Reproduce “Pivot-SD: Efficient Self-Distillation for Masked 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
Seo Hyun Kim, Sunwoo Hong, Younwoo Choi, Chen-Hao Chao, Se-Young Yun, Rahul G. Krishnan
Pivot-SD trains masked diffusion language models by selecting denoising commitments that sharply reduce uncertainty over still-masked positions. It reinforces selected commitments from verified correct answers and suppresses them from failed answers, leaving other failed-trajectory positions without loss. On LLaDA-8B-Instruct, it reaches 40.43% HumanEval+ accuracy at 256 tokens, but is not best on HumanEval+ at 512 tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.