@inproceedings{yang-etal-2026-factorized,
title = "Factorized Soft Prompts with Inference-Time Latent Control for {LLM} Personalization",
author = "Yang, Likun and
Xu, Pei and
Cao, Shiyue and
Wang, Tong and
Wu, Haoran and
Chen, Xiaotang and
Huang, Kaiqi",
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.1974/",
doi = "10.18653/2026.findings-emnlp.1974",
pages = "42093--42116",
ISBN = "979-8-89176-462-0",
abstract = "Personalizing large language models (LLMs) requires reflecting stable user preferences while adapting their expression to each target context. Soft-prompt personalization has become a practical route to this goal, but existing methods produce soft prompts in which stable preferences, target-induced content, and context-dependent expression are bundled into an entangled conditioning signal without separately addressable components, leaving inference with no principled space to intervene on individual aspects of personalization. We propose Factorized soft prompts with Latent Control (FaLCon), an LLM personalization framework built on two coupled stages. First, factorized soft prompts replace this entangled conditioning with a set of separately addressable preference factors, forming a structured control interface. Second, inference-time latent control treats the resulting factor weights as a continuous action space and adapts them per instance from teacher-forced likelihood feedback, keeping the LLM frozen and updating only a low-dimensional factor-weight posterior without back-propagating through the LLM. Experiments on multiple personalized generation tasks reveal that FaLCon consistently outperforms strong baselines, validating the individual necessity and joint effectiveness of the structured interface and the inference-time controller."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-factorized">
<titleInfo>
<title>Factorized Soft Prompts with Inference-Time Latent Control for LLM Personalization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Likun</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pei</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shiyue</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haoran</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaotang</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaiqi</namePart>
<namePart type="family">Huang</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>Personalizing large language models (LLMs) requires reflecting stable user preferences while adapting their expression to each target context. Soft-prompt personalization has become a practical route to this goal, but existing methods produce soft prompts in which stable preferences, target-induced content, and context-dependent expression are bundled into an entangled conditioning signal without separately addressable components, leaving inference with no principled space to intervene on individual aspects of personalization. We propose Factorized soft prompts with Latent Control (FaLCon), an LLM personalization framework built on two coupled stages. First, factorized soft prompts replace this entangled conditioning with a set of separately addressable preference factors, forming a structured control interface. Second, inference-time latent control treats the resulting factor weights as a continuous action space and adapts them per instance from teacher-forced likelihood feedback, keeping the LLM frozen and updating only a low-dimensional factor-weight posterior without back-propagating through the LLM. Experiments on multiple personalized generation tasks reveal that FaLCon consistently outperforms strong baselines, validating the individual necessity and joint effectiveness of the structured interface and the inference-time controller.</abstract>
<identifier type="citekey">yang-etal-2026-factorized</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1974</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1974/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>42093</start>
<end>42116</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Factorized Soft Prompts with Inference-Time Latent Control for LLM Personalization
%A Yang, Likun
%A Xu, Pei
%A Cao, Shiyue
%A Wang, Tong
%A Wu, Haoran
%A Chen, Xiaotang
%A Huang, Kaiqi
%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 yang-etal-2026-factorized
%X Personalizing large language models (LLMs) requires reflecting stable user preferences while adapting their expression to each target context. Soft-prompt personalization has become a practical route to this goal, but existing methods produce soft prompts in which stable preferences, target-induced content, and context-dependent expression are bundled into an entangled conditioning signal without separately addressable components, leaving inference with no principled space to intervene on individual aspects of personalization. We propose Factorized soft prompts with Latent Control (FaLCon), an LLM personalization framework built on two coupled stages. First, factorized soft prompts replace this entangled conditioning with a set of separately addressable preference factors, forming a structured control interface. Second, inference-time latent control treats the resulting factor weights as a continuous action space and adapts them per instance from teacher-forced likelihood feedback, keeping the LLM frozen and updating only a low-dimensional factor-weight posterior without back-propagating through the LLM. Experiments on multiple personalized generation tasks reveal that FaLCon consistently outperforms strong baselines, validating the individual necessity and joint effectiveness of the structured interface and the inference-time controller.
%R 10.18653/2026.findings-emnlp.1974
%U https://aclanthology.org/2026.findings-emnlp.1974/
%U https://doi.org/10.18653/2026.findings-emnlp.1974
%P 42093-42116
Markdown (Informal)
[Factorized Soft Prompts with Inference-Time Latent Control for LLM Personalization](https://aclanthology.org/2026.findings-emnlp.1974/) (Yang et al., Findings 2026)
ACL