United States

Multilingual large language models exhibit systematic differences in their outputs across languages, even when representing the same underlying knowledge. Prior work has primarily focused on evaluating or reducing such inconsistencies. In this work, we instead study whether cross-lingual behavior can be controlled: specifically, whether answer distributions associated with other languages can be expressed under English prompting. To this end, we construct a human-annotated factual dataset and a cultural scenarios dataset, and compare intervention methods including persona prompting, activation steering, and preference-based fine-tuning. We evaluate how these methods affect answer distributions and their generalization to culturally grounded settings. Our results show that answer distributions can be systematically shifted toward those observed in other languages, with persona prompting consistently outperforming more complex intervention methods.

ACL 2026

Controlling Cross-Lingual Answer Distributions in Language Models: Enabling Transfer of Factual Preferences

multilingual large language models

cross-lingual knowledge transfer

factual consistency

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

workshop paper

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. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<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>

To access this event page, you need to log in with the **email address you registered with**. 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.

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.

Large language models (LLMs) are increasingly explored for patient-facing medical advice and symptom triage, yet their responses may shift when identical clinical evidence is paired with culturally marked patient descriptors. 
We present a counterfactual audit framework for evaluating cross-cultural variation in LLM-generated medical advice by isolating identity-related cues while holding clinical evidence constant.
Our evaluation uses matched clinical vignettes, cross-regional and culturally marked prompt variants, repeated sampling, and structured comparison of urgency framing, safety recommendations, empathy, and escalation advice.
Across multiple commercial and open-weight LLMs, we observe measurable identity-conditioned variation in both triage decisions and interactional framing. In several cases, culturally marked descriptors shift urgency assessments or escalation recommendations despite unchanged clinical evidence. While the magnitude and direction of these effects differ across models, the results suggest that LLM-generated medical advice remains sensitive to culturally linked identity cues in ways that may affect safety-critical guidance.
Our results demonstrate how culturally grounded counterfactual auditing can help identify clinically unsupported variation while distinguishing potentially harmful shifts from appropriate communication adaptation in patient-facing medical advice.

Counterfactual Auditing of Cross-Cultural Variation in LLM-Generated Medical Advice

Large language models (LLMs) perpetuate cultural stereotypes not only through biased associations but through systematic omission and orthographic erasure of underrepresented languages. We present empirical evidence of two compounding failure modes affecting Northeast Indian languages: (1) entity-level invisibility, where state-of-the-art NER systems score F1=0.000 on culturally critical named entities such as Khasi surnames, Garo festivals, and tribal names; and (2) orthographic corruption, where LLM tokenizers corrupt semantically meaningful diacritics (ï, ñ) and the Garo morpheme boundary marker (U+00B7) at rates of 18.8–50% across four of five evaluated models. Drawing on NortheastNER (F1=0.964, six entity categories, XLM-RoBERTa-base) and a systematic tokenization study across Khasi and Garo, we argue that stereotype-by-omission constitutes a distinct and measurable harm to indigenous language communities. We further show that a custom multilingual tokenizer achieves 26–50% token reduction over five baseline LLMs, demonstrating that culturally grounded infrastructure can partially remediate these failures. Our findings call for cultural representation audits as a standard component of multilingual NLP evaluation.

Stereotyped by Silence: How LLMs Erase Northeast Indian Languages Through Omission and Orthographic Corruption

Stereotypes associate a group (G) with an attribute (A), often through broad generalizations like "Gs are A". In the age of large language models, many existing datasets and methods measure bias by considering changes in $P(A|G)$ between different social groups ($G_{alt}$). Motivated by literature on generics, we argue that constructing minimal pairs along the group dimension is not enough: explicitly considering the alternate attribute ($A_{alt}$) is necessary for a robust score. We first show how to adapt existing log-probability metrics to this setup, but we find that aggregating them across different stereotypes is inherently flawed. We propose a new metric representing the mutual information (MI) between the social groups and attributes. We argue that it is better suited for aggregation across individual points, allowing for robust comparison of stereotype strength across different languages and models. We demonstrate this on a set of paired English and Russian stereotypes from the BiasShades dataset.

The Missing Minimal Pair: Mutual Information for Stereotype Evaluation in LLMs

Real-world social judgments often depend on local norms, values, and situational expectations. We study whether multilingual language models use such cultural context robustly, or whether their judgments are overly sensitive to surface cues such as prompt language, country labels, and externally supplied contextual descriptions. Building on NormAd, a benchmark for cultural norm adherence, we keep the underlying scenario fixed while perturbing language, context granularity, and contextual consistency. We evaluate 12 models from four model families: Cohere Command, Cohere Tiny-Aya, Alibaba Qwen, and Google Gemma. Across three countries, five language conditions, and 55,000 raw generations, models benefit most from explicit norm descriptions. Adding these descriptions improves accuracy by 17.8 percentage points over country-only prompts. In contrast, adding externally supplied rules that are not applicable to the scenario is the largest harmful perturbation, reducing accuracy by 24.5 points. Language changes produce 9–14% prediction flips, but these flips are directionally balanced. These results show why flip rate alone is insufficient: it does not distinguish between correct-to-wrong (C→W) and wrong-to-correct (W→C) transitions. We therefore report cultural robustness using signed accuracy change, C→W/W→C decompositions, and perturbation-specific vulnerability alongside clean accuracy.

Robustness of Cultural Norm Reasoning Under Language and Context Perturbations

As language models scaled to internet data, exhaustive curation became infeasible and stereotype learning became an expected side effect of training. A common response is post-training local editing: identify an internal direction linked to harmful behavior and try to remove it. However, the causal effectiveness of this method is unclear. In this paper, we present a causal audit of local stereotype editing across seven open-weight LLMs using matched inject-versus-remove interventions, component-level diagnostics, and cross-source transfer stress tests. The main result is asymmetric controllability: stereotype preference is easier to increase than to reliably decrease. In matched tests, score-level inject-minus-remove asymmetry is significant in 5/7 models, and margin-level asymmetry is significant in 7/7. Supporting diagnostics show model-dependent sign reliability, suppressor contamination in ranked intervention sets, and no reliable large-effect cross-source transfer under current settings. Overall, localization provides control leverage but not a debiasing guarantee; deployment-facing claims should be validated with direct causal intervention and transfer stress tests.

Easy to Add, Hard to Erase: Causal Evidence on Stereotype Editing in LLMs

MLLM-as-a-Judge is conventionally validated by agreement with human annotations, but agreement against a culturally heterogeneous pool is pool-dependent: it silently conditions on whose judgments the aggregate represents. We introduce VOIR DIRE, a multimodal benchmark of 626 culturally paired image–prompt items spanning U.S. and mainland Chinese contexts across food, fashion, and architecture, with annotator pools that are within-pool reliable (α = 0.86/0.74) but cross-pool divergent on evaluation (Q1 r = −0.12). Across six MLLMs, the bias decomposes into two failures: a positivity-floor calibration failure (compressed scale use) and an orientation failure (default to one cultural norm). On this corpus, where contested items are sampled to split the two pools, the floor mechanically validates the more permissive Chinese reading; persona prompting partially recovers calibration, but the orientation residual survives, evidence the tilt is not reducible to scale compression. Reference pool in-context demonstrations deepen the orientation residual and inflate the high end rather than restoring use of the low end. Model origin adds a small additive tilt (~ 0.10 MAE) that is approximately invariant under demonstration. We recommend reporting alignment against each reference pool separately and treating cross-pool divergence as a judge property.

Jury Duty: Uncovering Latent Cultural Biases in MLLM-as-a-Judge Under Ambiguity

We present an open-source measurement protocol for stereotype interpretation that quantifies how users translate or interprets provocative discourse without assuming a normative direction. Building on Deleuze and Guattari's rhizomatic framework, we operationalize three modes of semantic movement —Reaffirm, De-signify, and Escape (RDE)— through an abstract-machine operator detector that combines transparent linguistic patterns (526 patterns across 8 languages) with optional contextual embeddings. The protocol is direction-agnostic: it measures equally well a user who reproduces their own semantic territory and one who departs from it, capturing diasporic, assimilationist, and escape trajectories that English-centric, Chomskyan-hierarchical taxonomies obscure. We demonstrate the protocol on five extreme user profiles (Russian conservative, Russian diaspora, trans Russian exile, Mexican malinchista, Mapuche speaker), each producing coherent and distinct RDE signatures. Deployed in a free-tier web service, the protocol enables both individual reflective use and corporate calibration of tolerable territoriality ranges for personnel engaged in intercultural translation and interpretation tasks.

Measuring Semantic Flow Without Direction: A Rhizomatic Protocol for Stereotype Translation in Cross-Cultural Language Technology

Socio-cultural stereotypical bias is an important consideration in the development and deployment of NLP systems. It is however often considered only at the national level, despite rich subnational socio-cultural structures. We present AmchiBias, the first benchmark for enmeasuring socio-cultural stereotypical bias for the Indian state of Goa with its unique historically multicultural setting. It covers various Goan identity groups and comprises 313 minimal pairs across eight sociodemographic dimensions in both English and Devanagari Konkani. 
We then evaluate stereotypical bias in five multilingual encoder models on this benchmark. We find near-chance scores in Konkani, reflecting language incompetence for general multilingual models and a lack of Goan cultural competence for Indian language models. 
Queried in English, models with a stronger Indian language coverage show higher bias for pan-Indian groups than hyperlocal Goan groups. This suggests the English signal reflects pan-Indian pretraining associations rather than genuine Goan cultural knowledge. Our findings highlight a critical gap in low-resource multilingual NLP evaluation for hyperlocal community identities.

AmchiBias: Measuring Stereotypical Bias in Goan Identity Groups with a Minimal Pair Dataset in English and Konkani

While the large language models (LLMs) are considered to be largely proficient in translation, especially among relatively high-resource languages, we posit that they are inadequate at translating colloquial and informal cultural usage. Translation of cultural involves localization at various levels of granularity, identifying equivalents for comparable values and concepts, and elaborating on fundamental conceptual differences. In this work, we show that even for a narrow instantiation of this task that focuses only on proverb translation, LLMs exhibit significant asymmetry in translating cultural terms used by English speakers into other languages compared to the translation of proverbs in the other (even high-resource) languages to English. We compare several different families of language models and observe this asymmetry even extends to model-based translation evaluation metrics like COMET. Further analysis reveals that LLMs typically tend to be literal in their translations and often fail to find distinct but equivalent concepts across languages as often as human translators, thus raising deployment concerns around differential service across language users worldwide.

Unveiling Asymmetry in Large Language Models' Ability to Translate Culture

Foundation models pretrained on internet-scale data benefit substantially from post-training alignment with human preferences. However, existing alignment techniques typically assume a homogeneous notion of human preference, overlooking the cross-regional diversity of human values. As LLMs are increasingly used for everyday and professional decision-making, ideation, and opinion formation, homogeneous preference modeling may drive convergence toward dominant values, suppressing diverse perspectives and reinforcing stereotypes.
Pluralistic reward modeling is therefore essential for building AI systems that better reflect the pluralistic values of human society. In this work, we propose **PAL**, a general reward modeling framework for pluralistic alignment. **PAL** captures *cross-user commonalities* with a shared reward backbone, while modeling *individual variation* through personalized weights over learnable group-level preference prototypes. By adapting only these weights, **PAL** enables sample-efficient preference localization for new users.
Extensive empirical evaluation demonstrates that **PAL** ***matches or outperforms state-of-the-art methods on both text-to-text and text-to-image tasks***. Finally, we provide a theoretical analysis for **PAL**, showing that its shared structure ***reduces the per-user sample complexity of personalized reward learning***.

Downloads

Next from ACL 2026

Counterfactual Auditing of Cross-Cultural Variation in LLM-Generated Medical Advice

Similar lecture

Disentangling Language Understanding and Reasoning Structures in Cross-lingual Chain-of-Thought Prompting

Stay up to date with the latest Underline news!

PRODUCT

COMPANY

RESOURCES

.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

Counterfactual Auditing of Cross-Cultural Variation in LLM-Generated Medical Advice

Similar lecture

Disentangling Language Understanding and Reasoning Structures in Cross-lingual Chain-of-Thought Prompting

Stay up to date with the latest Underline news!

PRODUCT

COMPANY

RESOURCES

Downloads