Bit-Level Discrete Diffusion with Markov Probabilistic Models: An Improved Framework with Sharp Convergence Bounds under Minimal Assumptions

Le Tuyet Nhi Pham, Dario Shariatian, Antonio Ocello, Giovanni Conforti, Alain Durmus · 2025

Read Bit-Level Discrete Diffusion with Markov Probabilistic Models: An Improved Framework with Sharp Convergence Bounds under Minimal Assumptions by Le Tuyet Nhi Pham, Dario Shariatian, Antonio Ocello, Giovanni Conforti, Alain Durmus (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-2502-07939.txt.

Description

Le-Tuyet-Nhi Pham and co-authors build a generator for data made of bits, with the math to prove it converges. Their method adds noise by flipping bits at random over time, then learns to run the film backward, guided by a discrete cousin of the score function used in image generators. They prove that this guide is exactly the expected value of a certain forward-process quantity, which ties their method to the score-based family.

The team derives sharp error bounds under weak assumptions. Experiments on simple random data and on binary handwritten digits matched leading rivals. The authors aim to give discrete generation the same firm footing as its continuous cousins.

Catalog record

CREATORLe Tuyet Nhi Pham, Dario Shariatian, Antonio Ocello, Giovanni Conforti, Alain Durmus
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE241 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-07939

Browse the full library · About the Tower of Babel