sbert-default-pooling-mean
IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s2-related-work.md
Created 2026-08-25T02:58:27+00:00
SBERT's default sentence pooling strategy is MEAN over all token vectors, with CLS-token and MAX-over-time as alternatives.
Summary
When SBERT compresses a whole sentence into a single comparison-ready vector, it does so by averaging every token's embedding together rather than relying on a special start-of-sentence token or a peak-value trick. This means every word in the sentence contributes equally to the final meaning vector, so a system built on this default will weigh rare and common terms the same unless explicitly reconfigured.