Reproduce “ShortListing Model: A Streamlined SimplexDiffusion for Discrete Variable Generation” 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
Yuxuan Song, Zhe Zhang, Yu Pei, Jingjing Gong, Qiying Yu, Zheng Zhang, Mingxuan Wang, Hao Zhou and 2 more
Researchers at Tsinghua University and ByteDance Seed developed the Shortlisting Model (SLM), a streamlined simplex diffusion approach for discrete variable generation. SLM views generation as progressive candidate pruning, achieving state-of-the-art performance in protein and DNA design, and competitive results in large-vocabulary language modeling by simplifying continuous simplex diffusion.
Key insights
- Reframing discrete variable generation as progressive candidate pruning significantly simplifies continuous simplex-based diffusion while maintaining strong generative capabilities.
- Restricting transitions to centroids within the simplex enables a more streamlined and scalable approach compared to modeling trajectories over the entire continuous space.
- Simplified cross-entropy-based training objectives effectively address gradient vanishing, improving stability and performance for discrete diffusion models in large vocabulary contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.