The Diffusion Duality
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917157698273280 |
|---|---|
| author | Sahoo, Subham Sekhar Deschenaux, Justin Gokaslan, Aaron Wang, Guanghan Chiu, Justin Kuleshov, Volodymyr |
| author_facet | Sahoo, Subham Sekhar Deschenaux, Justin Gokaslan, Aaron Wang, Guanghan Chiu, Justin Kuleshov, Volodymyr |
| contents | Uniform-state discrete diffusion models hold the promise of fast text generation due to their inherent ability to self-correct. However, they are typically outperformed by autoregressive models and masked diffusion models. In this work, we narrow this performance gap by leveraging a key insight: Uniform-state diffusion processes naturally emerge from an underlying Gaussian diffusion. Our method, Duo, transfers powerful techniques from Gaussian diffusion to improve both training and sampling. First, we introduce a curriculum learning strategy guided by the Gaussian process, doubling training speed by reducing variance. Models trained with curriculum learning surpass autoregressive models in zero-shot perplexity on 3 of 7 benchmarks. Second, we present Discrete Consistency Distillation, which adapts consistency distillation from the continuous to the discrete setting. This algorithm unlocks few-step generation in diffusion language models by accelerating sampling by two orders of magnitude. We provide the code, model checkpoints, and video tutorials on the project page: http://s-sahoo.github.io/duo |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_10892 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | The Diffusion Duality Sahoo, Subham Sekhar Deschenaux, Justin Gokaslan, Aaron Wang, Guanghan Chiu, Justin Kuleshov, Volodymyr Machine Learning Artificial Intelligence Computation and Language Uniform-state discrete diffusion models hold the promise of fast text generation due to their inherent ability to self-correct. However, they are typically outperformed by autoregressive models and masked diffusion models. In this work, we narrow this performance gap by leveraging a key insight: Uniform-state diffusion processes naturally emerge from an underlying Gaussian diffusion. Our method, Duo, transfers powerful techniques from Gaussian diffusion to improve both training and sampling. First, we introduce a curriculum learning strategy guided by the Gaussian process, doubling training speed by reducing variance. Models trained with curriculum learning surpass autoregressive models in zero-shot perplexity on 3 of 7 benchmarks. Second, we present Discrete Consistency Distillation, which adapts consistency distillation from the continuous to the discrete setting. This algorithm unlocks few-step generation in diffusion language models by accelerating sampling by two orders of magnitude. We provide the code, model checkpoints, and video tutorials on the project page: http://s-sahoo.github.io/duo |
| title | The Diffusion Duality |
| topic | Machine Learning Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2506.10892 |