Accelerating Portuguese Masked Diffusion Models through Representation Alignment

Adalberto Ferreira Barbosa Junior, Lucas Lima Neves, Adriano César Santana


Abstract
Masked Diffusion Language Models (MDLM) have recently demonstrated that discrete diffusion can achieve competitive performance in text generation. However, training these models remains computationally expensive, particularly for lower-resourced languages like Portuguese. In this work, we adapt REPresentation Alignment (REPA), a technique originally proposed for vision, to the textual domain. We systematically evaluate the impact of aligning the internal representations of a Portuguese MDLM with those of pretrained teacher encoders (e.g., Qwen, BERTimbau). Our experiments show that REPA significantly accelerates training and improves final perplexity by 28.6% compared to a baseline without alignment. We also identify optimal hyperparameters, finding that mid-level alignment with modern teacher encoders yields the best results.
Anthology ID:
2026.propor-1.97
Volume:
Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1
Month:
April
Year:
2026
Address:
Salvador, Brazil
Editors:
Marlo Souza, Iria de-Dios-Flores, Diana Santos, Larissa Freitas, Jackson Wilke da Cruz Souza, Eugénio Ribeiro
Venue:
PROPOR
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
968–973
Language:
URL:
https://aclanthology.org/2026.propor-1.97/
DOI:
Bibkey:
Cite (ACL):
Adalberto Ferreira Barbosa Junior, Lucas Lima Neves, and Adriano César Santana. 2026. Accelerating Portuguese Masked Diffusion Models through Representation Alignment. In Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1, pages 968–973, Salvador, Brazil. Association for Computational Linguistics.
Cite (Informal):
Accelerating Portuguese Masked Diffusion Models through Representation Alignment (Junior et al., PROPOR 2026)
Copy Citation:
PDF:
https://aclanthology.org/2026.propor-1.97.pdf