@inproceedings{jang-etal-2026-doc2weight,
title = "{D}oc2{W}eight: In-Weights Retrieval-Augmented Generation via Parameter Injection",
author = "Jang, GyuSik and
Lee, Wonkee and
Na, Seung-Hoon",
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.744/",
pages = "15657--15674",
ISBN = "979-8-89176-462-0",
abstract = "Retrieval-augmented generation (RAG) typically incorporates external knowledge by appending retrieved documents to the input prompt. While effective for knowledge-intensive tasks, this in-context formulation increases memory use and serving cost as the retrieved context grows. Motivated by this limitation, we propose Doc2Weight, an in-weights RAG framework that injects retrieved information into model parameters rather than the textual context. Doc2Weight converts retrieved documents into parameter-efficient low-rank updates and fuses them in weight space according to their relevance to the query. This enables retrieval augmentation with a query-only prompt, decoupling retrieval depth from prompt length while allowing multiple retrieved passages to influence generation within a single forward pass. We evaluate Doc2Weight on instruction-tuned language models of different scales, including LLaMA-3.2-1B, Qwen2.5-1.5B, and LLaMA-3-8B, across major QA benchmarks. Experiments show that Doc2Weight achieves the best average F1 among standard in-context RAG and prior in-weights RAG baselines across backbones and prompting settings, while matching the throughput of in-context RAG once document signatures are cached. Our code is publicly available."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jang-etal-2026-doc2weight">
<titleInfo>
<title>Doc2Weight: In-Weights Retrieval-Augmented Generation via Parameter Injection</title>
</titleInfo>
<name type="personal">
<namePart type="given">GyuSik</namePart>
<namePart type="family">Jang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wonkee</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seung-Hoon</namePart>
<namePart type="family">Na</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>Retrieval-augmented generation (RAG) typically incorporates external knowledge by appending retrieved documents to the input prompt. While effective for knowledge-intensive tasks, this in-context formulation increases memory use and serving cost as the retrieved context grows. Motivated by this limitation, we propose Doc2Weight, an in-weights RAG framework that injects retrieved information into model parameters rather than the textual context. Doc2Weight converts retrieved documents into parameter-efficient low-rank updates and fuses them in weight space according to their relevance to the query. This enables retrieval augmentation with a query-only prompt, decoupling retrieval depth from prompt length while allowing multiple retrieved passages to influence generation within a single forward pass. We evaluate Doc2Weight on instruction-tuned language models of different scales, including LLaMA-3.2-1B, Qwen2.5-1.5B, and LLaMA-3-8B, across major QA benchmarks. Experiments show that Doc2Weight achieves the best average F1 among standard in-context RAG and prior in-weights RAG baselines across backbones and prompting settings, while matching the throughput of in-context RAG once document signatures are cached. Our code is publicly available.</abstract>
<identifier type="citekey">jang-etal-2026-doc2weight</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.744/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>15657</start>
<end>15674</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Doc2Weight: In-Weights Retrieval-Augmented Generation via Parameter Injection
%A Jang, GyuSik
%A Lee, Wonkee
%A Na, Seung-Hoon
%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 jang-etal-2026-doc2weight
%X Retrieval-augmented generation (RAG) typically incorporates external knowledge by appending retrieved documents to the input prompt. While effective for knowledge-intensive tasks, this in-context formulation increases memory use and serving cost as the retrieved context grows. Motivated by this limitation, we propose Doc2Weight, an in-weights RAG framework that injects retrieved information into model parameters rather than the textual context. Doc2Weight converts retrieved documents into parameter-efficient low-rank updates and fuses them in weight space according to their relevance to the query. This enables retrieval augmentation with a query-only prompt, decoupling retrieval depth from prompt length while allowing multiple retrieved passages to influence generation within a single forward pass. We evaluate Doc2Weight on instruction-tuned language models of different scales, including LLaMA-3.2-1B, Qwen2.5-1.5B, and LLaMA-3-8B, across major QA benchmarks. Experiments show that Doc2Weight achieves the best average F1 among standard in-context RAG and prior in-weights RAG baselines across backbones and prompting settings, while matching the throughput of in-context RAG once document signatures are cached. Our code is publicly available.
%U https://aclanthology.org/2026.findings-emnlp.744/
%P 15657-15674
Markdown (Informal)
[Doc2Weight: In-Weights Retrieval-Augmented Generation via Parameter Injection](https://aclanthology.org/2026.findings-emnlp.744/) (Jang et al., Findings 2026)
ACL