Coupling Models for One-Step Discrete Generation

Fred Zhangzhi Peng, Avishek Joey Bose, Anru R. Zhang, Alexander Tong · 2026

Read Coupling Models for One-Step Discrete Generation by Fred Zhangzhi Peng, Avishek Joey Bose, Anru R. Zhang, Alexander Tong (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-2605-07193.txt.

Description

Fred Zhangzhi Peng and three co-authors generate discrete data, like text or designed sequences, in a single step. Generation usually needs step-by-step decoding or many refinement rounds. Their Coupling Models learn a direct pairing between discrete sequences and Gaussian noise variables, then train a decoder to invert that pairing in one pass.

Unlike distillation methods, nothing starts from a pretrained multi-step sampler, and no continuous flows over probability simplices are needed. Across text generation, fly brain enhancer design, and binary image generation, the model improved on the strongest one-step baselines in each domain. The gains suggest the pairing between data and noise matters greatly.

Catalog record

CREATORFred Zhangzhi Peng, Avishek Joey Bose, Anru R. Zhang, Alexander Tong
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE96 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.07193

Browse the full library · About the Tower of Babel