Reproduce “SPARSEUP: Towards Better Recipes for Training Modern Learned Sparse Retrievers” 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

Thibault Formal, Pierre Erbacher, Cosimo Rulli

SPARSEUP trains sparse retrieval encoders by shifting logits before activation, capping each input position’s vocabulary expansion, and merging case and word-boundary variants. This recipe improves the effectiveness–sparsity trade-off over stronger FLOPS regularization alone at comparable document sparsity in an ablation. It averages 56.4 nDCG@10 on BEIR-13, highest the authors know among public vocabulary-based sparse encoders below 150M parameters, but trails same-family dense and late-interaction models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.