Structure Retention in Embedding Spaces as a Predictor of Benchmark Performance

Amanda Myntti [email protected], Jenna Kanerva [email protected], Veronika Laippala [email protected], Filip Ginter [email protected] · 2026

Read Structure Retention in Embedding Spaces as a Predictor of Benchmark Performance by Amanda Myntti [email protected], Jenna Kanerva [email protected], Veronika Laippala [email protected], Filip Ginter [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-22202.txt.

Description

Amanda Myntti and co-authors show high-performing models organize their vector spaces consistently. They evaluate 25 contemporary models on five MTEB tasks across retrieval, bitext mining, pair classification, and summarization in English and other-language settings. Nearest-neighbor overlap and magnitude differences in ICA between paired texts correlate up to 0.97 with task performance. Tasks thus vary in linearity and reliance on local information retention. The results point toward future training objectives and conditional vector design.

Catalog record

CREATORAmanda Myntti [email protected], Jenna Kanerva [email protected], Veronika Laippala [email protected], Filip Ginter [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63438
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2605.22202

Browse the full library · About the Tower of Babel