Reproduce “Discrete Diffusion Language Modeling by Estimating the Ratios of the Data Distribution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aaron Lou, Chenlin Meng, Stefano Ermon
SEDD learns ratios between probabilities of nearby token sequences, then uses them to reverse sequence-wide corruption. Its score-entropy objective enforces positive ratio estimates and supports likelihood bounds. On GPT-2-scale language-modeling tasks, SEDD reports better perplexity bounds than GPT-2 on three of five tasks, and better sample generative perplexity with fewer network evaluations; results do not establish faster elapsed generation or performance beyond this scale.
Key insights
- A novel "score entropy" loss function, derived from Bregman divergence, provides a theoretically sound way to learn concrete scores (ratios) while inherently encouraging positive predictions.
- The denoising score entropy variant of the loss makes training discrete diffusion models computationally tractable for high-dimensional data like natural language.
- By modeling probability ratios, the framework naturally supports flexible conditional generation tasks such as infilling and arbitrary prompting, without requiring specialized training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.