CRoCoDiL: Continuous and Robust Conditioned Diffusion for Language

Roy Uziel, Omer Belhasin, Itay Levi, Akhiad Bercovich, Ran El-Yaniv, Ran Zilberstein, Michael Elad, NVIDIA · 2026

Read CRoCoDiL: Continuous and Robust Conditioned Diffusion for Language by Roy Uziel, Omer Belhasin, Itay Levi, Akhiad Bercovich, Ran El-Yaniv, Ran Zilberstein, Michael Elad, NVIDIA (2026) 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-2603-20210.txt.

Description

Uziel and co-authors move masked diffusion for language into a continuous semantic space. Masked diffusion models generate fast but can lose track of token links and meaning. CRoCoDiL fine-tunes an encoder and demasker together around latent sentence representations. Decoding then runs through the masked diffusion algorithm, forming a novel autoencoder. Two sampling routes follow: one generates latents first, then decodes them. The other refines latents during the discrete sampling itself. On LLaDA tests, quality rose and sampling ran over ten times faster in unconditional settings.

Catalog record

CREATORRoy Uziel, Omer Belhasin, Itay Levi, Akhiad Bercovich, Ran El-Yaniv, Ran Zilberstein, Michael Elad, NVIDIA
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78032
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.20210

Browse the full library · About the Tower of Babel