SPLADE Deep Dive SPLADE vs BM25 in One Sentence BM25 scores documents by raw term frequency weighted by IDF. SPLADE uses a masked-language-model head to predict a learned weight for every term in the vocabulary — including terms that never appeared in the document. This is called learned term expansion: a document about "token refresh" gets non-zero weights for "oauth", "bearer", "jwt", "access" even if those words are absent in the surface text. The FLOPS Regularizer Without constraint, SPLADE would fill the whole vocabulary with small non-zero weights, which is slow to index and search. SPL…