Diffusion language-model glossary

Definitions, common confusions, and primary sources: masking, ELBO, NFE, latency, conditioning, and controllable text generation.

Follow a learning path

Forward process

A specified corruption process that moves clean data toward noise or masks.

Do not confuse it with generating text.

Austin et al. · NeurIPS 2021

Reverse process

Learned transitions that reconstruct data-like states from corrupted states.

This is a distribution, not recovery of one unique original sentence.

Austin et al. · NeurIPS 2021

Absorbing mask

In the forward process a masked position cannot leave the mask state.

This restriction describes forward corruption, not an inability to unmask during generation.

Austin et al. · NeurIPS 2021

Discrete score / score entropy

SEDD models ratios of probabilities of discrete states; score entropy is its training loss.

It is not a derivative with respect to the integer token ID.

Lou, Meng & Ermon · ICML 2024

ELBO / negative ELBO

A lower bound on log likelihood; its negative is an upper bound on negative log likelihood.

A bound and an exact likelihood are not interchangeable.

Austin et al. · NeurIPS 2021 · Sahoo et al. · NeurIPS 2024

Continuous versus discrete state

Diffusion-LM denoises real-valued word representations; categorical diffusion operates on discrete states.

Continuous time does not imply continuous states, and embeddings do not imply a compressed autoencoder latent.

Li et al. · NeurIPS 2022 · Austin et al. · NeurIPS 2021

Network function evaluations (NFE)

The number of network evaluations, not simply the number of scheduled denoising steps.

Caching may avoid evaluations; guidance or a solver may add work. State the counting convention.

Lou, Meng & Ermon · ICML 2024 · Sahoo et al. · NeurIPS 2024 · Li et al. · NeurIPS 2022

Latency versus throughput

Latency is elapsed time for a request; throughput is completed work per unit time.

Compare hardware, length, batching, precision, cache, and quality. NFE alone is insufficient.

Sahoo et al. · NeurIPS 2024 · Arriola et al. · ICLR 2025

Conditioning

Generating with an input sequence or other supplied information.

A conditioning input is not automatically a hard output constraint.

Gong et al. · ICLR 2023

Guidance

Steering generation with an additional signal; Diffusion-LM uses gradients on continuous intermediate states.

Include controller computation and validate the decoded output.

Li et al. · NeurIPS 2022

Hard constraint

A requirement every accepted output must satisfy, such as unchanged protected tokens.

An improved control score alone does not establish this guarantee.

Li et al. · NeurIPS 2022 · Lou, Meng & Ermon · ICML 2024

Remasking versus carry-over

Remasking hides a revealed token again; MDLM carry-over instead copies unmasked inputs.

Do not assume every masked sampler revises all previous decisions.

Sahoo et al. · NeurIPS 2024

Block diffusion

Autoregression across token blocks with diffusion inside each conditional block.

Dependencies between blocks remain sequential.

Arriola et al. · ICLR 2025

Evaluation protocol

The dataset/split, metric definition, decoding settings, and compute conditions behind a result.

If conditions are missing or differ, do not treat numbers as directly comparable.

Sahoo et al. · NeurIPS 2024 · Lou, Meng & Ermon · ICML 2024

Evidence on this site

An attributable source passage supporting a specific claim.

A model interpretation is not an independently measured result.

Core corpus on this site

Work selected as directly relevant to diffusion language modeling.

A relevance label is not a guarantee of scientific quality.