Reproduce “Where Beam Search Loses Top-Ranked Items:First-Step Coverage in 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
Heng Yao, Hanwei Zhu
The authors reallocate beam capacity to the first token position in TIGER decoding, preserving the model and identifiers. In Sports with RQ-VAE identifiers, Front-151 raises Search Recall@10 from 0.8839 to 0.9929 at nominal budget 151. Across three domains and 77,373 users, Front-151 matches Beam-100’s ordered top ten at budgets 151 versus 301, while measured decoder latency falls 6.35–6.39%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.