sbert-siamese-tied-weights

IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s2-related-work.md

Created 2026-08-25T02:58:27+00:00

In SBERT's siamese network structure, the two BERT encoders share (are tied to) the same weights rather than being separate models.

Summary

Rather than running two separate transformer models side by side, SBERT runs the same single model twice on two different inputs. This ensures both sentences are projected into one shared meaning space, so the distance between their outputs is a consistent, symmetric similarity measure rather than a comparison between two unrelated coordinate systems.

Dependents

These beliefs depend on this one:

IN directionality-as-universal-geometric-property — Asymmetry (A→B ≠ B→A) appears independently in ROME's separate-edit constraint, Park's signed binary contrasts (ℓ̄_{w0⇒w1} = ℓ̄_{w1} − ℓ̄_{w0}), and SBERT's tied-weight siamese architecture, establishing directionality as an invariant geometric property of the superposed space rather than a per-implementation artifact.