United States

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders (SAEs), we identify interpretable features that encode culturally salient information and aggregate them into $\textbf{Cu}$ltural $\textbf{E}$mbeddings ($\textit{CuE}$). We use $\textit{CuE}$ both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that $\textit{CuE}$ -based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, $\textit{CuE}$ -based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don&#39;t necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

ACL 2026

Steering LLMs for Culturally Localized Generation

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders (SAEs), we identify interpretable features that encode culturally salient information and aggregate them into $\textbf{Cu}$ltural $\textbf{E}$mbeddings ($\textit{CuE}$). We use $\textit{CuE}$ both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that $\textit{CuE}$ -based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, $\textit{CuE}$ -based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

workshop paper

The rapid evolution of text-to-image generation has blurred the perceptual boundary between natural and synthetic imagery. However, it remains questionable whether the statistical structure of generated visual content mirrors the information density of the physical visual world. Drawing upon principles from statisti-cal linguistics, this study investigates the visual language of generative models through the lens of Zipfian dynamics. By analyzing a large-scale corpus of real and synthetic images, we uncover a fundamental divergence between visual syntax and semantics. We find that while generative models have successfully replicated the low-level physics of light, their high-level texture vocabulary exhibits distinct statistical signatures. Our analysis reveals a spectrum of entropy, identifying architectural fingerprints unique to each model. Furthermore, we investigate the relationship between generated images and prompt complexity, and find that increasing the semantic specificity of text prompts systematically degrades the statistical realism of the generated output.

A Zipfian Analysis of Visual Token Distributions for AI-Generated Images

Large language models (LLMs) are increasingly deployed in multilingual settings, yet little is known about whether their moral and social judgments remain consistent across languages. In particular, when faced with moral and social dilemmas, LLMs must often implicitly or explicitly assign responsibility — to an individual, to broader social forces, or across multiple parties — a process known as responsibility attribution. This study investigates whether responsibility attributions vary across languages, whether any observed variation persists across thematic domains, and whether the degree of variation differs across LLMs. We evaluate three models (GPT-5.2, Gemini-2.5-Pro, and LLaMA-3.3-70B) across 12 scenarios spanning six thematic domains (marriage, career, authority, gender, elder care, and family). Each model was prompted to attribute responsibility for each scenario by selecting from four options: the primary individual, a secondary interpersonal actor, a broader societal factor, or distributed responsibility shared across multiple parties. Results reveal a significant overall association between language and responsibility attribution (Cramér's $V = 0.24$) that persists within every thematic domain ($V = 0.26$--$0.53$). The magnitude of cross-language variation is strongly model-dependent: GPT-5.2 and Gemini-2.5-Pro show modest shifts ($V \approx 0.19$), while LLaMA-3.3-70B exhibits substantially stronger divergence ($V = 0.52$). These findings suggest that normative consistency across languages cannot be assumed and should be treated as a distinct dimension of model evaluation.

Lost in Translation? How Language Shapes Responsibility Attribution in Large Language Models

This paper proposes a semi-automatic lexico-semantic modeling framework for Chinese chéngyǔ containing body-part and animal lexemes. The framework combines manual semantic annotation, lightweight RDF/OWL formalization and semantic classification in order to investigate whether lexical mediators such as 心 xīn “heart/mind”, 口 kǒu “mouth” or 马 mǎ “horse” are sufficient to predict idiomatic semantic interpretation. Based on 440 annotated chéngyǔ normalized into 18 semantic categories, we compare three classification approaches: a rule-based keyword baseline, character n-gram TF-IDF with logistic regression, and BERT-base-chinese. The results show that lexical mediators cannot be directly equated with semantic categories and that TF-IDF achieves the best overall performance, suggesting that lightweight character-level representations remain robust for very short idioms in low-resource settings. The study contributes an interpretable RDF/OWL-compatible resource for culture-aware modeling of Chinese idioms.

Ontology-oriented lexico-semantic modeling and neural classification of Chinese chéngyǔ: A culture-aware NLP approach

Cross-cultural psychology has shown that moral judgments about failures to help vary systematically across cultures. In a landmark study, Miller, Bersoff, and Harwood (1990) found that while Indian and American participants agreed that failures to help are undesirable, they differed in whether they considered helping a moral obligation subject to social sanction or a personal decision. We adapt Miller et al.'s paradigm---nine scenarios crossing need severity (life-threatening, moderate, minor) with role relationship (parent, friend, stranger) and their original probe questions---to a cross-lingual LLM setting, presenting them to four LLMs (GPT-5.4, Claude-Opus-4.6, DeepSeek-V3.1, Qwen3-235B) across ten languages. We find that language significantly shapes how LLMs categorize failures to help as moral violations, social conventions, personal-moral concerns, or personal decisions ($\chi^2(27) = 116.14$, $p < .001$, Cramer's $V = 0.147$). Models agree across languages that failures to help are undesirable, but diverge substantially in how they classify them, with the primary divergence falling between moral violations and personal decisions. The proportion of responses classifying failures as moral violations decreases as need severity decreases and the role relationship becomes more distant. Cross-lingual variation differs substantially across models, with open-weight models showing significantly stronger variation than closed-weight models. These findings indicate that users consulting LLMs in different languages may receive substantively different moral guidance, underscoring the need for cross-lingual normative auditing as a component of multilingual LLM evaluation.

Sorry, Can't Help You: How Large Language Models Judge Failures to Help Across Languages

Verbal humor involves reasoning through complex conversational contexts. Although LLMs have achieved strong performance on English humor datasets, their ability to interpret humor in Hindi remains unexplored. In this paper, we evaluate Hindi humor for which we extract dialogues from humorous video clips. We use a pipeline that transforms video content into detailed textual streams, including dialogue transcripts and scene descriptions, allowing reasoning over inputs exceeding 2,000 words. We test various LLMs, from efficient edge models (Qwen-2.5-7B, Qwen-3-7B, Gemma-3-27B) to Indic-focused models (Sarvam-M-24B) and large frontier models (Llama-3.1-70B, Gemini-2.0-Flash). Our findings show a concave performance pattern in long-context understanding, with reasoning quality peaking at moderate lengths (250–750 words) and declining at higher context lengths. We also show that standard metrics overstate pragmatic competence. While increasing model size generally improves performance, we also observe distinct failures in smaller LLMs due to instructional and linguistic issues, necessitating diversity metrics to capture hallucinations. Smaller, Hindi-focused models can compete with much larger generalist models. Importantly, our evaluation reveals that conversational humor is a challenge for even specialized models, making HinS a valuable benchmark for advancing research in Hindi Long-Context Humor Reasoning.

Does Reasoning Kill the Joke? Long-Context Humor Understanding in Hindi

Conversational AI systems trained on large-scale web corpora inevitably encode the cultural values and interactional norms embedded in their training data, yet our understanding of how deployed LLMs reflect or reinforce culture-specific social expectations remains limited. This study examined how supportive versus challenging chatbot interaction styles shape user experience and continuance intention, and whether people-pleasing tendency (PPT) moderates these effects across cultures. Taiwanese (N = 49) and Korean (N = 52) participants completed a collaborative tourism-planning task. Results showed that: (1) supportive chatbots consistently led to higher continuance intention, satisfaction, and trust; (2) PPT did not moderate these effects; and (3) cultural variation emerged only in perceived threat, where higher PPT was associated with greater baseline threat in the Taiwanese but not the Korean sample. These findings reveal how a general-purpose LLM style may differentially activate culturally situated social scripts, raising implications for culturally inclusive conversational AI design.

One Style Fits All? Cultural Values Embedded in Conversational AI via a People-Pleasing Lens

Culture shapes how people interpret language, especially in online reviews containing culture-specific items (CSIs). Yet, most existing evaluations treat culture as a monolithic construct, offering no insight into which cultural dimensions pose difficulty for readers, or how large language models (LLMs), which power AI reading assistants, perform across them. This gap limits our ability to obtain reliable, cross-cultural estimates of model performance. To address this, we analyze CSIs in English Goodreads reviews across Newmark's cultural dimensions (e.g., material, ecology, customs, habits, social) and evaluate six LLMs of varying sizes on their ability to identify CSIs within each dimension. We find that readers struggle most with CSIs from the material, customs, and social dimensions, while models underperform on more localized ones (e.g., habits), revealing systematic cultural blind spots. To support further research on culturally representative benchmarking, we release an expert-annotated dataset of CSIs labeled by cultural dimension. Empirical analysis shows our dataset as more challenging and of higher quality than existing cultural benchmarks, enabling finer-grained evaluation of cultural understanding in models.

Beyond Monolithic Culture: Evaluating Understandability of Online Text Across Cultural Dimensions

Current Vision-Language Models (VLMs) are celebrated for their multilingual capabilities, yet they operate under a flawed assumption: that one language corresponds to a single writing system. This overlooks billions of users of multi-script languages like Punjabi, Serbian, Hindi-Urdu, Kurdish, among many others, for whom a model's capability may be fractured by orthographic bias. We introduce PuMVR (Punjabi Multimodal Visual Reasoning), the first benchmark designed to quantify script-dependent bias through 375 culturally grounded image-reasoning tasks across Punjabi's three active scripts (Gurmukhi, Shahmukhi, Roman). Evaluating 10 state-of-the-art VLMs, we expose a substantial Script Gap: models frequently solve visual puzzles in one script while failing identical tasks in another, with accuracy deltas reaching 16% and Script Consistency Rates (SCR) as low as 24.8%. Crucially, visual input boosts absolute performance but does not close this gap, the relative bias persists. Our analysis suggests reasoning patterns show limited cross-script transferability, and Chain-of-Thought pathways diverge based on script alone. We propose SCR as a core metric for script-agnostic evaluation, challenging current multilingual assessment paradigms and providing a framework for equitable AI.

Beyond 'One Language, One Script': Quantifying Orthographic Bias in Multilingual VLMs with PuMVR

Theory of Mind (ToM) - the ability to attribute beliefs and intents to others - is fundamental for social intelligence, yet Vision-Language Model (VLM) evaluations remain largely Western-centric. In this work, we introduce CulturalToM-VQA, a benchmark of 5,095 visually situated ToM probes across diverse cultural contexts, rituals, and social norms. Constructed through a frontier proprietary MLLM, human-verified pipeline, the dataset spans a taxonomy of six ToM tasks and four complexity levels. We benchmark 10 VLMs (2023-2025) and observe a significant performance leap: while earlier models struggle, frontier models achieve high accuracy (>93%). However, significant limitations persist: models struggle with false belief reasoning (19-83% accuracy) and show high regional variance (20-30% gaps). Crucially, we find that SOTA models exhibit social desirability bias - systematically favoring semantically positive answer choices over negative ones. Ablation experiments reveal that some frontier models rely heavily on parametric social priors, frequently defaulting to safety-aligned predictions. Furthermore, while Chain-of-Thought prompting aids older models, it yields minimal gains for newer ones. Overall, our work provides a testbed for cross-cultural social reasoning, underscoring that despite architectural gains, achieving robust, visually grounded understanding remains an open challenge.

Are Vision Language Models Cross-Cultural Theory of Mind Reasoners?

Stereotype repositories are critical to assess generative AI model safety, but currently lack adequate global coverage. It is imperative to prioritize targeted expansion, strategically addressing existing deficits, over merely increasing data volume. This work introduces a multilingual stereotype resource covering four sub-Saharan African countries that are severely underrepresented in NLP resources: Ghana, Kenya, Nigeria, and South Africa. By utilizing socioculturally-situated, community-engaged methods, including telephonic surveys moderated in native languages, we establish a reproducible methodology that is sensitive to the region's complex linguistic diversity and traditional orality. By deliberately balancing the sample across diverse ethnic and demographic backgrounds, we ensure broad coverage, resulting in a dataset of 3,534 stereotypes in English and 3,206 stereotypes across 15 native languages.

Content not yet available

Downloads

Next from ACL 2026

A Zipfian Analysis of Visual Token Distributions for AI-Generated Images

Similar lecture

Localized Cultural Knowledge is Conserved and Controllable in Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Content not yet available

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

A Zipfian Analysis of Visual Token Distributions for AI-Generated Images

Similar lecture

Localized Cultural Knowledge is Conserved and Controllable in Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads