokens toward
uniform noise; Duo's argmax duality is used for a training curriculum and for
Discrete Consistency Distillation, while generation remains discrete
substitution; and (4) Continuous Latent Diffusion, which embeds discrete
tokens into a low-dimensional codebook and applies continuous Gaussian
diffusion with PF-ODE sampling. For each paradigm, we derive its training
objectives, analyze its theoretical properties, and present its core sampling
mechanics in a clear, step-by-step manner.
Moving from theory to practice, we examine how these formulations scale to
billions of parameters. We trace the shift from full-sequence bidirectional
diffusion, which preserves global context but incurs a memory wall, to block
diffusion, which restores prefix Key-Value caching and enables continued
pretraining from mature autoregressive checkpoints. We further review how
masked diffusion can be reparameterized as purely causal attention for
standard serving engines, together with adaptive block sizing that avoids
wasted computation on already-resolved tokens.
Finally, we discuss key open challenges, including whether more flexible
modeling paradigms can surpass mask-based block diffusion, high- concurrency
serving economics, the factorization gap in parallel decoding, and
post-training without exact sequence likelihoods. By clarifying the core
mechanics and practical trade-offs of diffusion language models, this survey
aims to provide a clear and rigorous foundation for research in
non-autoregressive sequence generation.
Date: Monday, 28 September 2026
Time: 3:00pm - 5:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Binhang Yuan (Supervisor/Chairperson)
Dr. Chaojian Li
Dr. Shuai Wang