Cultural Knowledge in Language Models

Recent advancements in Natural Language Processing and Deep Learning have enabled the development of Large Language Models (LLMs), which have significantly advanced the state-of-the-art across a wide range of tasks, including Question Answering (QA). Despite these advancements, research on LLMs has primarily targeted high-resourced languages (e.g., English), and only recently has attention shifted toward multilingual models. However, these models demonstrate a training data bias towards a small number of popular languages or rely on transfer learning from high- to under-resourced languages; this may lead to a misrepresentation of social, cultural, and historical aspects. To address this challenge, monolingual LLMs have been developed for under-resourced languages; however, their effectiveness remains less studied when compared to multilingual counterparts on language-specific tasks. In this study, we address this research gap in Greek QA by contributing: (i) DemosQA, a novel dataset, which is constructed using social media user questions and community-reviewed answers to better capture the Greek social and cultural zeitgeist; (ii) a memory-efficient LLM evaluation framework adaptable to diverse QA datasets and languages; and (iii) an extensive evaluation of 11 monolingual and multilingual LLMs on 6 human-curated Greek QA datasets using 3 different prompting strategies. We release our code and data to facilitate reproducibility.

Large language models (LLMs) are trained on web-scale corpora that exhibit steep power-law distributions, in which the distribution of knowledge is highly long-tailed, with most appearing infrequently. While scaling has improved average-case performance, persistent failures on low-frequency, domain-specific, cultural, and temporal knowledge remain poorly characterized. This paper develops a structured taxonomy and analysis of long-Tail Knowledge in large language models, synthesizing prior work across technical and sociotechnical perspectives. We introduce a structured analytical framework that synthesizes prior work across four complementary axes: how long-Tail Knowledge is defined, the mechanisms by which it is lost or distorted during training and inference, the technical interventions proposed to mitigate these failures, and the implications of these failures for fairness, accountability, transparency, and user trust. We further examine how existing evaluation practices obscure tail behavior and complicate accountability for rare but consequential failures. The paper concludes by identifying open challenges related to privacy, sustainability, and governance that constrain long-Tail Knowledge representation. Taken together, this paper provides a unifying conceptual framework for understanding how long-Tail Knowledge is defined, lost, evaluated, and manifested in deployed language model systems.

Large Language Models (LLMs) exhibit inequalities with respect to various cultural contexts. Most prominent open-weights models are trained on Global North data and show prejudicial behavior towards other cultures. Moreover, there is a notable lack of resources to detect biases in non-English languages, especially from Latin America (Latam), a continent containing various cultures, even though they share a common cultural ground. We propose to leverage the content of Wikipedia, the structure of the Wikidata knowledge graph, and expert knowledge from social science in order to create a dataset of question/answer (Q/As) pairs, based on the different popular and social cultures of various Latin American countries. We create the LatamQA database of over 26k questions and associated answers extracted from 26k Wikipedia articles, and transformed into multiple-choice questions (MCQ) in Spanish and Portuguese, in turn translated to English. We use this MCQ to quantify the degree of knowledge of various LLMs and find out (i) a discrepancy in performances between the Latam countries, ones being easier than others for the majority of the models, (ii) that the models perform better in their original language, and (iii) that Iberian Spanish culture is better known than Latam one.

The global deployment of large language models (LLMs) has raised concerns about cultural misalignment, yet the linguistic properties of fine-tuning datasets used for cultural adaptation remain poorly understood. We adopt a dataset-centric view of cultural alignment and investigate which properties of post-training data are associated with cultural performance, whether they can guide data selection before fine-tuning, and how their effects vary across languages and model families. We compute lightweight linguistic, semantic, and structural metrics for Arabic, Chinese, and Japanese datasets and apply principal component analysis (PCA) separately within each language. The resulting components form broadly interpretable axes: PC1 is generally dominated by semantic structure, PC2 captures diversity and lexical variation, and PC3 reflects more language-specific organization. We fine-tune LLaMA, Mistral, and DeepSeek models and evaluate them on benchmarks of cultural knowledge, values, and norms. Although the PCA-derived dataset descriptors are associated with downstream performance, the strongest relationships vary across models, benchmarks, and languages, indicating that no single component serves as a universal predictor. Controlled, size-matched subset interventions further show that PCA-guided selection can improve cultural performance when the relevant component and direction are validated against random sampling. PC3 provides the strongest signal for Arabic, while High-PC1 is most effective for Japanese, particularly for LLaMA. Chinese results are weaker and more model-specific and remain exploratory because of smaller subset sizes. Overall, our findings show that lightweight dataset descriptors can support pre-training data diagnostics, but effective cultural adaptation requires language- and architecture-aware selection rather than a universal linguistic criterion.

In recent years, multilingual Large Language Models (LLMs) have become an inseparable part of daily life, making it crucial for them to master the rules of conversational language in order to communicate effectively with users. While previous work has evaluated LLMs' understanding of figurative language in high-resource languages, their performance in low-resource languages remains underexplored. In this paper, we introduce MasalBench, a comprehensive benchmark for assessing LLMs' contextual and cross-cultural understanding of Persian proverbs, which are a key component of conversation in this low-resource language. We evaluate eight state-of-the-art LLMs on MasalBench and find that they perform well in identifying Persian proverbs in context, achieving accuracies above 0.90. However, their performance drops considerably when tasked with identifying equivalent English proverbs, with the best model achieving 0.79 accuracy. Our findings highlight the limitations of current LLMs in cultural knowledge and analogical reasoning, and they provide a framework for assessing cross-cultural understanding in other low-resource languages. MasalBench is available at https://github.com/kalhorghazal/MasalBench.

Internet audio-visual clips convey meaning through time-varying sound and motion, which extend beyond what text alone can represent. To examine whether AI models can understand such signals in human cultural contexts, we introduce AVMeme Exam, a human-curated benchmark of over one thousand iconic Internet sounds and videos spanning speech, songs, music, and sound effects. Each meme is paired with a unique Q&A assessing levels of understanding from surface content to context and emotion to usage and world knowledge, along with metadata such as original year, transcript, summary, and sensitivity. We systematically evaluate state-of-the-art multimodal large language models (MLLMs) alongside human participants using this benchmark. Our results reveal a consistent limitation: current models perform poorly on textless music and sound effects, and struggle to think in context and in culture compared to surface content. These findings highlight a key gap in human-aligned multimodal intelligence and call for models that can perceive contextually and culturally beyond the surface of what they hear and see. Project page: avmemeexam.github.io/public

We present $\textbf{Korean SimpleQA (KoSimpleQA)}$, a benchmark for evaluating factuality in large language models (LLMs) with a focus on Korean cultural knowledge. KoSimpleQA is designed to be challenging yet easy to grade, consisting of 938 short, fact-seeking questions with unambiguous answers. We conduct a comprehensive evaluation across a diverse set of open-source LLMs of varying sizes that support Korean, and find that even the strongest model generates correct answer only 31.6% of the time, underscoring the challenging nature of KoSimpleQA. Notably, performance rankings on KoSimpleQA differ substantially from those on the English SimpleQA, highlighting the unique value of our dataset. Furthermore, we observe that reasoning helps mitigate the cross-lingual knowledge gap in LLMs, which refers to disparities in their ability to manifest knowledge across languages. KoSimpleQA can be found at https://github.com/naver-ai/KoSimpleQA.

What happens inside a language model when alignment training conflicts with a cultural value it encodes? Across 16 MENA countries, 26 models, and 1.53M human survey responses, we show the answer is suppression, not erasure: at the moment of refusal, a model's internal logit distribution correlates with human survey data more strongly than its freely generated answers. We call this the alignment veto. We distinguish suppression failures (accurate internal distributions blocked at output) from representational bias failures (the encoding itself diverges from human values), and show the two require different interventions. The gate is inequitable: the safety tax reaches 37.6%, with a 19.8% alignment-quality gap between best- and worst-served nations, and native-language prompting widens rather than closes it. Sparse autoencoder analysis corroborated by comparisons across alignment stages identifies a candidate DPO-stage feature mediating suppression in Tulu-3-8B. The gate is real, its costs are unequal, and deciding what it should protect is not a technical question.

As Large Language Models (LLMs) develop stronger multilingual capabilities, their sensitivity to culturally diverse entities becomes increasingly important. Prior work by Naous et al. (2024) has shown that LLMs often favor Western-associated entities in Arabic. Due to the lack of entity-centric multilingual benchmarks, it remains unclear if such biases also manifest in various non-Western languages. In this paper, we introduce Camellia, a benchmark for evaluating entity-centric cultural biases in nine Asian languages, spanning six Asian cultures. Camellia includes 19,530 manually annotated entities associated with the covered Asian or Western cultures, as well as 2,173 masked contexts for these entities derived from social media posts. Using Camellia, we evaluate cultural biases in four recent multilingual LLMs across three tasks: cultural context adaptation, sentiment association, and entity extractive QA. Our analyses show that LLMs struggle with cultural adaptation across these languages, with performance differing across models developed in different regions. We further observe that different LLM families can hold distinct biases, reflected in the ways they link cultures to particular sentiments. Lastly, we find that LLMs can struggle with context understanding in some Asian languages, creating performance gaps between cultures in entity extraction.

Generative AI (GenAI) is entering K-12 classrooms, but its default culture does not always fit local teaching. Existing accounts either treat this misfit as macro-level structural bias or examine it through single-case studies; neither captures how much work teachers must do to make outputs usable across tasks. This exploratory, theory-building qualitative study introduces cultural distance, the gap between GenAI's default cultural repertoire and the situated demands of teachers' practice. Through interviews with 30 K-12 teachers who frequently use GenAI, 10 each from South Africa, Taiwan, and the United States, we identified three levels defined by the effort teachers invested and whether it yielded a usable output: low (minor edits), mid (sustained prompting and revision), and high (structural gaps rooted in unsupported languages, undigitized knowledge, or policy filters that no teacher effort can close). We present cultural distance as a promising framework requiring further examination and draw implications for designers, policymakers, and educators.

Aligning LLMs with Chinese values requires culturally grounded resources that connect abstract values to concrete behavioral guidance. We introduce C-VARC, a Chinese Value Rule Corpus containing 257,609 Rules of Thumb (RoTs) organized into three dimensions, 12 core values, and 50 derived values. Among them, 36,000 rules underwent independent human review by two annotators (Cohen's kappa>0.85), with disagreements adjudicated by a third. C-VARC supports reusable rule-driven pipelines for both value evaluation and model alignment. Compared with abstract-value guidance, RoT guidance produces greater semantic and character-level n-gram diversity. Across five non-political sensitive themes, the seven-model average selection rate for C-VARC-consistent options is at least 72.0%, while five Chinese annotators show overall rates of at least 87.5%. Conflicting RoT pairs further enable the construction of 404,505 moral dilemmas, with 7,191 of 10,998 evaluated cases eliciting cross-model disagreement. For alignment, C-VARC-Query consistently outperforms matched Value-only supervision on CMoralEval across both Qwen3-8B and Llama-3.1-8B-Instruct, with average gains of 2.97 and 12.82 percentage points, respectively, and positive paired gains across all three seeds for both backbones. It also improves CValues and C-Eval over Value-only supervision on both backbones. C-VARC thus provides a scalable Chinese normative resource for constructing both value-evaluation and alignment supervision.

Often, multilingual language models are trained with the objective to map semantically similar content (in different languages) in the same latent space. In this paper, we show a nuance in this training objective, and find that by changing the language of the input query, we can improve the question answering ability of language models. We make two main contributions. First, we introduce the term Language Specific Knowledge (LSK) to denote queries that are best answered in an ``expert language'' for a given LLM, thereby enhancing its question-answering ability. We introduce the problem of language selection -- for some queries, language models can perform better when queried in languages other than English, sometimes even better in low-resource languages -- and the goal is to select the optimal language for the query. Second, we introduce a variety of simple to strong baselines to empirically motivate the language selection problem (including one of our own methods called LSKExtractor). During our evaluation, we employ three datasets that contain knowledge about both cultural and social behavioral norms. Overall, the results show that principled language selection can improve the performance of a language model, and that the expected question-to-language map is not always intuitive: Gemma models know most about China and Middle East in Spanish; Qwen models know most about authority and responsibility in Arabic and Chinese. Broadly, our research contributes to the open-source development of language models that are inclusive and more aligned with the cultural and linguistic contexts in which they are deployed.