sbert-poly-encoders-unsuitable-on2
IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s2-related-work.md
Created 2026-08-25T02:58:28+00:00
Poly-encoders (Humeau et al., 2019) use an asymmetric score function and high computational cost, making them unsuitable for symmetric O(n²) tasks like clustering where SBERT's symmetric cosine approach is preferred.
Summary
If a task requires comparing every item against every other item on equal footing, such as clustering, you need an encoder that treats both sides the same way and can do so affordably. Poly-encoders fail on both counts because they score a query and an item differently and are expensive to run, so in practice SBERT's symmetric and cheaper cosine comparison is the workable choice for those workloads.