Attn-GS: Attention-Guided Context Compression for Efficient Personalized LLMs

Shenglai Zeng, Tianqi Zheng, Chuan Tian, Dante Everaert, Yau-Shian Wang, Yupin Huang, Michael J. Morais, Rohit Patki, Jinjin Tian, Xinnan Dai, Kai Guo, Monica Xiao Cheng, Hui Liu, Michigan State University · 2026

Read Attn-GS: Attention-Guided Context Compression for Efficient Personalized LLMs by Shenglai Zeng, Tianqi Zheng, Chuan Tian, Dante Everaert, Yau-Shian Wang, Yupin Huang, Michael J. Morais, Rohit Patki, Jinjin Tian, Xinnan Dai, Kai Guo, Monica Xiao Cheng, Hui Liu, Michigan State 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-2602-07778.txt.

Description

Personalized language models should remember a user's history, but stuffing long profiles into the prompt wastes tokens and money. Shenglai Zeng and twelve co-authors ask whether the model's own attention patterns can reveal which profile pieces actually matter, then use that signal to compress context intelligently.

Their Attn-GS framework uses attention feedback from a marking model to flag important personalization signals and drop the rest, replacing crude heuristics like keeping only recent interactions. Fine-tuning further sharpens the model's ability to separate relevant from irrelevant details. The paper gives builders of personalized assistants a principled way to cut costs and latency without losing what makes the response personal.

Catalog record

CREATORShenglai Zeng, Tianqi Zheng, Chuan Tian, Dante Everaert, Yau-Shian Wang, Yupin Huang, Michael J. Morais, Rohit Patki, Jinjin Tian, Xinnan Dai, Kai Guo, Monica Xiao Cheng, Hui Liu, Michigan State University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-07778

Browse the full library · About the Tower of Babel