Diffusion Language Models are Provably Optimal Parallel Samplers

Haozhe Jiang, Nika Haghtalab, Lijie Chen · 2025

Read Diffusion Language Models are Provably Optimal Parallel Samplers by Haozhe Jiang, Nika Haghtalab, Lijie Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-25014.txt.

Description

Haozhe Jiang, Nika Haghtalab, and Lijie Chen prove diffusion language models are provably optimal parallel samplers. Diffusion models create tokens in parallel, hopeful faster inference than autoregressive models that go one by one. Formalizing parallel sampling, they show such models with polynomial-length chain-of-thought can mimic any parallel sampling method using an optimal number of stepwise steps.

Whenever a target distribution needs few stepwise steps, a diffusion model can match that optimum. Machine learning theory readers get a strict foundation for the speed edge, not just benchmarks. Parallel generation is shown to lose nothing in stepwise efficiency, under the stated model.

Catalog record

CREATORHaozhe Jiang, Nika Haghtalab, Lijie Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-25014

Browse the full library · About the Tower of Babel