Query as Anchor: Scenario-Adaptive User Representation via Large Language Model

Jiahao Yuan 1,2,∗ , Yike Xu 1,∗ , Jinyong Wen 1 , Baokun Wang 1,† , Ziyi Gao 1 , Xiaotong Lin 1, Yun Liu 1 , Xing Fu 1 , Yu Cheng 1 , Yongchao Liu 1 , Weiqiang Wang 1 , Zhongle Xie 3 · 2026

Read Query as Anchor: Scenario-Adaptive User Representation via Large Language Model by Jiahao Yuan 1,2,∗ , Yike Xu 1,∗ , Jinyong Wen 1 , Baokun Wang 1,† , Ziyi Gao 1 , Xiaotong Lin 1, Yun Liu 1 , Xing Fu 1 , Yu Cheng 1 , Yongchao Liu 1 , Weiqiang Wang 1 , Zhongle Xie 3 (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-2602-14492.txt.

Description

Jiahao Yuan and eleven co-authors offer a 2026 arXiv setup for industrial user representation that tune to each query. Old paradigms yield static, task-blind codes that strain to square varied downstream scenes in one vector space, while mixed multi-source data add noise and mode clashes.

Query-as-Anchor shifts user modeling from static coding to live, query-aware building. A pre-training set called UserU lines up multi-mode act runs with user-knowing meanings, and a Q-Anchor code design joins ranked coarse-to-fine coders into dual-tower language models via joint contrastive-autoregressive tuning. Cluster-based soft prompt tuning then bridges broad pre-training and special business logic, forcing scene-aware tuning for downstream tasks.

Catalog record

CREATORJiahao Yuan 1,2,∗ , Yike Xu 1,∗ , Jinyong Wen 1 , Baokun Wang 1,† , Ziyi Gao 1 , Xiaotong Lin 1, Yun Liu 1 , Xing Fu 1 , Yu Cheng 1 , Yongchao Liu 1 , Weiqiang Wang 1 , Zhongle Xie 3
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-14492

Browse the full library · About the Tower of Babel