Reproduce “Alpha Diffusion Language Models: Factorization Alone Is Not the Problem” 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

Nikita Gushchin, Dmitry Baranchuk, Alexander Korotin

AlphaDLM trains a factorized masked diffusion language model with a sequence-level generalized cross-entropy loss on the probability of an entire masked completion, rather than separate token losses. This encourages more consistent parallel predictions. Trained on TinyGSM, one checkpoint reaches 34.6% GSM8K accuracy in four evaluations. This is a fixed-evaluation result, not a claim of autoregressive-level accuracy or lower total compute.

Summary by alphaXiv.

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