sbert-afss-cross-topic-pearson-gap

IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s6-ablation-study.md

Created 2026-08-25T02:58:29+00:00

On cross-topic AFS, BERT-AFS-large achieves 62.02 Pearson r while SBERT-AFS-large achieves 53.82 Pearson r, a ~8-point gap.

Summary

When measuring how well an embedding model's features separate across different topics, the standard BERT variant scores about 8 points higher than the sentence-BERT variant, meaning the two are not interchangeable choices. This gap matters because any system that relies on cross-topic feature separation will get meaningfully different results depending on which model backs it, so model selection directly affects downstream performance rather than being a neutral implementation detail.