@inproceedings{ziaei-2026-limits,
title = "The Limits of Simulated Societies: How Post-Training and Survey Fine-Tuning Erase Cross-Cultural Variance",
author = "Ziaei, Rojin",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.144/",
pages = "3152--3166",
ISBN = "979-8-89176-439-2",
abstract = "Using large language models (LLMs) to simulate diverse human populations promises to transform computational social science. However, the accuracy of these simulations remains unclear because current evaluations prioritize predicting an average response over capturing the rich diversity of opinions within real human groups. Here, we develop a diagnostic framework measuring both accuracy and viewpoint retention across 10,000 respondent-question pairs from the World Values Survey, spanning twelve countries and six continents. By evaluating nine LLMs across four alignment paradigms (zero-shot, SFT, DPO, GRPO) against ground-truth human responses, we identify a failure mode we term consensus collapse. While training alignment improves point accuracy, we find that it compresses the range of outputs toward a single demographic stereotype. The highest-performing model (Tulu 3 70B DPO, 57.9{\%} accuracy) retains only 50{\%} of real human variance ({\ensuremath{\sigma}}r = 0.50). We demonstrate this collapse is within the models' underlying parameters rather than decoding procedures; temperature scaling up to T = 1.0 alters fewer than 3.7{\%} of predictions. Furthermore, this diversity loss is inequitable, erasing significantly more variation for Global South populations like Nigeria ({\ensuremath{\sigma}}r = 0.11) and Ethiopia ({\ensuremath{\sigma}}r = 0.16) compared to Western democracies ({\ensuremath{\sigma}}r = 0.70{--}0.87), with accuracy-rewarded GRPO unable to reverse this effect. While mixing an aligned model with an unaligned prior partially mitigates the collapse (reducing W1 from 0.55 to 0.52), substantial gaps remain (Nigeria recovering from {\ensuremath{\sigma}}r = 0.14 to 0.47). Ultimately, we find that modern AI alignment methods trade distributional diversity for point accuracy, challenging existing assumptions for using LLMs as substitutes for human populations. Yet, by exposing the mechanics of this collapse, our work provides insight for future methods designed to value and preserve diverse human viewpoints."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ziaei-2026-limits">
<titleInfo>
<title>The Limits of Simulated Societies: How Post-Training and Survey Fine-Tuning Erase Cross-Cultural Variance</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rojin</namePart>
<namePart type="family">Ziaei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Using large language models (LLMs) to simulate diverse human populations promises to transform computational social science. However, the accuracy of these simulations remains unclear because current evaluations prioritize predicting an average response over capturing the rich diversity of opinions within real human groups. Here, we develop a diagnostic framework measuring both accuracy and viewpoint retention across 10,000 respondent-question pairs from the World Values Survey, spanning twelve countries and six continents. By evaluating nine LLMs across four alignment paradigms (zero-shot, SFT, DPO, GRPO) against ground-truth human responses, we identify a failure mode we term consensus collapse. While training alignment improves point accuracy, we find that it compresses the range of outputs toward a single demographic stereotype. The highest-performing model (Tulu 3 70B DPO, 57.9% accuracy) retains only 50% of real human variance (\ensuremathσr = 0.50). We demonstrate this collapse is within the models’ underlying parameters rather than decoding procedures; temperature scaling up to T = 1.0 alters fewer than 3.7% of predictions. Furthermore, this diversity loss is inequitable, erasing significantly more variation for Global South populations like Nigeria (\ensuremathσr = 0.11) and Ethiopia (\ensuremathσr = 0.16) compared to Western democracies (\ensuremathσr = 0.70–0.87), with accuracy-rewarded GRPO unable to reverse this effect. While mixing an aligned model with an unaligned prior partially mitigates the collapse (reducing W1 from 0.55 to 0.52), substantial gaps remain (Nigeria recovering from \ensuremathσr = 0.14 to 0.47). Ultimately, we find that modern AI alignment methods trade distributional diversity for point accuracy, challenging existing assumptions for using LLMs as substitutes for human populations. Yet, by exposing the mechanics of this collapse, our work provides insight for future methods designed to value and preserve diverse human viewpoints.</abstract>
<identifier type="citekey">ziaei-2026-limits</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.144/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>3152</start>
<end>3166</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Limits of Simulated Societies: How Post-Training and Survey Fine-Tuning Erase Cross-Cultural Variance
%A Ziaei, Rojin
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F ziaei-2026-limits
%X Using large language models (LLMs) to simulate diverse human populations promises to transform computational social science. However, the accuracy of these simulations remains unclear because current evaluations prioritize predicting an average response over capturing the rich diversity of opinions within real human groups. Here, we develop a diagnostic framework measuring both accuracy and viewpoint retention across 10,000 respondent-question pairs from the World Values Survey, spanning twelve countries and six continents. By evaluating nine LLMs across four alignment paradigms (zero-shot, SFT, DPO, GRPO) against ground-truth human responses, we identify a failure mode we term consensus collapse. While training alignment improves point accuracy, we find that it compresses the range of outputs toward a single demographic stereotype. The highest-performing model (Tulu 3 70B DPO, 57.9% accuracy) retains only 50% of real human variance (\ensuremathσr = 0.50). We demonstrate this collapse is within the models’ underlying parameters rather than decoding procedures; temperature scaling up to T = 1.0 alters fewer than 3.7% of predictions. Furthermore, this diversity loss is inequitable, erasing significantly more variation for Global South populations like Nigeria (\ensuremathσr = 0.11) and Ethiopia (\ensuremathσr = 0.16) compared to Western democracies (\ensuremathσr = 0.70–0.87), with accuracy-rewarded GRPO unable to reverse this effect. While mixing an aligned model with an unaligned prior partially mitigates the collapse (reducing W1 from 0.55 to 0.52), substantial gaps remain (Nigeria recovering from \ensuremathσr = 0.14 to 0.47). Ultimately, we find that modern AI alignment methods trade distributional diversity for point accuracy, challenging existing assumptions for using LLMs as substitutes for human populations. Yet, by exposing the mechanics of this collapse, our work provides insight for future methods designed to value and preserve diverse human viewpoints.
%U https://aclanthology.org/2026.aacl-long.144/
%P 3152-3166
Markdown (Informal)
[The Limits of Simulated Societies: How Post-Training and Survey Fine-Tuning Erase Cross-Cultural Variance](https://aclanthology.org/2026.aacl-long.144/) (Ziaei, AACL-IJCNLP 2026)
ACL