Most hallucination evaluations focus on English, leaving it unclear whether findings transfer to lower-resource languages. We investigate faithfulness hallucinations, defined as model-generated content that is fluent and plausible but diverges from the provided input or is internally inconsistent. Leveraging the multilingual MultiWikiQA dataset, we utilize the LettuceDetect framework to create synthetic hallucination datasets for 21 European languages, which are then used to create token-level hallucination classifiers. In this work, we present evaluations of model hallucinations on a selection of languages: English, Danish, German, and Icelandic. Using these classifiers, we evaluate the hallucination rates for Qwen3-0.6B, Qwen3-14B, Gemma-3-12B-IT, cogito-v1-preview-qwen-32B, and cogito-v1-preview-llama-70B. Our classifiers reveal notably higher hallucination rates for Qwen3-0.6B (up to 60% of answers containing at least one hallucination, peaking in Icelandic) and generally lower rates for larger models, with cogito-v1-preview-qwen-32B and cogito-v1-preview-llama-70B performing best on most languages. Hallucination rates are consistently higher for lower-resource languages, particularly Icelandic.

We create high-quality datasets for LLM evaluation of logical reasoning skills across nine different languages, which have been manually checked by fluent speakers. The datasets consist of so-called zebra puzzles, and we analyse different ways of tuning the difficulty of the puzzles to fit modern LLMs. This includes the size of the puzzle (number of objects and number of clues), as well as a novel addition of red herring clues containing only irrelevant information. We show that presence of red herrings indeed makes the puzzles significantly harder for the models, and we find puzzle sizes 2×3 and 4×5 are sufficiently challenging for GPT-4o mini (a non-reasoning model) and o3-mini (a reasoning model), respectively. We analyse whether LLM performance of these are sensitive to the language, the cultural sensitivity of the puzzle theme, and the choice of clue types. These analyses are conducted with English and Danish, where we show that there is no significant difference for either of these three aspects, at least for the OpenAI models GPT-4o mini and o3-mini, chosen as representative non-reasoning and reasoning models, respectively. We publish the datasets for each of the nine languages for the identified sizes 2×3 and 4×5. We also publish the code used to generate the puzzles, which can be used to extend the benchmark into more languages.

We introduce a new reading comprehension dataset, dubbed MultiWikiQA, which covers 306 languages and has 1,220,757 samples in total. We start with Wikipedia articles, which also provide the context for the dataset samples, and use an LLM to generate question/answer pairs related to the Wikipedia article, ensuring that the answer appears verbatim within the article. Next, the question is then rephrased to hinder simple word matching methods from performing well on the dataset. We conduct a crowdsourced human evaluation of the fluency of the generated questions, which included 156 respondents across 30 of the languages (both low- and high-resource). All 30 languages received a mean fluency rating above “mostly natural”, showing that the samples are of good quality. We evaluate 6 different language models, both decoder and encoder models of varying sizes, showing that the benchmark is sufficiently difficult and that there is a large performance discrepancy amongst the languages. Both the dataset and survey evaluations are publicly available.

Translated versions of English-origin benchmarks such as MMLU are widely used to evaluate large language models in non-English languages, yet it remains unclear whether they measure the same constructs as natively created benchmarks. We frame this as a *proxy validity* question and compare DeepL-translated MMLU proxies against six native benchmarks in seven pairings across five Nordic and Baltic languages, evaluating 29 models on each pairing. Beyond ranking agreement, we test *incremental* proxy validity: does the translated score add predictive value for native performance after controlling for model size and language proficiency? Proxy validity varies substantially across benchmark pairings. Applying the same robustness criteria to all seven pairings, we find four robust effects, two marginal effects, and one null result. The strongest evidence occurs for curriculum-overlap benchmarks such as Skolprov (𝛥 R²=.128, 95% CI [.038,.253]), whereas translated MMLU provides no significant incremental value for Sweden-specific factual knowledge in SwedishFacts (p=.242). Because SwedishFacts is also the only open-ended benchmark, we cannot fully separate the effects of local content and response format. Overall, the results indicate a clear pattern: translated benchmarks provide a better proxy when their source content overlaps with the native evaluation target, while locally anchored knowledge is less reliably captured.