@inproceedings{xu-etal-2026-tokenpilot,
title = "{T}oken{P}ilot: Cache-Efficient Context Management for {LLM} Agents",
author = "Xu, Buqiang and
Xue, Zirui and
Chen, Dianmou and
Fu, Chenyang and
Wu, Chiyu and
Huang, Caiying and
Jiang, Chen and
Fang, Jizhan and
Deng, Xinle and
Chen, Yijun and
Yao, Yunzhi and
Wang, Xuehai and
Shang, Jin and
Chen, Huajun and
Zhang, Ningyu",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.899/",
pages = "19034--19049",
ISBN = "979-8-89176-462-0",
abstract = "As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present **TokenPilot**, a dual-granularity context management framework. Globally, *Ingestion-Aware Compaction* acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, *Lifecycle-Aware Eviction* monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on `PinchBench{`} and `Claw-Eval{`} under both isolated and continuous modes demonstrate that TokenPilot reduces costs by **61{\%}** and **56{\%}** in isolated mode, and **61{\%}** and **87{\%}** in continuous mode, while maintaining competitive performance compared to prior systems."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-tokenpilot">
<titleInfo>
<title>TokenPilot: Cache-Efficient Context Management for LLM Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Buqiang</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zirui</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dianmou</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chenyang</namePart>
<namePart type="family">Fu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chiyu</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Caiying</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chen</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jizhan</namePart>
<namePart type="family">Fang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinle</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yijun</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunzhi</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuehai</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jin</namePart>
<namePart type="family">Shang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huajun</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ningyu</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present **TokenPilot**, a dual-granularity context management framework. Globally, *Ingestion-Aware Compaction* acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, *Lifecycle-Aware Eviction* monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on ‘PinchBench‘ and ‘Claw-Eval‘ under both isolated and continuous modes demonstrate that TokenPilot reduces costs by **61%** and **56%** in isolated mode, and **61%** and **87%** in continuous mode, while maintaining competitive performance compared to prior systems.</abstract>
<identifier type="citekey">xu-etal-2026-tokenpilot</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.899/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19034</start>
<end>19049</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T TokenPilot: Cache-Efficient Context Management for LLM Agents
%A Xu, Buqiang
%A Xue, Zirui
%A Chen, Dianmou
%A Fu, Chenyang
%A Wu, Chiyu
%A Huang, Caiying
%A Jiang, Chen
%A Fang, Jizhan
%A Deng, Xinle
%A Chen, Yijun
%A Yao, Yunzhi
%A Wang, Xuehai
%A Shang, Jin
%A Chen, Huajun
%A Zhang, Ningyu
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F xu-etal-2026-tokenpilot
%X As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present **TokenPilot**, a dual-granularity context management framework. Globally, *Ingestion-Aware Compaction* acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, *Lifecycle-Aware Eviction* monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on ‘PinchBench‘ and ‘Claw-Eval‘ under both isolated and continuous modes demonstrate that TokenPilot reduces costs by **61%** and **56%** in isolated mode, and **61%** and **87%** in continuous mode, while maintaining competitive performance compared to prior systems.
%U https://aclanthology.org/2026.findings-emnlp.899/
%P 19034-19049
Markdown (Informal)
[TokenPilot: Cache-Efficient Context Management for LLM Agents](https://aclanthology.org/2026.findings-emnlp.899/) (Xu et al., Findings 2026)
ACL
- Buqiang Xu, Zirui Xue, Dianmou Chen, Chenyang Fu, Chiyu Wu, Caiying Huang, Chen Jiang, Jizhan Fang, Xinle Deng, Yijun Chen, Yunzhi Yao, Xuehai Wang, Jin Shang, Huajun Chen, and Ningyu Zhang. 2026. TokenPilot: Cache-Efficient Context Management for LLM Agents. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 19034–19049, Budapest, Hungary. Association for Computational Linguistics.