@inproceedings{sarangi-etal-2026-fine,
title = "Fine-Tuning for Social-Media User Simulation: A Study Across Models and Evaluators",
author = {Sarangi, Sneheel and
B{\"u}ck-Kaeffer, Aur{\'e}lien and
Puelma Touzel, Maximilian and
Yang, Zachary and
Godbout, Jean-Fran{\c{c}}ois and
Rabbany, Reihaneh},
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.2251/",
pages = "51204--51232",
ISBN = "979-8-89176-459-0",
abstract = "Does fine-tuning on a user{'}s social-media history make an LLM a better simulator of that user? We study this question using QLoRA, conditioning models on a user{'}s chronological action history and a reply thread before generating that user{'}s reply. Across 12 matched baseline{--}QLoRA pairs from four model families (0.8B{--}32B), plus a baseline-only 122B model, LLM-as-judge believability evaluations show that QLoRA improves several smaller models, has mixed effects on stronger models, and does not lift a sub-10B model into the strongest large-model tier. However, a real-versus-generated classifier gives a different result: QLoRA makes Qwen, OLMo, and Llama replies much less detectable, while Gemma remains readily detectable. Contextual believability and detector AUC are uncorrelated across systems. We further explore LLM evaluations: strong LLM judges agree on broad quality tiers, and a blind 200-comparison audit finds human{--}judge agreement of 66{\%} when three strong judges agree and 47{\%} when they split. The measured effect of QLoRA therefore depends on the evaluation axis."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sarangi-etal-2026-fine">
<titleInfo>
<title>Fine-Tuning for Social-Media User Simulation: A Study Across Models and Evaluators</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sneheel</namePart>
<namePart type="family">Sarangi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aurélien</namePart>
<namePart type="family">Bück-Kaeffer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maximilian</namePart>
<namePart type="family">Puelma Touzel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zachary</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jean-François</namePart>
<namePart type="family">Godbout</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Reihaneh</namePart>
<namePart type="family">Rabbany</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>Does fine-tuning on a user’s social-media history make an LLM a better simulator of that user? We study this question using QLoRA, conditioning models on a user’s chronological action history and a reply thread before generating that user’s reply. Across 12 matched baseline–QLoRA pairs from four model families (0.8B–32B), plus a baseline-only 122B model, LLM-as-judge believability evaluations show that QLoRA improves several smaller models, has mixed effects on stronger models, and does not lift a sub-10B model into the strongest large-model tier. However, a real-versus-generated classifier gives a different result: QLoRA makes Qwen, OLMo, and Llama replies much less detectable, while Gemma remains readily detectable. Contextual believability and detector AUC are uncorrelated across systems. We further explore LLM evaluations: strong LLM judges agree on broad quality tiers, and a blind 200-comparison audit finds human–judge agreement of 66% when three strong judges agree and 47% when they split. The measured effect of QLoRA therefore depends on the evaluation axis.</abstract>
<identifier type="citekey">sarangi-etal-2026-fine</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2251/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51204</start>
<end>51232</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Fine-Tuning for Social-Media User Simulation: A Study Across Models and Evaluators
%A Sarangi, Sneheel
%A Bück-Kaeffer, Aurélien
%A Puelma Touzel, Maximilian
%A Yang, Zachary
%A Godbout, Jean-François
%A Rabbany, Reihaneh
%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 sarangi-etal-2026-fine
%X Does fine-tuning on a user’s social-media history make an LLM a better simulator of that user? We study this question using QLoRA, conditioning models on a user’s chronological action history and a reply thread before generating that user’s reply. Across 12 matched baseline–QLoRA pairs from four model families (0.8B–32B), plus a baseline-only 122B model, LLM-as-judge believability evaluations show that QLoRA improves several smaller models, has mixed effects on stronger models, and does not lift a sub-10B model into the strongest large-model tier. However, a real-versus-generated classifier gives a different result: QLoRA makes Qwen, OLMo, and Llama replies much less detectable, while Gemma remains readily detectable. Contextual believability and detector AUC are uncorrelated across systems. We further explore LLM evaluations: strong LLM judges agree on broad quality tiers, and a blind 200-comparison audit finds human–judge agreement of 66% when three strong judges agree and 47% when they split. The measured effect of QLoRA therefore depends on the evaluation axis.
%U https://aclanthology.org/2026.emnlp-main.2251/
%P 51204-51232
Markdown (Informal)
[Fine-Tuning for Social-Media User Simulation: A Study Across Models and Evaluators](https://aclanthology.org/2026.emnlp-main.2251/) (Sarangi et al., EMNLP 2026)
ACL