@inproceedings{shi-etal-2026-dearag,
title = "{D}e{ARAG}: Decoupled Agentic {RAG} for On-Device and Precise Multi-Hop Question Answering",
author = "Shi, Qiuchang and
Liu, Tianqi and
Zhang, Feng and
Chen, Peng and
Liu, Jianjun and
An, Gaoyun",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2320/",
pages = "52706--52727",
ISBN = "979-8-89176-459-0",
abstract = "Although cloud-based Large Language Models (LLMs) have significantly advanced complex multi-hop question answering, they pose severe data privacy risks. On the other hand, simply migrating such compute-intensive systems to edge devices with limited computational resources inevitably triggers context overload and the ``lost-in-the-middle'' phenomenon. To resolve the privacy-accuracy dilemma, we introduce DeARAG, a fully decoupled, local agentic framework. By shifting heavy hypergraph construction entirely offline, DeARAG substantially condenses the required context during inference. Consequently, the agent{'}s exploration is strictly grounded in the retrieved hypergraph topology, ensuring rigorous, evidence-based reasoning without overwhelming local resources. Evaluated on multiple datasets, our framework consistently outperforms massive, cloud-dependent baselines (e.g., achieving a gain of 6.47 F1 points on HotpotQA) using only consumer-grade GPUs, enabling highly precise on-device reasoning."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shi-etal-2026-dearag">
<titleInfo>
<title>DeARAG: Decoupled Agentic RAG for On-Device and Precise Multi-Hop Question Answering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Qiuchang</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianqi</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peng</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianjun</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gaoyun</namePart>
<namePart type="family">An</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Although cloud-based Large Language Models (LLMs) have significantly advanced complex multi-hop question answering, they pose severe data privacy risks. On the other hand, simply migrating such compute-intensive systems to edge devices with limited computational resources inevitably triggers context overload and the “lost-in-the-middle” phenomenon. To resolve the privacy-accuracy dilemma, we introduce DeARAG, a fully decoupled, local agentic framework. By shifting heavy hypergraph construction entirely offline, DeARAG substantially condenses the required context during inference. Consequently, the agent’s exploration is strictly grounded in the retrieved hypergraph topology, ensuring rigorous, evidence-based reasoning without overwhelming local resources. Evaluated on multiple datasets, our framework consistently outperforms massive, cloud-dependent baselines (e.g., achieving a gain of 6.47 F1 points on HotpotQA) using only consumer-grade GPUs, enabling highly precise on-device reasoning.</abstract>
<identifier type="citekey">shi-etal-2026-dearag</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2320/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>52706</start>
<end>52727</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DeARAG: Decoupled Agentic RAG for On-Device and Precise Multi-Hop Question Answering
%A Shi, Qiuchang
%A Liu, Tianqi
%A Zhang, Feng
%A Chen, Peng
%A Liu, Jianjun
%A An, Gaoyun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F shi-etal-2026-dearag
%X Although cloud-based Large Language Models (LLMs) have significantly advanced complex multi-hop question answering, they pose severe data privacy risks. On the other hand, simply migrating such compute-intensive systems to edge devices with limited computational resources inevitably triggers context overload and the “lost-in-the-middle” phenomenon. To resolve the privacy-accuracy dilemma, we introduce DeARAG, a fully decoupled, local agentic framework. By shifting heavy hypergraph construction entirely offline, DeARAG substantially condenses the required context during inference. Consequently, the agent’s exploration is strictly grounded in the retrieved hypergraph topology, ensuring rigorous, evidence-based reasoning without overwhelming local resources. Evaluated on multiple datasets, our framework consistently outperforms massive, cloud-dependent baselines (e.g., achieving a gain of 6.47 F1 points on HotpotQA) using only consumer-grade GPUs, enabling highly precise on-device reasoning.
%U https://aclanthology.org/2026.emnlp-main.2320/
%P 52706-52727
Markdown (Informal)
[DeARAG: Decoupled Agentic RAG for On-Device and Precise Multi-Hop Question Answering](https://aclanthology.org/2026.emnlp-main.2320/) (Shi et al., EMNLP 2026)
ACL