Saved in:
Bibliographic Details
Main Authors: Chen, Yujie, Chen, Tailai, Gao, Yifeng, He, Zoe Wanying, Xu, Yijue, Wang, Shaobo, Zhang, Linfeng
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.18103
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913162207428608
author Chen, Yujie
Chen, Tailai
Gao, Yifeng
He, Zoe Wanying
Xu, Yijue
Wang, Shaobo
Zhang, Linfeng
author_facet Chen, Yujie
Chen, Tailai
Gao, Yifeng
He, Zoe Wanying
Xu, Yijue
Wang, Shaobo
Zhang, Linfeng
contents Prefilling computational costs pose a significant bottleneck for Large Language Models (LLMs) and Large Multimodal Models (LMMs) in long-context settings. While token pruning reduces sequence length, prior methods rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. In this work, we observe that tokens evolve toward \textit{semantic fixing points}, making further processing redundant. To this end, we introduce Delta Attention Selective Halting (DASH), a training-free policy that monitors the layer-wise update dynamics of the self-attention mechanism to selectively halt stabilized tokens. Extensive evaluation confirms that DASH generalizes across language and vision benchmarks, delivering significant prefill speedups while preserving model accuracy and hardware efficiency. Code will be released at https://github.com/verach3n/DASH.git.
format Preprint
id arxiv_https___arxiv_org_abs_2604_18103
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling
Chen, Yujie
Chen, Tailai
Gao, Yifeng
He, Zoe Wanying
Xu, Yijue
Wang, Shaobo
Zhang, Linfeng
Artificial Intelligence
Prefilling computational costs pose a significant bottleneck for Large Language Models (LLMs) and Large Multimodal Models (LMMs) in long-context settings. While token pruning reduces sequence length, prior methods rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. In this work, we observe that tokens evolve toward \textit{semantic fixing points}, making further processing redundant. To this end, we introduce Delta Attention Selective Halting (DASH), a training-free policy that monitors the layer-wise update dynamics of the self-attention mechanism to selectively halt stabilized tokens. Extensive evaluation confirms that DASH generalizes across language and vision benchmarks, delivering significant prefill speedups while preserving model accuracy and hardware efficiency. Code will be released at https://github.com/verach3n/DASH.git.
title Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling
topic Artificial Intelligence
url https://arxiv.org/abs/2604.18103