Reproduce “Bidirectional Centered Score Matching for Discrete Diffusion Distillation” 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
Peng Tu, Baoyou Chen, Haoyuan Xia, Zilong Dong, Weihao Yuan, Siyu Zhu
The authors introduce BCSM, a distillation method that matches relative preferences among dynamically selected teacher and student tokens, using centering to make pairwise comparison linear in active-set size. On zero-shot HumanEval, LLaDA-8B scores 42.1 with 42.9 evaluations per block, versus 36.0 with 256 for its base; this improves scores but does not restore cross-position dependencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.