sbert-inference-no-combination-vector

IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s8-conclusion.md

Created 2026-08-25T02:58:30+00:00

At inference time, SBERT uses only cosine similarity on the two raw sentence embeddings; the combination vector (u, v, |u−v|) is used exclusively during classifier training.

Summary

At prediction time, SBERT skips the heavier classifier step entirely and just compares two sentence vectors with a simple cosine similarity, meaning inference is fast and cheap. The richer combination-vector machinery (concatenating the two vectors and their element-wise difference) is only a training-time tool for shaping the embedding space, not something the running system actually computes.