Multilingual Language Model Evaluation

Idioms remain a persistent challenge in natural language processing due to their figurative and culturally grounded meanings, which distinguish them from literal expressions. Although recent advances in large language models (LLMs) have improved idiom handling across several languages, limited attention has been given to low resource languages such as Urdu. In this work, we present a comprehensive benchmark for Urdu to English idiomatic translation, consisting of a manually verified dataset of 4,000 aligned idiom sentence pairs in both Perso Arabic (native Urdu script) and Romanized Urdu. We evaluate multiple tasks, including translation, paraphrasing, idiom span detection, and back-translation, using diverse prompting strategies such as cultural prompting, idiomatic prompting, and few-shot learning. Our findings show that frontier LLMs consistently outperform traditional neural machine translation systems across all evaluation settings, particularly in preserving figurative and metaphorical meaning. While models demonstrate relatively stable performance on native Urdu script, the absence of standardized orthography in Romanized Urdu introduces substantial challenges for consistency and idiom span detection. This work establishes a high quality benchmark for cross script idiomatic evaluation in Urdu and underscores the importance of prompt engineering in preserving figurative language meaning across languages.

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.

Confidence calibration, the alignment of a model's predicted confidence with its actual accuracy, is crucial for the reliable deployment of Large Language Models (LLMs). However, this critical property remains largely under-explored in multilingual contexts. In this work, we conduct the first large-scale, systematic studies of multilingual calibration across six model families and over 100 languages, revealing that non-English languages suffer from systematically worse calibration. To diagnose this, we investigate the model's internal representations and find that the final layer, biased by English-centric training, provides a poor signal for multilingual confidence. In contrast, our layer-wise analysis uncovers a key insight that late-intermediate layers consistently offer a more reliable and better-calibrated signal. Building on this, we introduce a suite of training-free methods, including Language-Aware Confidence Ensemble (LACE), which adaptively selects an optimal ensemble of layers for each specific language. Our study highlights the hidden costs of English-centric alignment and offer a new path toward building more globally equitable and trustworthy LLMs by looking beyond the final layer.

Large Language Models (LLMs) are increasingly deployed across multilingual and multicultural settings, yet it remains unclear whether changing language leads models to adopt community-specific moral reasoning or merely changes how shared learned abstractions are expressed. We conduct a controlled multilingual evaluation across six geographically, culturally, and linguistically diverse languages (Arabic, Chinese, English, Hindi, Russian, and Spanish), using parallel moral reasoning benchmarks with English-origin, Chinese-origin, and natively elicited ground-truth judgments. Across 13 open-weight LLMs spanning 2B-70B parameters, we find substantial cross-lingual divergence in moral judgments, with English generally achieving the highest performance even when ground-truth judgments originate in Chinese or are collected natively in each language. Yet the reasoning underlying these divergent judgments is considerably more convergent: Utilitarianism dominates in five of six languages, reasoning follows broadly shared stages, and language-specific moral-value associations correspond only sparsely and inconsistently to values measured in the corresponding human communities. Finally, a large-scale OLMoTrace analysis of pretraining data sources reveals little direct reproduction of training text across languages, while the corpus composition, training stage, and cultural provenance of retrieved training evidence vary substantially by response language. Thus, similar moral reasoning structures emerge even from heterogeneous and often linguistically localized training evidence. Our findings, collectively, reveal a central disconnect in multilingual moral reasoning: language changes models' moral judgments and the training evidence associated with their reasoning, but does not correspondingly localize the moral abstractions they apply.

Instruction following is a critical ability for Large Language Models to be used directly by humans. This often requires supervised fine-tuning on curated instruction datasets, sometimes complemented with an alignment step. However, in multilingual scenarios, obtaining high-quality data for these stages remains challenging, motivating the exploration of In-Context Learning (ICL) as a possible alternative. In this work, we study whether ICL can serve as a substitute for Instruction Tuning in multilingual language models, while also examining how the comparison changes with model scale. Our results indicate that a gap remains between ICL and Instruction Tuning, motivating further research to reduce it.

Multilingual watermarking aims to make large language model (LLM) outputs traceable across languages, yet current methods still fall short. Despite claims of cross-lingual robustness, they are evaluated only on high-resource languages. We show that existing multilingual watermarking methods are not truly multilingual: they fail to remain robust under translation attacks in medium- and low-resource languages. We trace this failure to semantic clustering, which fails when the tokenizer vocabulary contains too few full-word tokens for a given language. To address this, we introduce STEAM, a detection method that uses Bayesian optimisation to search among 126 candidate languages for the back-translation that best recovers the watermark strength. It is compatible with any watermarking method, robust across different tokenizers and languages, non-invasive, and easily extendable to new languages. With average gains of +0.23 AUC and +37%p TPR@1%, STEAM provides a scalable approach toward fairer watermarking across the diversity of languages.

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.

Meaningful multilingual evaluation must test models in the target language and educational context. Urdu, spoken by more than 230 million people, lacks a broad MMLU-style benchmark built from native educational sources. We introduce UrduMMLU, a benchmark of 26,389 Urdu MCQs across 26 subjects and five domains, collected from native Urdu MCQ banks and public examination PDFs. Unlike translation-based benchmarks, UrduMMLU combines academic subjects with content specific to Urdu and regional education. We label the exam-derived portion through dual human annotation with strict consensus filtering. We evaluate 30 LLMs under English and Urdu prompts, yielding 60 zero-shot evaluations, and further evaluate four open-source LLMs under multiple few-shot settings across both prompt languages. Gemini-3.5-Flash performs best, reaching 90.23% and 90.45% accuracy, while no other model exceeds 85%. The strongest open-source model trails by 7.78 and 9.12 points, and many models lose 25 to 40 points on Urdu-centered Humanities subjects compared with STEM. Few-shot prompting yields only modest gains. Results on UrduMMLU show that current LLMs have uneven Urdu knowledge, particularly for content grounded in the regional context.

Urdu, the world's tenth most spoken language with 246 million speakers, remains almost entirely absent from mainstream LLM safety evaluation and nine years of WOAH proceedings. To investigate whether this absence has measurable consequences for content moderation reliability, five large language models, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen-2.5, and Llama-3.1, were tested across six datasets spanning Nastaliq Urdu, Roman Urdu, English, and code-switched Urdu-English. Across the five Urdu-script datasets, label instability between original-script and English-translation classification ranged from 15.9% (Gemini 2.5 Flash) to 31.6% (Qwen-2.5), with a 'Missed-in-Urdu' rate, content flagged as harmful in English translation but passed as normal in the original script, ranging from 2.4% to 9.9% (median 4.3%). A complete enumeration of all 205 papers across nine ALW/WOAH editions via the ACL Anthology API confirms zero dedicated Urdu papers across the entire period. Results indicate that current LLMs provide uneven safety assurance across Urdu's script varieties, with smaller open-weight models showing substantially higher instability and missed-harm rates than frontier closed models.