Reproduce “ZeroGR: Filtering-First Constrained Decoding for Generative Recommendation” 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
Xiangrui Yang, Jiawei Yang, Liang Zeng, Daqian Ding, Ziqian Liu, Bing Tian, Cheng Peng, Yunfeng Zhao and 4 more
ZeroGR accelerates constrained recommendation decoding by looking up catalog-valid next tokens before scoring, rather than scoring the whole vocabulary. Its step-sliced index and shared candidate slots support selective projection, while full projection is used when faster. On MovieLens-1M, returned items overlap 99.90% on average with scoring-first results. Filtering, scoring, and selection latency falls by up to 9.9×, excluding model inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.