MaDiS: Taming Masked Diffusion Language Models for Sign Language Generation
Abstract
Sign language generation (SLG) aims to translate written texts into expressive sign motions, bridging communication barriers for the Deaf and Hard-of-Hearing communities. Recent studies formulate SLG within the language modeling framework using autoregressive language models, which suffer from unidirectional context modeling and slow token-by-token inference. To address these limitations, we present MaDiS, a masked-diffusion-based language model for SLG that captures bidirectional dependencies and supports efficient parallel multi-token generation. We further introduce a tri-level cross-modal pretraining scheme that jointly learns from token-, latent-, and 3D physical-space objectives to leverage complementary, multi-level sign representations. To accelerate model convergence in the fine-tuning stage, we design a novel unmasking strategy with temporal checkpoints, which restructures generation in a coarse-to-fine manner and reduces the combinatorial complexity of unmasking orders by over times. In addition, a mixture-of-parts embedding layer is developed to effectively fuse information stored in different part-wise sign tokens through a learnable gate and well-optimized codebooks. Extensive experiments on CSL-Daily, Phoenix-2014T, and How2Sign demonstrate that MaDiS achieves superior performance across multiple metrics, including DTW error and two newly introduced metrics, SiBLEU and SiCLIP, while delivering a 40\% higher throughput. Code and models will be publicly released.
Cite
@article{arxiv.2601.19577,
title = {MaDiS: Taming Masked Diffusion Language Models for Sign Language Generation},
author = {Ronglai Zuo and Rolandos Alexandros Potamias and Qi Sun and Evangelos Ververas and Jiankang Deng and Stefanos Zafeiriou},
journal= {arXiv preprint arXiv:2601.19577},
year = {2026}
}