@inproceedings{nasim-etal-2026-llms,
title = "Do {LLM}s Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation",
author = {Nasim, Mehwish and
Selvaganapathy, Sanjeevan and
Sabhahit, Neel Ganapathi and
Griesbach, Marie and
Bhandari, Pranav and
Stockdiek, Janina L{\"u}tke and
Sch{\"a}permeier, Lennart and
Naseem, Usman and
Grimme, Christian},
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.2085/",
doi = "10.18653/2026.findings-emnlp.2085",
pages = "44383--44402",
ISBN = "979-8-89176-462-0",
abstract = "Many benchmarks show that large language models can answer direct questions about culture. We study a different question: do they also change how they speak when culture is only implied by the situation? We introduce a triad-diagnostic framework; applying it to four deployed endpoints to evaluate 60 culturally grounded conversational scenarios. We examine responses across five languages in three conditions: a neutral baseline (Prompt A), an explicit cultural instruction (Prompt B), and implicit situational cueing (Prompt C). We use an LLM-judge ensemble with explicit rubrics to score responses on 12 pragmatic features covering deference to authority, individual-versus-group framing, and uncertainty management. We define \textbf{\textit{Pragmatic Context Sensitivity} (PCS)} as the fraction of the judge-scored Prompt A$\rightarrow$B feature shift that reappears under Prompt A$\rightarrow$C. Across four deployed LLM endpoints and five languages (English, German, Hindi, Nepali, and Urdu), the primary stable-only PCS mean is 0.196 ($SD = 0.113$), indicating that the tested models recover only about one-fifth of the explicitly prompted rubric-aligned shift. Transfer is strongest for authority-related cues (0.299) and weakest for individual-versus-group framing (0.120). We argue that multilingual cultural pragmatics is an explicit-versus-implicit deployment problem, not only a factual knowledge problem."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nasim-etal-2026-llms">
<titleInfo>
<title>Do LLMs Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mehwish</namePart>
<namePart type="family">Nasim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sanjeevan</namePart>
<namePart type="family">Selvaganapathy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neel</namePart>
<namePart type="given">Ganapathi</namePart>
<namePart type="family">Sabhahit</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marie</namePart>
<namePart type="family">Griesbach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pranav</namePart>
<namePart type="family">Bhandari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Janina</namePart>
<namePart type="given">Lütke</namePart>
<namePart type="family">Stockdiek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lennart</namePart>
<namePart type="family">Schäpermeier</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Usman</namePart>
<namePart type="family">Naseem</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christian</namePart>
<namePart type="family">Grimme</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>Many benchmarks show that large language models can answer direct questions about culture. We study a different question: do they also change how they speak when culture is only implied by the situation? We introduce a triad-diagnostic framework; applying it to four deployed endpoints to evaluate 60 culturally grounded conversational scenarios. We examine responses across five languages in three conditions: a neutral baseline (Prompt A), an explicit cultural instruction (Prompt B), and implicit situational cueing (Prompt C). We use an LLM-judge ensemble with explicit rubrics to score responses on 12 pragmatic features covering deference to authority, individual-versus-group framing, and uncertainty management. We define Pragmatic Context Sensitivity (PCS) as the fraction of the judge-scored Prompt A\rightarrowB feature shift that reappears under Prompt A\rightarrowC. Across four deployed LLM endpoints and five languages (English, German, Hindi, Nepali, and Urdu), the primary stable-only PCS mean is 0.196 (SD = 0.113), indicating that the tested models recover only about one-fifth of the explicitly prompted rubric-aligned shift. Transfer is strongest for authority-related cues (0.299) and weakest for individual-versus-group framing (0.120). We argue that multilingual cultural pragmatics is an explicit-versus-implicit deployment problem, not only a factual knowledge problem.</abstract>
<identifier type="citekey">nasim-etal-2026-llms</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2085</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2085/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>44383</start>
<end>44402</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Do LLMs Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation
%A Nasim, Mehwish
%A Selvaganapathy, Sanjeevan
%A Sabhahit, Neel Ganapathi
%A Griesbach, Marie
%A Bhandari, Pranav
%A Stockdiek, Janina Lütke
%A Schäpermeier, Lennart
%A Naseem, Usman
%A Grimme, Christian
%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 nasim-etal-2026-llms
%X Many benchmarks show that large language models can answer direct questions about culture. We study a different question: do they also change how they speak when culture is only implied by the situation? We introduce a triad-diagnostic framework; applying it to four deployed endpoints to evaluate 60 culturally grounded conversational scenarios. We examine responses across five languages in three conditions: a neutral baseline (Prompt A), an explicit cultural instruction (Prompt B), and implicit situational cueing (Prompt C). We use an LLM-judge ensemble with explicit rubrics to score responses on 12 pragmatic features covering deference to authority, individual-versus-group framing, and uncertainty management. We define Pragmatic Context Sensitivity (PCS) as the fraction of the judge-scored Prompt A\rightarrowB feature shift that reappears under Prompt A\rightarrowC. Across four deployed LLM endpoints and five languages (English, German, Hindi, Nepali, and Urdu), the primary stable-only PCS mean is 0.196 (SD = 0.113), indicating that the tested models recover only about one-fifth of the explicitly prompted rubric-aligned shift. Transfer is strongest for authority-related cues (0.299) and weakest for individual-versus-group framing (0.120). We argue that multilingual cultural pragmatics is an explicit-versus-implicit deployment problem, not only a factual knowledge problem.
%R 10.18653/2026.findings-emnlp.2085
%U https://aclanthology.org/2026.findings-emnlp.2085/
%U https://doi.org/10.18653/2026.findings-emnlp.2085
%P 44383-44402
Markdown (Informal)
[Do LLMs Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation](https://aclanthology.org/2026.findings-emnlp.2085/) (Nasim et al., Findings 2026)
ACL
- Mehwish Nasim, Sanjeevan Selvaganapathy, Neel Ganapathi Sabhahit, Marie Griesbach, Pranav Bhandari, Janina Lütke Stockdiek, Lennart Schäpermeier, Usman Naseem, and Christian Grimme. 2026. Do LLMs Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 44383–44402, Budapest, Hungary. Association for Computational Linguistics.