DySem: Uncovering Dynamic Semantic Components of Large Language Models for Calculating Semantic Textual Similarity

Kaijie Zheng, Weiqin Wang, Yile Wang, Hui HuangCollege of Computer Science and Software Engineering, Shenzhen University · 2026

Read DySem: Uncovering Dynamic Semantic Components of Large Language Models for Calculating Semantic Textual Similarity by Kaijie Zheng, Weiqin Wang, Yile Wang, Hui HuangCollege of Computer Science and Software Engineering, Shenzhen University (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-2605-29751.txt.

Description

Kaijie Zheng and three colleagues argue that standard similarity scores waste what language models know. Last-layer states mix broad knowledge with meaning, and their huge size adds noise and redundancy to the comparison. Their training-free frame, DySem, hunts meaning-bearing inner parts through agreement across tongues, then builds a text-dependent shared set of meaning dimensions for each pair. Similarity is then scored over that shared subset alone, not a fixed space. Tests across several models show wins over recent rivals with fewer dimensions for the score.

Catalog record

CREATORKaijie Zheng, Weiqin Wang, Yile Wang, Hui HuangCollege of Computer Science and Software Engineering, Shenzhen University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56705
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.29751

Browse the full library · About the Tower of Babel