Reproduce “IDLM: Inverse-distilled Diffusion Language Models” 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

David Li, Nikita Gushchin, Dmitry Abulkhanov, Eric Moulines, Ivan Oseledets, Maxim Panov, Alexander Korotin

IDLM trains a few-step diffusion language model by comparing a frozen teacher’s assessment of samples with that of a fake denoiser trained on those samples. This inverse-distillation objective targets a student distribution rather than skipped teacher steps. On OpenWebText, IDLM-MDLM uses 16 rather than 1,024 reverse steps, with GenPPL, entropy, and MAUVE metrics; this is a step-count reduction, not speedup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.