kuleshov-group/bd3lms
A family of block discrete denoising diffusion language models that interpolate between autoregressive and diffusion approaches for text generation.
Not currently ranked — collecting fresh signals.
star history
BD3-LMs decompose token sequences into blocks and perform discrete diffusion within each block to generate text. By tuning block size, the model trades off between quality and sample efficiency. The work proposes an efficient training algorithm, gradient variance estimators, and data-driven noise schedules to minimize variance. The models achieve state-of-the-art likelihoods among diffusion models while enabling generation of arbitrary-length sequences.
Frequently asked
- What is kuleshov-group/bd3lms?
- A family of block discrete denoising diffusion language models that interpolate between autoregressive and diffusion approaches for text generation.
- Is bd3lms open source?
- Yes — kuleshov-group/bd3lms is open source, released under the Apache-2.0 license.
- What language is bd3lms written in?
- kuleshov-group/bd3lms is primarily written in Python.
- How popular is bd3lms?
- kuleshov-group/bd3lms has 1k stars on GitHub.
- Where can I find bd3lms?
- kuleshov-group/bd3lms is on GitHub at https://github.com/kuleshov-group/bd3lms.