Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state using any measurement system of their choice. Being available to users from diverse cultural backgrounds, Large Language Models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is truly the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs’ answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.

With Large Language Models (LLMs) becoming increasingly multilingual, effective knowledge editing (KE) needs to propagate edits across languages. Evaluation of the existing methods for cross-lingual knowledge editing (CKE) is limited both w.r.t. edit effectiveness: benchmarks do not account for entity aliases and use faulty entity translations; as well as robustness: existing work fails to report on downstream generation and task-solving abilities of LLMs after editing. In this work, we aim to (i) maximize the effectiveness of CKE while at the same time (ii) minimizing the extent of downstream model collapse due to the edits. To accurately measure the effectiveness of CKE methods, we introduce BabelEdits, a new CKE benchmark covering 60 languages that combines high-quality multilingual synsets from BabelNet with marker-based translation to ensure entity translation quality. Unlike existing CKE benchmarks, BabelEdits accounts for the rich variety of entity aliases within and across languages. We then propose BabelReFT, a modular CKE approach based on representation fine-tuning (ReFT) which learns entity-scope ReFT modules, applying them to all multilingual aliases at inference. Our experimental results show that not only is BabelReFT more effective in CKE than state-of-the-art methods, but, owing to its modular design, much more robust against downstream model collapse when subjected to many sequential edits.

Existing multilingual vision-language (VL) benchmarks often only cover a handful of languages. Consequently, evaluations of large vision-language models (LVLMs) predominantly target high-resource languages, underscoring the need for evaluation data for low-resource languages. To address this limitation, we introduce MVL-SIB, a massively multilingual vision-language benchmark that evaluates both cross-modal and text-only topical matching across 205 languages – over 100 more than the most multilingual existing VL benchmarks encompass. We then benchmark a range of of open-weight LVLMs together with GPT-4o(-mini) on MVL-SIB. Our results reveal that LVLMs struggle in cross-modal topic matching in lower-resource languages, performing no better than chance on languages like N’Koo. Our analysis further reveals that VL support in LVLMs declines disproportionately relative to textual support for lower-resource languages, as evidenced by comparison of cross-modal and text-only topical matching performance. We further observe that open-weight LVLMs do not benefit from representing a topic with more than one image, suggesting that these models are not yet fully effective at handling multi-image tasks. By correlating performance on MVL-SIB with other multilingual VL benchmarks, we highlight that MVL-SIB serves as a comprehensive probe of multilingual VL understanding in LVLMs.

LLMs have become a go-to solution not just for text generation, but also for natural language understanding (NLU) tasks. Acquiring extensive knowledge through language modeling on web-scale corpora, they excel on English NLU, yet struggle to extend their NLU capabilities to underrepresented languages. In contrast, machine translation models (MT) produce excellent multilingual representations, resulting in strong translation performance even for low-resource languages. MT encoders, however, lack the knowledge necessary for comprehensive NLU that LLMs obtain through language modeling training on immense corpora. In this work, we get the best both worlds by integrating MT encoders directly into LLM backbones via sample-efficient self-distillation. The resulting MT-LLMs preserve the inherent multilingual representational alignment from the MT encoder, allowing lower-resource languages to tap into the rich knowledge embedded in English-centric LLMs. Merging the MT encoder and LLM in a single model, we mitigate the propagation of translation errors and inference overhead of MT decoding inherent to discrete translation-based cross-lingual transfer (e.g., translate-test). Evaluation spanning three prominent NLU tasks and 127 predominantly low-resource languages renders MT-LLMs highly effective in cross-lingual transfer. MT-LLMs substantially and consistently outperform translation-test based on the same MT model, showing that we truly unlock multilingual language understanding for LLMs.

Massively multilingual language models have displayed strong performance in zero-shot (ZS-XLT) and few-shot (FS-XLT) cross-lingual transfer setups, where models fine-tuned on task data in a source language are transferred without any or with only a few annotated instances to the target language(s). However, current work typically overestimates model performance as fine-tuned models are frequently evaluated at model checkpoints that generalize best to validation instances in the target languages. This effectively violates the main assumptions of ‘true’ ZS-XLT and FS-XLT. Such XLT setups require robust methods that do not depend on labeled target language data for validation and model selection. In this work, aiming to improve the robustness of ‘true’ ZS-XLT and FS-XLT, we propose a simple and effective method that averages different checkpoints (i.e., model snapshots) during task fine-tuning. We conduct exhaustive ZS-XLT and FS-XLT experiments across higher-level semantic tasks (NLI, extractive QA) and lower-level token classification tasks (NER, POS). The results indicate that averaging model checkpoints yields systematic and consistent performance gains across diverse target languages in all tasks. Importantly, it simultaneously substantially desensitizes XLT to varying hyperparameter choices in the absence of target language validation. We also show that checkpoint averaging benefits performance when further combined with run averaging (i.e., averaging the parameters of models fine-tuned over independent runs).

Multilingual language models enable zero-shot cross-lingual transfer (ZS-XLT): fine-tuned on sizable source-language task data, they perform the task in target languages without labeled instances. The effectiveness of ZS-XLT hinges on the linguistic proximity between languages and the amount of pretraining data for a language. Because of this, model selection based on source-language validation is unreliable: it picks model snapshots with suboptimal target-language performance. As a remedy, some work optimizes ZS-XLT by extensively tuning hyperparameters: the follow-up work then routinely struggles to replicate the original results. Other work searches over narrower hyperparameter grids, reporting substantially lower performance. In this work, we therefore propose an unsupervised evaluation protocol for ZS-XLT that decouples performance maximization from hyperparameter tuning. As a robust and more transparent alternative to extensive hyperparameter tuning, we propose to accumulatively average snapshots from different runs into a single model. We run broad ZS-XLT experiments on both higher-level semantic tasks (NLI, extractive QA) and a lower-level token classification task (NER) and find that conventional model selection based on source-language validation quickly plateaus to suboptimal ZS-XLT performance. On the other hand, our accumulative run-by-run averaging of models trained with different hyperparameters boosts ZS-XLT performance and closely correlates with “oracle” ZS-XLT, i.e., model selection based on target-language validation performance.

A large body of recent work highlights the fallacies of zero-shot cross-lingual transfer (ZS-XLT) with large multilingual language models. Namely, their performance varies substantially for different target languages and is the weakest where needed the most: for low-resource languages distant to the source language. One remedy is few-shot transfer (FS-XLT), where leveraging only a few task-annotated instances in the target language(s) may yield sizable performance gains. However, FS-XLT also succumbs to large variation, as models easily overfit to the small datasets. In this work, we present a systematic study focused on a spectrum of FS-XLT fine-tuning regimes, analyzing key properties such as effectiveness, (in)stability, and modularity. We conduct extensive experiments on both higher-level (NLI, paraphrasing) and lower-level tasks (NER, POS), presenting new FS-XLT strategies that yield both improved and more stable FS-XLT across the board. Our findings challenge established FS-XLT methods: e.g., we propose to replace sequential fine-tuning with joint fine-tuning on source and target language instances, offering consistent gains with different number of shots (including resource-rich scenarios). We also show that further gains can be achieved with multi-stage FS-XLT training in which joint multilingual fine-tuning precedes the bilingual source-target specialization.

Large multilingual language models generally demonstrate impressive results in zero-shot cross-lingual transfer, yet often fail to successfully transfer to low-resource languages, even for token-level prediction tasks like named entity recognition (NER). In this work, we introduce a simple yet highly effective approach for improving zero-shot transfer for NER to low-resource languages. We observe that NER fine-tuning in the source language decontextualizes token representations, i.e., tokens increasingly attend to themselves. This increased reliance on token information itself, we hypothesize, triggers a type of overfitting to properties that NE tokens within the source languages share, but are generally not present in NE mentions of target languages. As a remedy, we propose a simple yet very effective sliced fine-tuning for NER (SLICER) that forces stronger token contextualization in the Transformer: we divide the transformed token representations and classifier into disjoint slices that are then independently classified during training. We evaluate SLICER on two standard benchmarks for NER that involve low-resource languages, WikiANN and MasakhaNER, and show that it (i) indeed reduces decontextualization (i.e., extent to which NE tokens attend to themselves), consequently (ii) yielding consistent transfer gains, especially prominent for low-resource target languages distant from the source language.

We introduce Seagle, a platform for comparative evaluation of semantic text encoding models on information retrieval (IR) tasks. Seagle implements (1) word embedding aggregators, which represent texts as algebraic aggregations of pretrained word embeddings and (2) pretrained semantic encoders, and allows for their comparative evaluation on arbitrary (monolingual and cross-lingual) IR collections. We benchmark Seagle’s models on monolingual document retrieval and cross-lingual sentence retrieval. Seagle functionality can be exploited via an easy-to-use web interface and its modular backend (micro-service architecture) can easily be extended with additional semantic search models.