Reproduce “BYOD: Build Your Own DLM - Efficient Autoregressive-to-Diffusion Language Model Conversion” 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
Ruurd J.A. Kuiper, Robert A. Bagheri
BYOD converts four instruction-tuned model families into diffusion language models by freezing pretrained weights and training LoRA adapters with a masked-token objective. Each 25,000-update conversion takes 3.1–4.0 hours on one GPU. BYOD-Gemma averages 50.6% across nine tasks versus 51.9% for rerun LLaDA. At 64 NFE, open-generation perplexity remains 8.27–11.12, versus 4.07 for LLaDA and 2.23–3.01 for paired autoregressive parents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.