Recent advances in Natural Language Processing (NLP) have underscored the crucial role of high-quality datasets in building large language models (LLMs). However, while extensive resources and analyses exist for English, the landscape for East Asian languages, particularly Chinese, Japanese, and Korean (CJK), remains fragmented and underexplored, despite these languages serving over 1.6 billion speakers. To address this gap, we investigate the HuggingFace ecosystem from a cross-linguistic perspective, focusing on how cultural norms, research environments, and institutional practices shape dataset availability and quality. Drawing on more than 3,300 datasets, we employ quantitative and qualitative methods to examine how these factors drive distinct creation and curation patterns across Chinese, Japanese, and Korean NLP communities. Our findings highlight the large-scale and often institution-driven nature of Chinese datasets, grassroots community-led development in Korean NLP, and an entertainment and subculture-focused emphasis on Japanese collections. By uncovering these patterns, we reveal practical strategies for enhancing dataset documentation, licensing clarity, and cross-lingual resource sharing, guiding more effective and culturally attuned LLM development in East Asia. We conclude by discussing best practices for future dataset curation and collaboration, aiming to strengthen resource development across all three languages.

Natural Language Inference (NLI) is a fundamental task in natural language processing. While NLI has developed many sub-directions such as sentence-level NLI, document-level NLI and cross-lingual NLI, Cross-Document Cross-Lingual NLI (CDCL-NLI) remains largely unexplored. In this paper, we propose a novel paradigm: CDCL-NLI, which extends traditional NLI capabilities to multi-document, multilingual scenarios. To support this task, we construct a high-quality CDCL-NLI dataset including 25,410 instances and spanning 26 languages. To address the limitations of previous methods on CDCL-NLI task, we further propose an innovative method that integrates RST-enhanced graph fusion with interpretability-aware prediction. Our approach leverages RST (Rhetorical Structure Theory) within heterogeneous graph neural networks for cross-document context modeling, and employs a structure-aware semantic alignment based on lexical chains for cross-lingual understanding. For NLI interpretability, we develop an EDU (Elementary Discourse Unit)-level attribution framework that produces extractive explanations. Extensive experiments demonstrate our approach”s superior performance, achieving significant improvements over both conventional NLI models as well as large language models. Our work sheds light on the study of NLI and will bring research interest on cross-document cross-lingual context understanding, hallucination elimination and interpretability inference. Our dataset and code are available at https://anonymous.4open.science/r/CDCL-NLI-637E/ for peer review.

Grammatical gender is a fundamental linguistic feature that varies across languages, and its cross-linguistic correspondence has been a central question in disciplines such as cognitive science and linguistic typology. This study takes an information-theoretic approach to investigate the extent to which variational usable information about grammatical gender encoded by a large language model generalizes across languages belonging to different language families. Using mBERT as a case study, we analyze how grammatical gender is encoded and transferred across languages based on the usable information of the intermediate representations. The empirical results provide evidence that gender mechanisms are driven by abstract semantic features largely shared across languages, and that the information becomes more accessible at the higher layers of the language model.

Multilingual large language models have emerged as a promising solution for resource-constrained settings, with significant efforts aimed towards improving multilingual capabilities of English-centric pretrained models. However, the broader cross-lingual implications of fine-tuning interventions remain understudied. This work examines instruction tuning (IT) over the BLOOMZ model for Question Answering (QA) in low-resource settings, with special emphasis on transfer dynamics across several languages. Our findings reveal two critical insights: first, IT on the target language can negatively impact its own performance in constrained short-span generation tasks due to overgeneration tendencies; second, in QA tasks, IT appears to suppress performance in some interfering languages, thereby enhancing capabilities in some target Indic languages by extbfmore than doubling QA performance. These results highlight important trade-offs in multilingual LLM adaptation and enhance our understanding of cross-lingual transfer mechanisms.

Multilingual Large Language Models (LLMs) develop cross-lingual abilities despite being trained on limited parallel data. However, they often struggle to generate responses in the intended language, favoring high-resource languages such as English. In this work, we introduce CoCo-CoLa (Correct Concept - Correct Language), a novel metric to evaluate language adherence in multilingual LLMs. Using fine-tuning experiments on a closed-book QA task across seven languages, we analyze how training in one language affects others’ performance. Our findings reveal that multilingual models share task knowledge across languages but exhibit biases in the selection of output language. We identify language-specific layers, showing that final layers play a crucial role in determining output language. Accordingly, we propose a partial training strategy that selectively fine-tunes key layers, improving language adherence while reducing computational cost. Our method achieves comparable or superior performance to full fine-tuning, particularly for low-resource languages, offering a more efficient multilingual adaptation.

Large language models (LLMs) demonstrate exceptional performance on complex reasoning tasks. However, despite their strong reasoning capabilities in high-resource languages (e.g., English and Chinese), a significant performance gap persists in other languages. To investigate this gap in Korean, we introduce HRM8K, a benchmark comprising 8,011 English-Korean parallel bilingual math problems. Through systematic analysis of model behaviors, we identify a key finding: these performance disparities stem primarily from difficulties in comprehending non-English inputs, rather than limitations in reasoning capabilities. Based on these findings, we propose UST(Understand, Solve, and Translate), a method that strategically uses English as an anchor for reasoning and solution generation. By fine-tuning the model on 130k synthetically generated data points, method achieves a 10.91% improvement on the HRM8K benchmark and reduces the multilingual performance gap from 11.6%% to 0.7%%. Additionally, we show that improvements from method generalize effectively to different Korean domains, demonstrating that capabilities acquired from machine-verifiable content can be generalized to other areas. We publicly release the benchmark, training dataset, and models.

Recent advances in LLMs have enhanced AI capabilities, but also increased the risk posed by malicious requests, highlighting the need for effective LLM safeguards to detect such queries. Existing approaches largely rely on classifier-based methods that lack interpretability and perform poorly on low-resource languages. To address these limitations, we propose ConsistentGuard, a novel reasoning-based multilingual safeguard, which enhances explainability via reasoning and boosts knowledge transfer between languages through alignment. With only 1,000 training samples, our method demonstrates superior performance on three datasets across six languages, outperforming larger models trained with significantly more data, and exhibits strong interpretability and generalization ability. We also contribute a multilingual benchmark extension and release our code to support future research.

Named-entity recognition (NER) in low-resource languages is usually tackled by finetuning very large multilingual LMs, an option that is often infeasible in memory- or latency-constrained settings. We ask whether small decoder LMs can be pretrained so that they adapt quickly and transfer zero-shot to languages unseen during pretraining. To this end we replace part of the autoregressive objective with first-order model-agnostic meta-learning (MAML). Tagalog and Cebuano are typologically similar yet structurally different in their actor/non-actor voice systems, and hence serve as a challenging test-bed. Across four model sizes (11 M – 570 M) MAML lifts zero-shot micro-F1 by 2–6 pp under head-only tuning and 1–3 pp after full tuning, while cutting convergence time by up to 8%. Gains are largest for single-token person entities that co-occur with Tagalog case particles si/ni, highlighting the importance of surface anchors.

Do independently trained monolingual language models converge on shared linguistic principles? To explore this question, we propose to analyze a suite of models trained separately on single languages but with identical architectures and budgets. We train sparse autoencoders (SAEs) on model activations to obtain interpretable latent features, then align them across languages using activation correlations. We do pairwise analyses to see if feature spaces show non-trivial convergence, and we identify universal features that consistently emerge across diverse models. Positive results will provide evidence that certain high-level regularities in language are rediscovered independently in machine learning systems.

Large language models (LLMs) still struggle across tasks outside of high-resource languages. In this work, we investigate cross-lingual transfer to lower-resource languages where task-specific post-training data is scarce. Building on prior work, we first validate that the subsets of model parameters that matter most for mathematical reasoning and multilingual capabilities are distinctly non-overlapping. To exploit this implicit separability between task and target language parameterization, we develop and analyze numerous modular frameworks to improve the composition of the two during fine-tuning. These methods generally employ freezing parameters or post hoc model merging to assign math and language improvement to different key parts of the LLM. In the absence of in-language math data, we demonstrate that the modular approaches successfully improve upon baselines across three languages, four models, and two fine-tuning paradigms (full and LoRA). Furthermore, we identify the most consistently successful modular method to be fine-tuning separate language and math experts and model merging via Layer-Swapping, somewhat surprisingly. We offer possible explanations for this result via recent works on the linearity of task vectors. We further explain this by empirically showing that reverting less useful fine-tuning updates after training often outperforms freezing them from the start.

Automatic speech translation (AST) promotes seamless communication among speakers of different languages. While current state-of-the-art models excel with high-resource languages, their performance on low-resource languages (LRLs) is not well-established. We investigate this by evaluating state-of-the-art models on 10 LRLs with varying data amounts (10-30+ hours). Through six finetuning strategies and experimenting with three main AST paradigms, we observe that: (1) The latest Large Language Models (LLMs) may struggle with LRLs. (2) Comprehensive experiments suggest that for LRLs, more AST finetuning data is not always beneficial. (3) Our 2-Stage with ASR corrector finetuning recipe can substantially improve AST performance on LRLs, achieving up to a 5.8x BLEU score boost on translating related languages to English, while on par with the best monolingual finetuning in BLEU score when translating the target language to English. (4) We share effective engineering practices, including how to effectively adapt AST models to unseen languages.

Multilingual Retrieval-Augmented Generation (mRAG) often retrieves English documents and translates them into the query language for low-resource settings. However, poor translation quality degrades response generation performance. Existing approaches either assume sufficient translation quality or utilize the rewriting method, which introduces factual distortion and hallucinations. To mitigate these problems, we propose Quality-Aware Translation Tagging in mRAG (QTT-RAG), which explicitly evaluates translation quality along three dimensions-semantic equivalence, grammatical accuracy, and naturalness&fluency-and attach these scores as metadata without altering the original content. We evaluate QTT-RAG against CrossRAG and DKM-RAG as baselines in two open-domain QA benchmarks (XORQA, MKQA) using six instruction-tuned LLMs ranging from 2.4B to 14B parameters, covering two low-resource languages (Korean and Finnish) and one high-resource language (Chinese). QTT-RAG outperforms the baselines by preserving factual integrity while enabling generator models to make informed decisions based on translation reliability. This approach allows for effective usage of cross-lingual documents in low-resource settings with limited native language documents, offering a practical and robust solution across multilingual domains.

Existing multilingual neural machine translation (MNMT) approaches mainly focus on improving models with the encoder-decoder architecture to translate multiple languages. However, decoder-only architecture has been explored less in MNMT due to its underperformance when trained on parallel data solely. In this work, we attribute the issue of the decoder-only architecture to its lack of language transfer capability. Specifically, the decoder-only architecture is insufficient in encoding source tokens with the target language features. We propose dividing the decoding process into two stages so that target tokens are explicitly excluded in the first stage to implicitly boost the transfer capability across languages. Additionally, we impose contrastive learning on translation instructions, resulting in improved performance in zero-shot translation. We conduct experiments on TED-19 and OPUS-100 datasets, considering both training from scratch and fine-tuning scenarios.results show that, compared to the encoder-decoder architecture, our methods not only perform competitively in supervised translations but also achieve improvements of up to 3.39 BLEU, 6.99 chrF++, 3.22 BERTScore, and 4.81 COMET in zero-shot translations. We release our codes at https://github.com/zhiqu22/PhasedDecoder.

The extent to which individual language characteristics influence tokenization and language modeling is an open question. Differences in morphological systems have been suggested as both unimportant and crucial to consider (e.g., Cotterell et al., 2018; Park et al., 2021, Arnett & Bergen, 2025). We argue this conflicting evidence is due to confounding factors in experimental setups, making it hard to compare results and draw conclusions. We identify confounding factors in analyses trying to answer the question of whether, and how, morphology relates to language modeling. Next, we introduce token bigram metrics as an intrinsic way to predict the difficulty of causal language modeling, and find that they are gradient proxies for morphological complexity that do not require expert annotation. Ultimately, we outline necessities to reliably answer whether, and how, morphology relates to language modeling.

Retrieval-augmented generation (RAG) with large language models (LLMs) has demonstrated strong performance in multilingual question-answering (QA) tasks by leveraging relevant passages retrieved from corpora. In multilingual RAG (mRAG), the retrieved passages can be written in languages other than that of the query entered by the user, making it challenging for LLMs to effectively utilize the provided information. Recent research suggests that retrieving passages from multilingual corpora can improve RAG performance, particularly for low-resource languages. However, the extent to which LLMs can leverage different kinds of multilingual contexts to generate accurate answers, independently from retrieval quality, remains understudied. In this paper, we conduct an extensive assessment of LLMs’ ability to (i) make consistent use of a relevant passage regardless of its language, (ii) respond in the expected language, and (iii) focus on the relevant passage even when multiple ‘distracting passages’ in different languages are provided in the context. Our experiments with four LLMs across three QA datasets covering 48 languages reveal a surprising ability of LLMs to extract relevant information from passages in a different language than the query, but a much weaker ability to produce a full answer in the correct language. Our analysis, based on both accuracy and feature attribution techniques, further shows that distracting passages negatively impact answer quality regardless of their language. However, distractors in the query language exert a slightly stronger influence. Taken together, our findings deepen the understanding of how LLMs utilize context in mRAG systems, providing directions for future improvements. All codes and data are released at https://github.com/Betswish/mRAG-Context-Consistency.

Cross-lingual information retrieval (CLIR) helps users find documents in languages different from their queries. This is especially important in academic search, where key research is often published in non-English languages. We present CLIRudit, a novel English-French academic retrieval dataset built from Érudit, a Canadian publishing platform. Using multilingual metadata, we pair English author-written keywords as queries with non-English abstracts as target documents, a method that can be applied to other languages and repositories. We benchmark various first-stage sparse and dense retrievers, with and without machine translation. We find that dense embeddings without translation perform nearly as well as systems using machine translation, that translating documents is generally more effective than translating queries, and that sparse retrievers with document translation remain competitive while offering greater efficiency. Along with releasing the first English-French academic retrieval dataset, we provide a reproducible benchmarking method to improve access to non-English scholarly content.

Multilingual language models excel across languages, yet how they internally encode grammatical tense remains largely unclear. We investigate how decoder-only transformers represent, transfer, and control tense across eight typologically diverse languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai. We construct a synthetic tense-annotated dataset and combine probing, causal analysis, feature disentanglement, and model steering to LLaMA-3.1 8B. We show that tense emerges as a distinct signal from early layers and transfers most strongly within the same language family. Causal tracing reveals that attention outputs around layer 16 consistently carry cross-lingually transferable tense information. Leveraging sparse autoencoders in this subspace, we isolate and steer English tense-related features, improving target-tense prediction accuracy by up to 11%% in a downstream cloze task.

We propose an unsupervised method to disentangle sentence embeddings from multilingual sentence encoders into language-specific and language-agnostic representations. Such language-agnostic representations distilled by our method can estimate cross-lingual semantic sentence similarity by cosine similarity. Previous studies have trained individual extractors to distill each language-specific and -agnostic representation. This approach suffers from missing information resulting in the original sentence embedding not being fully reconstructed from both language-specific and -agnostic representations; this leads to performance degradation in estimating cross-lingual sentence similarity. We only train the extractor for language-agnostic representations and treat language-specific representations as differences from the original sentence embedding; in this way, there is no missing information. Experimental results for both tasks, quality estimation of machine translation and cross-lingual sentence similarity estimation, show that our proposed method outperforms existing unsupervised methods.

Developing a high-performing large language models (LLMs) for low-resource languages such as Urdu, present several challenges. These challenges include the scarcity of high-quality datasets, multilingual inconsistencies, and safety concerns. Existing multilingual LLMs often address these issues by translating large volumes of available data. However, such translations often lack quality and cultural nuance while also incurring significant costs for data curation and training. To address these issues, we propose Alif-1.0-8B-Instruct, a multilingual Urdu-English model, that tackles these challenges with a unique approach. We train the model on a high-quality, multilingual synthetic dataset (Urdu-Instruct), developed using a modified self-instruct technique. By using unique prompts and seed values for each task along with a global task pool, this dataset incorporates Urdu-native chain-of-thought based reasoning, bilingual translation, cultural relevance, and ethical safety alignments. This technique significantly enhances the comprehension of Alif-1.0-8B-Instruct model for Urdu-specific tasks. As a result, Alif-1.0-8B-Instruct, built upon the pretrained Llama-3.1-8B, demonstrates superior performance compared to Llama-3.1-8B-Instruct for Urdu specific-tasks. It also outperformed leading multilingual LLMs, including Mistral-7B-Instruct-v0.3, Qwen-2.5-7B-Instruct, and Cohere-Aya-Expanse-8B, all within a training budget of under $100. Our results demonstrate that high-performance and low-resource language LLMs can be developed efficiently and culturally aligned using our modified self-instruct approach.

The effectiveness of Large Language Models (LLMs) depends heavily on the availability of high-quality post-training data, particularly instruction-tuning and preference-based examples. Existing open-source datasets, however, often lack multilingual coverage, cultural grounding, and suffer from task diversity gaps that are especially pronounced for Indian languages. We introduce a human-in-the-loop pipeline that combines translations with synthetic expansion to produce reliable and diverse Indic post-training data. Using this pipeline, we curate two datasets: Pragyaan-IT (22.5K) and Pragyaan-Align (100K) across 10 Indian languages covering 13 broad and 56 sub-categories, leveraging 57 diverse datasets. Our dataset protocol incorporates several often-overlooked dimensions and emphasize task diversity, multi-turn dialogue, instruction fidelity, safety alignment, and preservation of cultural nuance, providing a foundation for more inclusive and effective multilingual LLMs.

This work investigates the impact of multi-task, multi-lingual, and multi-source learning approaches on the robustness and performance of pretrained language models. To enhance this analysis, we introduce Subsets of Interest (SOI), a novel categorization framework that identifies six distinct learning behavior patterns during training, including forgettable examples, unlearned examples, and always correct examples. Through SOI transition heatmaps and dataset cartography visualization, we analyze how examples shift between these categories when transitioning from single-setting to multi-setting configurations. We perform comprehensive experiments across three parallel comparisons: multi-task vs. single-task learning using English tasks (entailment, paraphrase, sentiment), multi-source vs. single-source learning using sentiment analysis datasets, and multi-lingual vs. single-lingual learning using intent classification in French, English, and Persian. Our results demonstrate that multi-source learning consistently improves out-of-distribution performance by up to 7%, while multi-task learning shows mixed results with notable gains in similar task combinations. We further introduce a two-stage fine-tuning approach where the second stage leverages SOI-based subset selection to achieve additional performance improvements. These findings provide new insights into training dynamics and offer practical approaches for optimizing multi-setting language model performance.

Multilingual Neural Machine Translation (MNMT) models enhance translation quality for low-resource languages by exploiting cross-lingual similarities during training—a process known as knowledge transfer. This transfer is particularly effective between languages that share lexical or structural features, often enabled by a common orthography. However, languages with strong phonetic and lexical similarities but distinct writing systems experience limited benefits, as the absence of a shared orthography hinders knowledge transfer. To address this limitation, we propose an approach based on phonetic information that enhances token-level alignment across scripts by leveraging transliterations. We systematically evaluate several phonetic transcription techniques and strategies for incorporating phonetic information into NMT models. Our results show that using a shared encoder to process orthographic and phonetic inputs separately consistently yields the best performance for Khmer, Thai, and Lao in both directions with English, and that our custom Cognate-Aware Transliteration (CAT) method consistently improves translation quality over the baseline.

Fine-tuning multilingual foundation models on specific languages often induces catastrophic forgetting, degrading performance on languages unseen in fine-tuning. While this phenomenon is widely-documented, the literature presents fragmented results about when forgetting occurs. To address this ambiguity, we conduct a systematic empirical study using machine translation as a testbed to identify the conditions that trigger catastrophic forgetting in multilingual fine-tuning. Through controlled experiments across different model architectures, data scales, and fine-tuning approaches, we reveal that the relative scale between model and data size is a primary determinant of forgetting. Moreover, we demonstrate that a model’s instruction-following ability is more critical for retaining multilingual knowledge than its architecture. Contrary to assumptions, parameter-efficient fine-tuning offers no clear advantage over full fine-tuning in mitigating forgetting. Lastly, we show that cross-lingual alignment can mitigate forgetting while also facilitating positive transfer to unseen target languages.

Multilingual alignment for low-resource languages is a challenge for embedding models. The scarcity of parallel datasets in addition to rich morphological diversity in languages adds to the complexity of training multilingual embedding models. To aid in the development of multilingual models for under-represented languages such as Sanskrit, we introduce GitaDB: a collection of 640 Sanskrit verses translated in 5 Indic languages and English. We benchmarked various state-of-the-art embedding models on our dataset in different bilingual and cross-lingual semantic retrieval tasks of increasing complexity and found a steep degradation in retrieval scores. We found a wide margin in the retrieval performance between English and Sanskrit targets. To bridge this gap, we introduce Monolingual Adapter Networks: a parameter-efficient method to bolster cross-lingual alignment of embedding models without the need for parallel corpora or full finetuning.

Culturally grounded commonsense reasoning is underexplored in low-resource languages due to scarce data and costly native annotation. We test whether large language models (LLMs) can generate culturally nuanced narratives for such settings. Focusing on Javanese and Sundanese, we compare three data creation strategies: (1) LLM-assisted stories prompted with cultural cues, (2) machine translation from Indonesian benchmarks, and (3) native-written stories. Human evaluation finds LLM stories match natives on cultural fidelity but lag in coherence and correctness. We fine-tune models on each dataset and evaluate on a human-authored test set for classification and generation. LLM-generated data yields higher downstream performance than machine-translated and Indonesian human-authored training data. We release a high-quality benchmark of culturally grounded commonsense stories in Javanese and Sundanese to support future work.

With the release of new large language models (LLMs) like Llama and Mistral, zero-shot cross-lingual transfer has become increasingly feasible due to their multilingual pretraining and strong generalization capabilities. However, adapting these decoder-only LLMs to new tasks across languages remains challenging. While parameter-efficient fine-tuning (PeFT) techniques like Low-Rank Adaptation (LoRA) are widely used, prefix-based techniques such as soft prompt tuning, prefix tuning, and Llama Adapter are less explored, especially for zero-shot transfer in decoder-only models. We present a comprehensive study of three prefix-based methods for zero-shot cross-lingual transfer from English to 35+ high- and low-resource languages. Our analysis further explores transfer across linguistic families and scripts, as well as the impact of scaling model sizes from 1B to 24B. With Llama 3.1 8B, prefix methods outperform LoRA-baselines by up to 6% on the Belebele benchmark. Similar improvements were observed with Mistral v0.3 7B as well. Despite using only 1.23M learning parameters with prefix tuning, we achieve consistent improvements across diverse benchmarks. These findings highlight the potential of prefix-based techniques as an effective and scalable alternative to LoRA, particularly in low-resource multilingual settings.

Decoder-only large language models (LLMs) excel in high-resource languages across various tasks through few-shot or even zero-shot in-context learning (ICL). However, their performance often does not transfer well to low-resource languages, especially those written in non-Latin scripts. Inspired by recent work that leverages transliteration in encoder-only models, we investigate whether transliteration is also effective in improving LLMs’ performance for low-resource languages written in non-Latin scripts. To this end, we propose three prompt templates, where the target-language text is represented in (1) its original script, (2) Latin script transliteration, or (3) combined. We apply these methods to several representative LLMs of different sizes on various tasks including text classification and sequential labeling. Our findings show that the effectiveness of transliteration varies by task type and model size. For instance, all models benefit from transliterations for sequential labeling (with increases of up to 25%%).

This work investigates how a multilingual transformer model represents morphosyntactic properties of questions. We introduce the Question Type and Complexity (QTC) dataset with sentences across seven languages, annotated with type information and complexity metrics including dependency length, tree depth, and lexical density. Our evaluation extends probing methods to regression labels with selectivity controls to quantify gains in generalizability. We compare layer-wise probes on frozen Glot500-m (Imani et al., 2023) representations against subword TF-IDF baselines, and a fine-tuned model. Results show that statistical features classify questions well in explicitly marked languages and structural complexity prediction, while neural probes lead on individual metrics. We use these results to evaluate when contextual representations outperform statistical baselines and whether parameter updates reduce availability of pre-trained linguistic information.

We introduce Entropy2Vec, a novel framework for deriving cross-lingual language representations by leveraging the entropy of monolingual language models. Unlike traditional typological inventories that suffer from feature sparsity and static snapshots, Entropy2Vec uses the inherent uncertainty in language models to capture typological relationships between languages. By training a language model on a single language, we hypothesize that the entropy of its predictions reflects its structural similarity to other languages: Low entropy indicates high similarity, while high entropy suggests greater divergence. This approach yields dense, non-sparse language embeddings that are adaptable to different timeframes and free from missing values. Empirical evaluations demonstrate that Entropy2Vec embeddings align with established typological categories and achieved competitive performance in downstream multilingual NLP tasks, such as those addressed by the LinguAlchemy framework.

Large Language Models (LLMs) have demonstrated remarkable generalization capabilities across tasks and languages, revolutionizing natural language processing. This paper investigates the naturally emerging representation alignment in LLMs, particularly in the middle layers, and its implications for disentangling language-specific and language-agnostic information. We empirically confirm the existence of this alignment, analyze its behavior in comparison to explicitly designed alignment models, and demonstrate its potential for language-specific manipulation without semantic degradation. Building on these findings, we propose Inference-Time Language Control (ITLC), a novel method that leverages latent injection to enable precise cross-lingual language control and mitigate language confusion in LLMs. Our experiments highlight ITLC’s strong cross-lingual control capabilities while preserving semantic integrity in target languages. Furthermore, we demonstrate its effectiveness in alleviating the cross-lingual language confusion problem, which persists even in current large-scale LLMs, leading to inconsistent language generation. This work advances our understanding of representation alignment in LLMs and introduces a practical solution for enhancing their cross-lingual performance.

Large Language Models (LLMs) show strong potential for reranking documents in information retrieval (IR), but training with monolingual data often leads to monolingual overfitting and lexical bias, limiting generalization in cross-lingual IR (CLIR). To overcome these issues, we investigate instruction-tuning LLaMA-3.1-8B-Instruct on English and multilingual code-switched data, and evaluate on mMARCO and XQuAD-R. Results show that instruction-tuning on code-switched data substantially improves CLIR performance, while monolingual tuning remains more effective for monolingual reranking. We introduce a novel measure to analyze the relationship between lexical overlap and reranking performance, showing that the two factors are correlated. We finally conduct a causal analysis using counterfactual examples, where we evaluate whether rewriting passages that share overlapping keywords with the query causes models to change their relevance predictions. Overall, we find that code-switching serves as an effective and lightweight strategy to improve cross-lingual generalization in LLM-based re-ranking, while our analyses show that lexical overlap remains a major factor that can mislead reranking models.

Generic statements like “birds fly” or “lions have manes” express generalizations about kinds that allow exceptions, yet language models tend to overgeneralize them to universal claims. While previous work showed that ASCENT KB could reduce this effect in English by 30-40%, the effectiveness of broader knowledge sources and the cross-lingual nature of this phenomenon remain unexplored. We investigate generic overgeneralization across English and four South African languages (isiZulu, isiXhosa, Sepedi, SeSotho), comparing the impact of ConceptNet and DBpedia against the previously used ASCENT KB. Our experiments show that ConceptNet reduces overgeneralization by 45-52%% for minority characteristic generics, while DBpedia achieves 48-58%% for majority characteristics, with combined knowledge bases reaching 67%% reduction. These improvements are consistent across all languages, though Nguni languages show higher baseline overgeneralization than Sotho-Tswana languages, potentially suggesting that morphological features may influence this semantic bias. Our findings demonstrate that commonsense and encyclopedic knowledge provide complementary benefits for multilingual semantic understanding, offering insights for developing NLP systems that capture nuanced semantics in low-resource languages.

How similar are model outputs across languages? In this work, we study this question using a recently proposed model similarity metric—𝜅_p—applied to 20 languages and 47 subjects in GlobalMMLU. Our analysis reveals that a model’s responses become increasingly consistent across languages as its size and capability grow. Interestingly, models exhibit greater cross-lingual consistency within themselves than agreement with other models prompted in the same language. These results highlight not only the value of 𝜅_p as a practical tool for evaluating multilingual reliability, but also its potential to guide the development of more consistent multilingual systems.

Despite the effective performance of multilingual large language models (LLMs), the mechanisms underlying their multilingual capabilities remain unclear. This study examines the intermediate representations of multilingual LLMs to determine if these models utilize human-like second language acquisition strategies: coordinate, sub-coordinate, or compound learning. Our investigations into the discriminative and generative aspects of these models indicate that coordinate learning is the dominant mechanism, with decoder-only models progressively developing distinct feature spaces for each language, while encoder-only models exhibit a mixture of coordinate and compound learning in their middle layers. We find little evidence for sub-coordinate learning. Moreover, the role of training data coverage in shaping multilingual representations is reflected in the fact that languages present in a model’s training data consistently exhibit stronger separation than those absent from it.

This work investigates the effectiveness of small autoregressive language models (SLMs) with up to one billion parameters (sub-1B) for natural language processing (NLP) tasks in low-resource languages, focusing on Basque. We analyze optimal training strategies by comparing training from scratch and continual pre-training using state-of-the-art SLM architectures. Our analysis considers factors such as model size and the extent of Basque presence in the pre-training corpus. To assess linguistic capabilities, models are evaluated on 12 NLP tasks using the Harness framework. We also conduct a manual evaluation of fine-tuned models on three downstream natural language generation (NLG) tasks: question answering (QA), summarization, and machine translation (MT). Our findings indicate that continual pre-training on a multilingual SLM substantially enhances linguistic performance compared to training from scratch, particularly in low-resource language settings where available corpora typically contain fewer than one billion words. Additionally, the presence of Basque during the pre-training and larger model sizes contribute positively to performance in NLG tasks.

We introduce jina-embeddings-v4, a 3.8 billion parameter embedding model that unifies text and image representations, with a novel architecture supporting both single-vector and multi-vector embeddings. It achieves high performance on both single-modal and cross-modal retrieval tasks, and is particularly strong in processing visually rich content such as tables, charts, diagrams, and mixed-media formats that incorporate both image and textual information. We also introduce JVDR, a novel benchmark for visually rich document retrieval that includes more diverse materials and query types than previous efforts. We use JVDR to show that jina-embeddings-v4 greatly improves on state-of-the-art performance for these kinds of tasks.

In recent years, large language models (LLMs) have demonstrated significant potential across various natural language processing (NLP) tasks. However, their performance in domainspecific applications and non-English languages remains less explored. This study introduces a novel Romanian-language dataset 1 for multiple-choice biology questions, carefully curated to assess LLM comprehension and reasoning capabilities in scientific contexts. Containing approximately 14,000 questions, the dataset provides a comprehensive resource for evaluating and improving LLM performance in biology. We benchmark several popular LLMs, analyzing their accuracy, reasoning patterns, and ability to understand domain-specific terminology and linguistic nuances. Additionally, we perform comprehensive experiments to evaluate the impact of prompt engineering, fine-tuning, and other optimization techniques on model performance. Our findings highlight both the strengths and limitations of current LLMs in handling specialized knowledge tasks in lowresource languages, offering valuable insights for future research and development.

We introduce MMCRICBENCH-3K, a benchmark for Visual Question Answering (VQA) on cricket scorecards, designed to evaluate large vision-language models (LVLMs) on complex numerical and cross-lingual reasoning over semi-structured tabular images. MMCRICBENCH-3K comprises 1,463 synthetically generated scorecard images from ODI, T20, and Test formats, accompanied by 1,500 English QA pairs. It includes two subsets: MMCRICBENCH-E-1.5K, featuring English scorecards, and MMCRICBENCH-H1.5K, containing visually similar Hindi scorecards, with all questions and answers kept in English to enable controlled cross-script evaluation. The task demands reasoning over structured numerical data, multi-image context, and implicit domain knowledge. Empirical results show that even state-of-the-art LVLMs, such as GPT-4o and Qwen2.5VL, struggle on the English subset despite it being their primary training language and exhibit a further drop in performance on the Hindi subset. This reveals key limitations in structure-aware visual text understanding, numerical reasoning, and cross-lingual generalization. The dataset is publicly available via Hugging Face at https://huggingface.co/ datasets/DIALab/MMCricBench, to promote LVLM research in this direction.

Evaluating text generation capabilities of large language models (LLMs) is challenging, particularly for low-resource languages where methods for direct assessment are scarce. We propose MUG-Eval, a novel framework that evaluates LLMs’ multilingual generation capabilities by transforming existing benchmarks into conversational tasks and measuring the LLMs’ accuracies on those tasks. We specifically designed these conversational tasks to require effective communication in the target language. Then, we simply use task success rate as a proxy for successful conversation generation. Our approach offers two key advantages: it is independent of language-specific NLP tools or annotated datasets, which are limited for most languages, and it does not rely on LLMs-as-judges, whose evaluation quality degrades outside a few high-resource languages. We evaluate 8 LLMs across 30 languages spanning high, mid, and low-resource categories, and we find that MUG-Eval correlates strongly with established benchmarks (r > 0.75) while enabling standardized comparisons across languages and models. Our framework provides a robust and resource-efficient solution for evaluating multilingual generation that can be extended to thousands of languages.

Most languages lack sufficient data for largescale monolingual pretraining, creating a “data wall.” Multilingual pretraining helps but is limited by language imbalance and the “curse of multilinguality.” An alternative is to translate high-resource text with machine translation (MT), which raises three questions: (1) How does MT-derived data scale with model capacity? (2) Can source-side transformations (e.g., simplifying English with an LLM) improve generalization to native text? (3) How well do models pretrained on MT-derived data adapt when continually trained on limited native text? We investigate these questions by translating English into Indonesian and Tamil—two typologically distant, lowerresource languages—and pretraining GPT-2 models (124M–774M) on native or MT-derived corpora from raw and LLM-simplified English. We evaluate cross-entropy loss on native text, along with accuracy on syntactic probes and downstream tasks. Our results show that (1) MT-pretrained models benefit from scaling; (2) source-side simplification harms generalization to native text; and (3) adapting MT-pretrained models on native text often yields better performance than native-only models, even with less native data. However, tasks requiring cultural nuance (e.g., toxicity detection) demand more exposure to native data.

Despite substantial efforts, detecting and preventing hate speech online remains an understudied task for marginalized communities, particularly in the Global South, which includes developing societies with increasing internet penetration. In this paper, we aim to provide marginalized communities in societies where the dominant language is low-resource with a privacy-preserving tool to protect themselves from online hate speech by filtering offensive content in their native languages. Our contributions are twofold: 1) we release REACT (REsponsive hate speech datasets Across ConTexts), a collection of high-quality, culturespecific hate speech detection datasets comprising multiple target groups and low-resource languages, curated by experienced data collectors; 2) we propose a few-shot hate speech detection approach based on federated learning (FL), a privacy-preserving method for collaboratively training a central model that exhibits robustness when tackling different target groups and languages. By keeping training local to user devices, we ensure data privacy while leveraging the collective learning benefits of FL. We experiment with both multilingual and monolingual pre-trained representation spaces as backbones to examine the interaction between FL and different model representations. Furthermore, we explore personalized client models tailored to specific target groups and evaluate their performance. Our findings indicate the overall effectiveness of FL across different target groups, and point to personalization as a promising direction.

Multilingual retrieval-augmented generation (MRAG) systems heavily rely on robust Information Retrieval (IR). Reranking as a key component optimizes the initially retrieved document set to present the most pertinent information to the generative model, addressing context limitations and minimizing hallucinations. We propose an approach that trains Large Language Models (LLMs) as multilingual listwise rerankers through supervised fine-tuning (SFT) on a diverse mixture of multilingual and extended English ranking examples, and enhancing reasoning capabilities through Direct Preference Optimization (DPO) from translated task-specific reasoning processes. Experiments demonstrate that the approach improves accuracy@5 by 20-30% across all six high- mediumand low-resource languages compared to the BM25. The posted training 1B models achieve comparable performance to 7B baseline models while enabling faster inference. Finally, we investigate the effectiveness of different reasoning strategies in DPO with crosslingual and monolingual thinking processes.