Reproduce “DNeo: Your Parallel Draft Model Needs Better Target Feature Encoding” 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

Jameson Sandler, Shane Bergsma, Mostafa Elhoushi, Bin Claire Zhang, Gavia Gray, Joel Hestness, Nolan Dey, Ferdinando Fioretto

DNeo accelerates large language model generation by having a small drafter propose token blocks for target-model verification. It encodes target features with MoFA and reuses features from rejected proposals as memory for the next draft. Rejection sampling preserves the target model’s output distribution. On Qwen3-4B GSM8K at temperature 1, it reports 3.20× speedup and 4.84 TPF under the tested conditions.

Summary by alphaXiv.

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