Sentinel: Decoding Context Utilization via Attention Probing for Efficient LLM Context Compression

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Yong, Li, Heng, Huang, Yanwen, Cheng, Ning, Guo, Yang, Zhu, Yun, Wang, Yanmeng, Wang, Shaojun, Xiao, Jing
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908784164601856
author Zhang, Yong
Li, Heng
Huang, Yanwen
Cheng, Ning
Guo, Yang
Zhu, Yun
Wang, Yanmeng
Wang, Shaojun
Xiao, Jing
author_facet Zhang, Yong
Li, Heng
Huang, Yanwen
Cheng, Ning
Guo, Yang
Zhu, Yun
Wang, Yanmeng
Wang, Shaojun
Xiao, Jing
contents Retrieval-augmented generation (RAG) often suffers from long and noisy retrieved contexts. Prior context compression methods rely on predefined importance metrics or supervised compression models, rather than on the model's own inference-time behavior. We propose Sentinel, a lightweight sentence-level compression framework that treats context compression as an understanding decoding problem. Sentinel probes native attention behaviors of a frozen LLM with a lightweight readout to decode which parts of the context are actually utilized when answering a query, rather than using attention as a direct relevance score. We empirically observe that decoded relevance signals exhibit sufficient consistency across model scales to support effective compression with compact proxy models. On LongBench, Sentinel with a 0.5B proxy model achieves up to 5x compression while matching the QA performance of 7B-scale baselines, and despite being trained only on English QA data, generalizes effectively to Chinese and out-of-domain settings.
format Preprint
id arxiv_https___arxiv_org_abs_2505_23277
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Sentinel: Decoding Context Utilization via Attention Probing for Efficient LLM Context Compression
Zhang, Yong
Li, Heng
Huang, Yanwen
Cheng, Ning
Guo, Yang
Zhu, Yun
Wang, Yanmeng
Wang, Shaojun
Xiao, Jing
Computation and Language
Artificial Intelligence
Retrieval-augmented generation (RAG) often suffers from long and noisy retrieved contexts. Prior context compression methods rely on predefined importance metrics or supervised compression models, rather than on the model's own inference-time behavior. We propose Sentinel, a lightweight sentence-level compression framework that treats context compression as an understanding decoding problem. Sentinel probes native attention behaviors of a frozen LLM with a lightweight readout to decode which parts of the context are actually utilized when answering a query, rather than using attention as a direct relevance score. We empirically observe that decoded relevance signals exhibit sufficient consistency across model scales to support effective compression with compact proxy models. On LongBench, Sentinel with a 0.5B proxy model achieves up to 5x compression while matching the QA performance of 7B-scale baselines, and despite being trained only on English QA data, generalizes effectively to Chinese and out-of-domain settings.
title Sentinel: Decoding Context Utilization via Attention Probing for Efficient LLM Context Compression
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2505.23277