| Average EM | The first main metric we use to quantify performance for MultiLoKo is the average Exact Match score across languages, which expresses how many of the answers match one of the gold standard answers verbatim (after post-processing the answers). |
| Gap | The second main metric is the gap between a model's best and worst performing language. We gap to quantify the extent to which a model has achieved parity across languages. Because a small gap can be achieved both through parity on high scores as parity on low scores, it is most informative in combination with average benchmark performance. |
| Mother tongue effect (MTE) | MTE expresses the impact of asking questions in a language in which the requested information is locally salient, compared to asking it in English. A positive MTE indicates information is more readily available in the language it was (likely) present in the training data, whereas a negative mother tongue effect indicates the information is more easily accessible in English. |
| Locality effect (LE) | LE quantifies the effect of using locally sourced vs translated data. It is measured by computing the difference between scores for locally sourced data and translated English-sourced data. A positive LE implies that using translated English data underestimates performance on a language, a negative LE that using translated English data overestimates performance. |
",
+ "bbox": [
+ 176,
+ 88,
+ 823,
+ 296
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Table 1: MultiLoKo metric cheatsheet. We use several metrics to quantify model performance using MultiLoKo. This table provides a cheatsheet for their meaning.",
+ "bbox": [
+ 169,
+ 301,
+ 823,
+ 333
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "LE The locality effect (LE) is a measure of how much performance on knowledge tasks is over- or underestimated through the use of using translated English data, as opposed to locally relevant data. We quantify the locality effect as the difference in EM for English translated data and locally sourced data. If for a language the English translated data has as a higher EM, the LE is positive, indicating that using English translated data likely overestimating a model's ability on providing knowledge for that language. If the LE is negative the English translated data may provide an underestimation of the score for that language. Note that because we often observe both positive and negative LEs for the 30 non-English languages in MultiLoKo, the average LE across languages may be small, even if the differences for individual languages may be large.",
+ "bbox": [
+ 169,
+ 369,
+ 826,
+ 494
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4 Experimental setup",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 516,
+ 372,
+ 535
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We test and showcase our benchmark by running experiments with 11 different models of varying sizes, that were all marketed to have multilingual abilities.",
+ "bbox": [
+ 169,
+ 547,
+ 823,
+ 578
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1 Models",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 589,
+ 266,
+ 604
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "To test the extent to which MultiLoKo provides useful signal across training stages, we consider both base and chat models. The base models we include in our experiments are Llama 3.1 70B and 405B (Dubey et al., 2024), Mixtral 8x22B (team, 2024), and Qwen 2.5 72B (Qwen et al., 2025), the seven chat models are Gemini 2.0 Flash (Google DeepMind, 2024), GPT4-o (OpenAI et al., 2024), Claude 3.5 Sonnet (Anthropic, 2025), Llama 3.1 70B and 405B Chat, Mixtral 8x22B-it, and Qwen 2.5 72B instruct. As mentioned before, we run chat and base models with separate prompts.",
+ "bbox": [
+ 169,
+ 614,
+ 823,
+ 700
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.2 Experimental setup",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 718,
+ 351,
+ 734
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We run all of our experiments with the generation temperature set to 0. To facilitate automatic evaluation, we include an instruction to answer questions curtly and precisely, producing only a number/name/location/etc. Full template information can be found in our github repository.",
+ "bbox": [
+ 169,
+ 744,
+ 823,
+ 787
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Few-shot prompting For base models we use a 5-shot prompt. For chat models, we use a 0-shot prompt, as this is the most likely use mode by chat model users.",
+ "bbox": [
+ 169,
+ 792,
+ 823,
+ 821
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Post-processing Because base models are good at following the instructions, minimal postprocessing is needed: we only lowercase the output and strip punctuation. Chat models often deviate from the required format, especially in English, in various ways that we discuss in Appendix B. To evaluate such models beyond their instruction-following issues, we perform more complex post-processing, aiming to remove any words resembling \"answer\" from the LLM output, as well as several special cases for English and Japanese. We provide full details about post-processing in Appendix C.",
+ "bbox": [
+ 169,
+ 828,
+ 826,
+ 912
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "5",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/187a2566712d7e7b5269313bad6b1bd1cade176fdb265f30224cbd55ff0b8bdc.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": " | Description | Example |
| 1. | Extremely local and relatively obscure. Content that is of interest only to a small, localized group, such as a specific town, region, or community. These topics are typically obscure and not widely known beyond their immediate area. | Local radio stations, small town historical events, regional businesses, or niche local cultural practices. |
| 2. | Regional interest. Topics that have some relevance beyond a specific locality but are still primarily of interest within a particular region or country. | State or provincial politicians, regional cuisine, local sports teams, or medium-sized companies with regional influence. |
| 3. | National Significance. Content that is widely recognized within a single country, but relatively un-known internationally. | National politicians (not internationally known), popular national media figures, major corporations within a country, or significant national historical events. |
| 4. | International recognition. Topics that are recognized and have relevance in multiple countries but may not be universally known across the globe. These topics often have international influence and are likely to be covered in international media, though their impact may vary by region. | International brands which may be recognized in more than one country, celebrities with some international reach, significant cultural movements, or political conflicts with some awareness on the international stage. |
| 5. | Global prominence. Content that is widely recognized and relevant across a large number of countries around the world. These topics have a global impact or appeal and are likely to be well-represented in media across diverse cultures and regions. | Globally famous celebrities (e.g., Cristiano Ronaldo), multinational corporations (e.g., Apple), major world events, or universally recognized cultural icons. |
",
+ "bbox": [
+ 173,
+ 333,
+ 823,
+ 747
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "Table 3: Rubric for locality rating task. In the locality rating task, we ask the annotators to rate paragraphs with respect to how locally relevant the topic is to the locale.",
+ "bbox": [
+ 169,
+ 750,
+ 823,
+ 779
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "D.2 Question generation",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 816,
+ 357,
+ 830
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "The second and main annotation step in our pipeline is the step in which we ask annotators to generate questions about sampled paragraphs. We ask annotators to generate a challenging question with a short answer. The answer should be easy to evaluate with string-matching metrics, the questions should not be open-ended or have many possible correct answers, be ambiguous or subjective, and the expected short answer should be concise. To ensure difficulty, we ask that answering the question",
+ "bbox": [
+ 169,
+ 842,
+ 825,
+ 912
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "page_number",
+ "text": "24",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "requires combining information from different parts in the accompanying text; It should not be answerable by mere regurgitation of a single sentence. We furthermore ask that the question is formulated such that its answer will not change over time (e.g. not 'How many medals has Sifan Hassan won', but 'How many medals has Sifan Hassan won between 2018 and 2022 (including)'), and that the question is answerable also without the article (e.g. not 'How many tv shows did the person in this article produce?'). To facilitate validation checks in the next round, we also ask that the question authors write a longer answer to explain how they arrived at the short answer. We also ask the question authors to annotate what is the type of the correct answer (e.g. number, name, date, etc) In the pilot, we observed that – for some languages – the vast majority of questions were questions that required some form of numerical reasoning. Because the intention of the benchmark is to address knowledge more than reasoning, we afterwards restricted the number of numerical questions to $10\\%$ . Similarly, we asked question authors to avoid yes/no questions.",
+ "bbox": [
+ 169,
+ 90,
+ 826,
+ 257
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "D.3 Question review",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 273,
+ 330,
+ 287
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "In the first round of question review, we asked annotators from a different provider to judge whether the questions abide by the rules provided to the question authors. All question reviewers are native speakers. Specifically, we ask them to check if:",
+ "bbox": [
+ 169,
+ 297,
+ 823,
+ 340
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- The question pertains to a locally relevant topic",
+ "- The question is clear and understandable, and not subjective",
+ "- The question has a clear and concise answer",
+ "- If there are multiple possible variations of the answer possible (e.g. 'Dick Schoof' / 'Minister Dick Schoof' / 'Prime Minister Dick Schoof' / etc), all versions of the answer are provided.",
+ "- The question and answer are in the correct language",
+ "- The question is understandable without the article",
+ "- That the answer to the question will not likely change in the near future"
+ ],
+ "bbox": [
+ 215,
+ 351,
+ 823,
+ 489
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "When a question can be fixed with a minor change (e.g. add a time indication to make sure an answer will not change in the near future, or add an extra answer version), we ask the question reviewers to implement this fix and describe it. In the pilot round, we use the annotator feedback to finetune our annotation protocol and provide feedback to the question-authors. During the rest of the data collection, we simply disregard questions that are not useable as is or can be corrected with minor changes.",
+ "bbox": [
+ 169,
+ 501,
+ 823,
+ 584
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "D.4 Validation through question answering",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 599,
+ 486,
+ 616
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "In the last stage of our question generation pipeline, we have additional annotators answer the sourced and reviewed question. The goal of this validation task is to confirm that the questions are answerable, correct, non-ambiguous when read by individuals other than the original question author, and that all possible versions of the answers are included. For each question, we ask two additional annotators to first answer the question, using the snippets the questions were sourced from for context. After they have answered the question, they are shown the list of reference answers written by the original author of the question as well as the rational they provided, and we ask them to reflect upon the answer they gave themselves. If their answer did not match any answer in the original reference list, we ask them to either add their answer to the list if it is semantically equivalent to their own answer or indicate which answer they believe to be correct, their own or the original answer. We disregard all questions where at least one annotator disagrees with the original question author.",
+ "bbox": [
+ 169,
+ 626,
+ 826,
+ 779
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "E Related work",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 797,
+ 321,
+ 813
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "In this paper, we introduce a new multilingual benchmark for LLMs, that we believe addresses gaps and pitfalls in existing benchmarks. We (concisely) outlined those gaps and pitfalls and mentioned several other works related to ours in the introduction of those paper. Here, we discuss multilingual evaluation of LLMs in more detail. Specifically, we discuss what datasets recent LLM releases have used for multilingual evaluation (Appendix E.1) and what other datasets and approaches they could have used but did not (Appendix E.2).",
+ "bbox": [
+ 169,
+ 828,
+ 823,
+ 912
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "page_number",
+ "text": "25",
+ "bbox": [
+ 488,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "table",
+ "img_path": "images/6097bbba21f178dc4be15a6d96ad58e0b1c3c7cc17e299a76a31147c2d105aca.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Claude 3.5 Sonnet | MGSM (Shi et al., 2023) |
| Gemini 2.0 Flash | Mentions multilingual audio, no multilingual benchmarks scores reported. |
| GPT4-o | ARC-Easy and TruthfulQA translated into five African languages (internal benchmark), Uhura-Eval (internal benchmark). |
| Llama 3.1 | MGSM (Shi et al., 2023), Multilingual MMLU (internal benchmark) |
| Mixtral 8x22B | translated ARC-C, HellaSwag and MMLU (internal benchmarks) |
| Qwen2.5 72B | M3Exam (Zhang et al., 2023), IndoMMLU (Koto et al., 2023), ruMMLU (Fenogenova et al., 2024), translated MMLU (Chen et al., 2023), Belebele (Bandarkar et al., 2024), XCOPA (Ponti et al., 2020), XWinograd (Muennighoff et al., 2023), XStoryClose (Lin et al., 2022), PAWS-X (Zhang et al., 2019), MGSM (Shi et al., 2023), Flores-101 (Goyal et al., 2022) |
",
+ "bbox": [
+ 176,
+ 88,
+ 823,
+ 275
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "Table 4: Multilingual evaluation of recent LLM releases, overview. We provide an overview table of the benchmark for which scores are reported in the release papers or notes of the LLMs we evaluated in this paper. Models are sorted alphabetically.",
+ "bbox": [
+ 169,
+ 279,
+ 823,
+ 324
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "E.1 Multilingual evaluation of LLMs in practice",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 358,
+ 522,
+ 373
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "While multilinguality is something frequently mentioned in the release papers or posts of recent LLM releases, the datasets for which they actual report scores is in most cases quite limited. Of the models that we evaluated for this paper, Gemini 2.0 Flash reported no multilingual scores at all; GPT4-o and Mixtral 8x22B report scores only on internally translated but not publicly available English benchmarks; Claude 3.5 Sonnet reports scores for only one benchmark - MGSM. MGSM is also the only publicly available benchmark for which Llama 3.1 reports scores, along with - also - an internally translated version of MMLU that is not publicly available. The only model that extensively reports multilingual benchmark values, on more than 10 benchmarks, is Qwen2.5 72B. We provide an overview of the multilingual benchmarks for which scores are reported for these models in Table 4.",
+ "bbox": [
+ 169,
+ 383,
+ 823,
+ 508
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "E.2 Multilingual evaluation options for LLMs",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 523,
+ 504,
+ 539
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "While, as we discuss below, there are gaps and challenges with multilingual evaluation for LLMs, there are in fact many more options than is suggested by what is reported in recent releases. Below, we discuss other options for multilingual LLM evaluation.",
+ "bbox": [
+ 169,
+ 549,
+ 826,
+ 592
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "Translated English benchmarks As mentioned earlier on, benchmarks used for LLM evaluation are often translated English benchmarks. In some cases, the benchmarks were designed to evaluate only English and translated later, such as translated MMLU (e.g. Li et al., 2024; Chen et al., 2023; OpenAI, 2025; Singh et al., 2024) or MMLU-ProX (Xuan et al., 2025), MGSM (Shi et al., 2023) or MLAMA (Kassner et al., 2021). In other cases, the benchmark was multilingual at the time of its creation, but means of creation of the non-English data was through translating English sourced data, such as Belebele Bandarkar et al. (2024), Mintaka (Sen et al., 2022), or X-FACTR (Jiang et al., 2020). Taken together, translated benchmarks span quite a range of tasks, such as question answering (Artetxe et al., 2020; Lewis et al., 2020; Qi et al., 2023; Ohmer et al., 2023), natural language inference (Conneau et al., 2018), paraphrase detection (Zhang et al., 2019), general linguistic competence (Jumelet et al., 2025), reading comprehension (Artetxe et al., 2020; Bandarkar et al., 2024) and commonsense reasoning (Ponti et al., 2020), and even instruction following (He et al., 2024). With the exception of question answering and of course instruction following, however, many of these tasks have gone (somewhat) out of fashion for LLM evaluation, a trend which is mirrored also in the usage of their multilingual counterparts. As mentioned before, translated benchmarks have the advantage of containing parallel data, allowing for some form of comparability across languages, but are English-centric in content and may suffer from translationese (see e.g. Romanou et al., 2024; Chen et al., 2024, for a recent discussion of this).",
+ "bbox": [
+ 169,
+ 604,
+ 826,
+ 854
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "Multilingual benchmarks sourced from scratch Though much rarer, there are also benchmarks that are created independently for each language they include. Clark et al. (2020) release a question answering dataset separately sourced for 11 different languages, with a protocol relatively similar",
+ "bbox": [
+ 169,
+ 869,
+ 823,
+ 912
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "page_number",
+ "text": "26",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "text",
+ "text": "to ours. In a different category, Hardalov et al. (2020), Zhang et al. (2023) and Romanou et al. (2024) and Sánchez et al. (2024) do not create benchmark data, but instead collect existing exam or competition questions from official human exams. In case of Zhang et al. (2023), the exams are graduation exams of primary, middle and high school; Hardalov et al. (2020) includes official state exams taken by graduating high school students, which may contain parallel pairs in case countries allow examinations to be taken in multiple languages; Romanou et al. (2024), cover academic exams at middle and high school and university level, professional certifications and licenses, and exams to obtain regional licenses. Sánchez et al. (2024) instead focus on questions from the International Linguistic Olympiad corpus. Lastly, as part of their study Ohmer et al. (2023) create a dataset called SIMPLE FACTS, containing factual questions created through a shared template filled in with language specific factual data.",
+ "bbox": [
+ 169,
+ 90,
+ 826,
+ 243
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "text",
+ "text": "Consistency evaluation A rather different approach to assess multilinguality in LLMs is to focus not on accuracy across different languages, but to consider whether predictions are consistent across languages. This tests knowledge and skill transfer between languages more explicitly. Two recent examples of studies incorporating consistency-based evaluations on factual knowledge questions are Qi et al. (2023) and Ohmer et al. (2023). Qi et al. (2023) focuseses specifically on sample-level consistency of answers across different languages, requiring existing parallel benchmarks. Ohmer et al. (2023), instead, ask models to translate benchmark questions themselves before answering them again. This can, with some caveats, be applied to any existing monolingual benchmark, but – requiring multiple steps – it is more involved as a paradigm, and is somewhat bottlenecked by the translation ability of the model to be evaluated.",
+ "bbox": [
+ 169,
+ 258,
+ 826,
+ 397
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "text",
+ "text": "Translation as a proxy for multilinguality Another, more implicit method to assess multilinguality in LLMs is to evaluate their ability to translate from one language to another. This approach was famously used by Brown et al. (2020), but has not been common since.",
+ "bbox": [
+ 169,
+ 411,
+ 823,
+ 454
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "text",
+ "text": "Monolingual non-English evaluation In our discussion, we have focussed on multilingual evaluation options that cover multiple other languages. After all, a benchmark to evaluate models on Bengali (e.g. Shafayat et al., 2024) or Arabic (e.g. Alwajih et al., 2024) can contribute to multilingual evaluation when combined with other benchmarks, but does not so on its own. Because such benchmarks are usually created by language experts for the respective languages, they usually target locally relevant skills and knowledge and are likely of higher quality than benchmarks created for many languages simultaneously (either through translation or from scratch). Yet, composing a suite including many languages that allows direct comparisons between languages remains challenging. We believe such benchmarks can be important for multilingual evaluation in LLMs, but will not further discuss benchmarks focussing on individual languages or very small sets of languages within one family here.",
+ "bbox": [
+ 169,
+ 468,
+ 826,
+ 621
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "page_number",
+ "text": "27",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 26
+ }
+]
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10356/15a8c18d-57d9-46bd-bbe1-f5ca7eeeb023_model.json b/data/2025/2504_10xxx/2504.10356/15a8c18d-57d9-46bd-bbe1-f5ca7eeeb023_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f1683abc24e9573d6fd2423d601b160bb68ff3e
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/15a8c18d-57d9-46bd-bbe1-f5ca7eeeb023_model.json
@@ -0,0 +1,3444 @@
+[
+ [
+ {
+ "type": "aside_text",
+ "bbox": [
+ 0.023,
+ 0.265,
+ 0.061,
+ 0.708
+ ],
+ "angle": 270,
+ "content": "arXiv:2504.10356v2 [cs.CL] 15 Apr 2025"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.228,
+ 0.123,
+ 0.771,
+ 0.175
+ ],
+ "angle": 0,
+ "content": "MultiLoKo: a multilingual local knowledge benchmark for LLMs spanning 31 languages"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.359,
+ 0.226,
+ 0.638,
+ 0.241
+ ],
+ "angle": 0,
+ "content": "Dieuwke Hupkes* Nikolay Bogoychev*"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.482,
+ 0.242,
+ 0.518,
+ 0.254
+ ],
+ "angle": 0,
+ "content": "Meta"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.365,
+ 0.255,
+ 0.635,
+ 0.269
+ ],
+ "angle": 0,
+ "content": "{dieuwkehupkes,nbogoych}@meta.com"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.46,
+ 0.305,
+ 0.538,
+ 0.32
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.23,
+ 0.335,
+ 0.768,
+ 0.64
+ ],
+ "angle": 0,
+ "content": "We present MultiLoKo, a new benchmark for evaluating multilinguality in LLMs covering 31 languages. MultiLoKo consists of three partitions: a main partition consisting of 500 questions per language, separately sourced to be locally relevant to the specific language, and two translated partitions, containing human-authored translations from 30 non-English languages to English and vice versa. For comparison, we also release corresponding machine-authored translations. The data is equally distributed over two splits: a dev split and a blind, out-of-distribution test split. MultiLoKo can be used to study a variety of questions regarding the multilinguality of LLMs as well as meta-questions about multilingual benchmark creation. We compute MultiLoKo scores for 11 base and chat models marketed to be multilingual and study their average performance, their performance parity across languages, how much their ability to answer questions depends on the question language, and which languages are most difficult. None of the models we studied performs well on MultiLoKo, as indicated by low average scores as well as large differences between the best and worst scoring languages. Furthermore, we find a substantial effect of the question language, indicating suboptimal knowledge transfer between languages. Lastly, we find that using local vs English-translated data can result in differences more than 20 points for the best performing models, drastically change the estimated difficulty of some languages. For using machines instead of human translations, we find a weaker effect on ordering of language difficulty, a larger difference in model rankings, and a substantial drop in estimated performance for all models."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.173,
+ 0.663,
+ 0.312,
+ 0.679
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.693,
+ 0.825,
+ 0.806
+ ],
+ "angle": 0,
+ "content": "With the growing presence and deployment of LLMs across the world, evaluating their abilities in languages other than English becomes more and more eminent. Yet, studying and evaluating multilinguality in LLMs remains a challenging enterprise, and it is hardly exaggerated to call the current state of multilingual evaluation in LLMs insufficient. Older multilingual benchmarks such as PAWS-X (Zhang et al., 2019), XNLI (Conneau et al., 2018) or XCOPA (Ponti et al., 2020) often do not fit the demands for evaluating auto-regressive models and are rarely used to evaluate recent models. Furthermore, their coverage of languages is relatively small compared to the number of languages in which LLMs are intended to be proficient."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.807,
+ 0.827,
+ 0.892
+ ],
+ "angle": 0,
+ "content": "More often used for LLM evaluation are benchmarks translated from English, such as MGSM (translated GSM8K, Shi et al., 2023), MMMLU (tranlated MMLU, OpenAI, 2025) or (less frequently) Belebele (Bandarkar et al., 2024). These benchmarks provide good coverage over many languages, but using translated data comes with its own set of issues. One such issue is that even when human-rather than machine-authored translations are used, translated data is known to differ from native text in several ways (Clark et al., 2020). Furthermore, using translated benchmarks imposes a strong"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.192,
+ 0.899,
+ 0.319,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "*Equal contributions"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.092,
+ 0.825,
+ 0.162
+ ],
+ "angle": 0,
+ "content": "English-centric bias: translated data may be multilingual on the surface, it is not in its content. The benchmarks MLQA (Lewis et al., 2020) and TidyQA (Clark et al., 2020) to some extent address the issue by sourcing data separately for different languages. Even in their sourcing protocols, however, there is no explicit focus on selecting locally relevant content for the chosen languages. In addition to that, their coverage is again small compared to the above mentioned translated benchmarks."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.164,
+ 0.827,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "In response to these issues, we introduce a wide-coverage multilingual benchmark with locally-sourced questions for 31 different languages. Because the benchmark targets multilingual local knowledge, we dub it MultiLoKo. The release of MultiLoKo serves two interconnected goals:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.209,
+ 0.209,
+ 0.625,
+ 0.222
+ ],
+ "angle": 0,
+ "content": "1) Provide a better means to evaluate multilinguality in LLMs;"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.208,
+ 0.225,
+ 0.782,
+ 0.239
+ ],
+ "angle": 0,
+ "content": "2) Provide data to study the effect of various design choices in multilingual evaluation."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.208,
+ 0.209,
+ 0.782,
+ 0.239
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.242,
+ 0.827,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "To address our first goal, we create 500 questions per language, written from scratch for each language, using a sourcing protocol specifically designed to ensure local relevance of the question topics. To also reap the benefits of parallel data, we commissioned both human and machine-authored translations for all non-English questions into English and vice versa, providing a total of 15500 parallel questions, sourced across the 31 languages in the benchmark. The translated data facilitates the study of transfer between languages and also serves our second goal. By comparing the English-translated data with the locally sourced data, we can explicitly compare the adequacy of using translated benchmarks; by comparing human- with machine-authored translations, we can better estimate the potential issues of the latter. To prevent quick overfitting and inadvertent contamination, we release a development set of the benchmark, while test scores can only be obtained through an external provider.| Average EM | The first main metric we use to quantify performance for MultiLoKo is the average Exact Match score across languages, which expresses how many of the answers match one of the gold standard answers verbatim (after post-processing the answers). |
| Gap | The second main metric is the gap between a model's best and worst performing language. We gap to quantify the extent to which a model has achieved parity across languages. Because a small gap can be achieved both through parity on high scores as parity on low scores, it is most informative in combination with average benchmark performance. |
| Mother tongue effect (MTE) | MTE expresses the impact of asking questions in a language in which the requested information is locally salient, compared to asking it in English. A positive MTE indicates information is more readily available in the language it was (likely) present in the training data, whereas a negative mother tongue effect indicates the information is more easily accessible in English. |
| Locality effect (LE) | LE quantifies the effect of using locally sourced vs translated data. It is measured by computing the difference between scores for locally sourced data and translated English-sourced data. A positive LE implies that using translated English data underestimates performance on a language, a negative LE that using translated English data overestimates performance. |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.171,
+ 0.303,
+ 0.825,
+ 0.334
+ ],
+ "angle": 0,
+ "content": "Table 1: MultiLoKo metric cheatsheet. We use several metrics to quantify model performance using MultiLoKo. This table provides a cheatsheet for their meaning."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.37,
+ 0.828,
+ 0.496
+ ],
+ "angle": 0,
+ "content": "LE The locality effect (LE) is a measure of how much performance on knowledge tasks is over- or underestimated through the use of using translated English data, as opposed to locally relevant data. We quantify the locality effect as the difference in EM for English translated data and locally sourced data. If for a language the English translated data has as a higher EM, the LE is positive, indicating that using English translated data likely overestimating a model's ability on providing knowledge for that language. If the LE is negative the English translated data may provide an underestimation of the score for that language. Note that because we often observe both positive and negative LEs for the 30 non-English languages in MultiLoKo, the average LE across languages may be small, even if the differences for individual languages may be large."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.517,
+ 0.373,
+ 0.536
+ ],
+ "angle": 0,
+ "content": "4 Experimental setup"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.549,
+ 0.825,
+ 0.579
+ ],
+ "angle": 0,
+ "content": "We test and showcase our benchmark by running experiments with 11 different models of varying sizes, that were all marketed to have multilingual abilities."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.59,
+ 0.267,
+ 0.605
+ ],
+ "angle": 0,
+ "content": "4.1 Models"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.616,
+ 0.825,
+ 0.701
+ ],
+ "angle": 0,
+ "content": "To test the extent to which MultiLoKo provides useful signal across training stages, we consider both base and chat models. The base models we include in our experiments are Llama 3.1 70B and 405B (Dubey et al., 2024), Mixtral 8x22B (team, 2024), and Qwen 2.5 72B (Qwen et al., 2025), the seven chat models are Gemini 2.0 Flash (Google DeepMind, 2024), GPT4-o (OpenAI et al., 2024), Claude 3.5 Sonnet (Anthropic, 2025), Llama 3.1 70B and 405B Chat, Mixtral 8x22B-it, and Qwen 2.5 72B instruct. As mentioned before, we run chat and base models with separate prompts."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.719,
+ 0.352,
+ 0.735
+ ],
+ "angle": 0,
+ "content": "4.2 Experimental setup"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.745,
+ 0.825,
+ 0.789
+ ],
+ "angle": 0,
+ "content": "We run all of our experiments with the generation temperature set to 0. To facilitate automatic evaluation, we include an instruction to answer questions curtly and precisely, producing only a number/name/location/etc. Full template information can be found in our github repository."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.794,
+ 0.825,
+ 0.823
+ ],
+ "angle": 0,
+ "content": "Few-shot prompting For base models we use a 5-shot prompt. For chat models, we use a 0-shot prompt, as this is the most likely use mode by chat model users."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.829,
+ 0.827,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "Post-processing Because base models are good at following the instructions, minimal postprocessing is needed: we only lowercase the output and strip punctuation. Chat models often deviate from the required format, especially in English, in various ways that we discuss in Appendix B. To evaluate such models beyond their instruction-following issues, we perform more complex post-processing, aiming to remove any words resembling \"answer\" from the LLM output, as well as several special cases for English and Japanese. We provide full details about post-processing in Appendix C."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.505,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "5"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.195,
+ 0.089,
+ 0.795,
+ 0.272
+ ],
+ "angle": 0,
+ "content": " | Description | Example |
| 1. | Extremely local and relatively obscure. Content that is of interest only to a small, localized group, such as a specific town, region, or community. These topics are typically obscure and not widely known beyond their immediate area. | Local radio stations, small town historical events, regional businesses, or niche local cultural practices. |
| 2. | Regional interest. Topics that have some relevance beyond a specific locality but are still primarily of interest within a particular region or country. | State or provincial politicians, regional cuisine, local sports teams, or medium-sized companies with regional influence. |
| 3. | National Significance. Content that is widely recognized within a single country, but relatively un-known internationally. | National politicians (not internationally known), popular national media figures, major corporations within a country, or significant national historical events. |
| 4. | International recognition. Topics that are recognized and have relevance in multiple countries but may not be universally known across the globe. These topics often have international influence and are likely to be covered in international media, though their impact may vary by region. | International brands which may be recognized in more than one country, celebrities with some international reach, significant cultural movements, or political conflicts with some awareness on the international stage. |
| 5. | Global prominence. Content that is widely recognized and relevant across a large number of countries around the world. These topics have a global impact or appeal and are likely to be well-represented in media across diverse cultures and regions. | Globally famous celebrities (e.g., Cristiano Ronaldo), multinational corporations (e.g., Apple), major world events, or universally recognized cultural icons. |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.171,
+ 0.75,
+ 0.825,
+ 0.78
+ ],
+ "angle": 0,
+ "content": "Table 3: Rubric for locality rating task. In the locality rating task, we ask the annotators to rate paragraphs with respect to how locally relevant the topic is to the locale."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.817,
+ 0.358,
+ 0.832
+ ],
+ "angle": 0,
+ "content": "D.2 Question generation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.843,
+ 0.826,
+ 0.914
+ ],
+ "angle": 0,
+ "content": "The second and main annotation step in our pipeline is the step in which we ask annotators to generate questions about sampled paragraphs. We ask annotators to generate a challenging question with a short answer. The answer should be easy to evaluate with string-matching metrics, the questions should not be open-ended or have many possible correct answers, be ambiguous or subjective, and the expected short answer should be concise. To ensure difficulty, we ask that answering the question"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.936,
+ 0.509,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "24"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.092,
+ 0.827,
+ 0.258
+ ],
+ "angle": 0,
+ "content": "requires combining information from different parts in the accompanying text; It should not be answerable by mere regurgitation of a single sentence. We furthermore ask that the question is formulated such that its answer will not change over time (e.g. not 'How many medals has Sifan Hassan won', but 'How many medals has Sifan Hassan won between 2018 and 2022 (including)'), and that the question is answerable also without the article (e.g. not 'How many tv shows did the person in this article produce?'). To facilitate validation checks in the next round, we also ask that the question authors write a longer answer to explain how they arrived at the short answer. We also ask the question authors to annotate what is the type of the correct answer (e.g. number, name, date, etc) In the pilot, we observed that – for some languages – the vast majority of questions were questions that required some form of numerical reasoning. Because the intention of the benchmark is to address knowledge more than reasoning, we afterwards restricted the number of numerical questions to \\(10\\%\\). Similarly, we asked question authors to avoid yes/no questions."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.274,
+ 0.331,
+ 0.289
+ ],
+ "angle": 0,
+ "content": "D.3 Question review"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.299,
+ 0.825,
+ 0.342
+ ],
+ "angle": 0,
+ "content": "In the first round of question review, we asked annotators from a different provider to judge whether the questions abide by the rules provided to the question authors. All question reviewers are native speakers. Specifically, we ask them to check if:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.352,
+ 0.546,
+ 0.366
+ ],
+ "angle": 0,
+ "content": "- The question pertains to a locally relevant topic"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.37,
+ 0.621,
+ 0.385
+ ],
+ "angle": 0,
+ "content": "- The question is clear and understandable, and not subjective"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.389,
+ 0.525,
+ 0.403
+ ],
+ "angle": 0,
+ "content": "- The question has a clear and concise answer"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.407,
+ 0.825,
+ 0.435
+ ],
+ "angle": 0,
+ "content": "- If there are multiple possible variations of the answer possible (e.g. 'Dick Schoof' / 'Minister Dick Schoof' / 'Prime Minister Dick Schoof' / etc), all versions of the answer are provided."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.44,
+ 0.575,
+ 0.454
+ ],
+ "angle": 0,
+ "content": "- The question and answer are in the correct language"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.458,
+ 0.561,
+ 0.472
+ ],
+ "angle": 0,
+ "content": "- The question is understandable without the article"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.476,
+ 0.701,
+ 0.491
+ ],
+ "angle": 0,
+ "content": "- That the answer to the question will not likely change in the near future"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.217,
+ 0.352,
+ 0.825,
+ 0.491
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.502,
+ 0.825,
+ 0.585
+ ],
+ "angle": 0,
+ "content": "When a question can be fixed with a minor change (e.g. add a time indication to make sure an answer will not change in the near future, or add an extra answer version), we ask the question reviewers to implement this fix and describe it. In the pilot round, we use the annotator feedback to finetune our annotation protocol and provide feedback to the question-authors. During the rest of the data collection, we simply disregard questions that are not useable as is or can be corrected with minor changes."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.601,
+ 0.488,
+ 0.617
+ ],
+ "angle": 0,
+ "content": "D.4 Validation through question answering"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.627,
+ 0.827,
+ 0.78
+ ],
+ "angle": 0,
+ "content": "In the last stage of our question generation pipeline, we have additional annotators answer the sourced and reviewed question. The goal of this validation task is to confirm that the questions are answerable, correct, non-ambiguous when read by individuals other than the original question author, and that all possible versions of the answers are included. For each question, we ask two additional annotators to first answer the question, using the snippets the questions were sourced from for context. After they have answered the question, they are shown the list of reference answers written by the original author of the question as well as the rational they provided, and we ask them to reflect upon the answer they gave themselves. If their answer did not match any answer in the original reference list, we ask them to either add their answer to the list if it is semantically equivalent to their own answer or indicate which answer they believe to be correct, their own or the original answer. We disregard all questions where at least one annotator disagrees with the original question author."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.798,
+ 0.322,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "E Related work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.829,
+ 0.825,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "In this paper, we introduce a new multilingual benchmark for LLMs, that we believe addresses gaps and pitfalls in existing benchmarks. We (concisely) outlined those gaps and pitfalls and mentioned several other works related to ours in the introduction of those paper. Here, we discuss multilingual evaluation of LLMs in more detail. Specifically, we discuss what datasets recent LLM releases have used for multilingual evaluation (Appendix E.1) and what other datasets and approaches they could have used but did not (Appendix E.2)."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.936,
+ 0.508,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "25"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.177,
+ 0.089,
+ 0.825,
+ 0.276
+ ],
+ "angle": 0,
+ "content": "| Claude 3.5 Sonnet | MGSM (Shi et al., 2023) |
| Gemini 2.0 Flash | Mentions multilingual audio, no multilingual benchmarks scores reported. |
| GPT4-o | ARC-Easy and TruthfulQA translated into five African languages (internal benchmark), Uhura-Eval (internal benchmark). |
| Llama 3.1 | MGSM (Shi et al., 2023), Multilingual MMLU (internal benchmark) |
| Mixtral 8x22B | translated ARC-C, HellaSwag and MMLU (internal benchmarks) |
| Qwen2.5 72B | M3Exam (Zhang et al., 2023), IndoMMLU (Koto et al., 2023), ruMMLU (Fenogenova et al., 2024), translated MMLU (Chen et al., 2023), Belebele (Bandarkar et al., 2024), XCOPA (Ponti et al., 2020), XWinograd (Muennighoff et al., 2023), XStoryClose (Lin et al., 2022), PAWS-X (Zhang et al., 2019), MGSM (Shi et al., 2023), Flores-101 (Goyal et al., 2022) |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.171,
+ 0.28,
+ 0.825,
+ 0.325
+ ],
+ "angle": 0,
+ "content": "Table 4: Multilingual evaluation of recent LLM releases, overview. We provide an overview table of the benchmark for which scores are reported in the release papers or notes of the LLMs we evaluated in this paper. Models are sorted alphabetically."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.359,
+ 0.523,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "E.1 Multilingual evaluation of LLMs in practice"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.384,
+ 0.825,
+ 0.509
+ ],
+ "angle": 0,
+ "content": "While multilinguality is something frequently mentioned in the release papers or posts of recent LLM releases, the datasets for which they actual report scores is in most cases quite limited. Of the models that we evaluated for this paper, Gemini 2.0 Flash reported no multilingual scores at all; GPT4-o and Mixtral 8x22B report scores only on internally translated but not publicly available English benchmarks; Claude 3.5 Sonnet reports scores for only one benchmark - MGSM. MGSM is also the only publicly available benchmark for which Llama 3.1 reports scores, along with - also - an internally translated version of MMLU that is not publicly available. The only model that extensively reports multilingual benchmark values, on more than 10 benchmarks, is Qwen2.5 72B. We provide an overview of the multilingual benchmarks for which scores are reported for these models in Table 4."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.524,
+ 0.506,
+ 0.54
+ ],
+ "angle": 0,
+ "content": "E.2 Multilingual evaluation options for LLMs"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.55,
+ 0.827,
+ 0.593
+ ],
+ "angle": 0,
+ "content": "While, as we discuss below, there are gaps and challenges with multilingual evaluation for LLMs, there are in fact many more options than is suggested by what is reported in recent releases. Below, we discuss other options for multilingual LLM evaluation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.606,
+ 0.827,
+ 0.856
+ ],
+ "angle": 0,
+ "content": "Translated English benchmarks As mentioned earlier on, benchmarks used for LLM evaluation are often translated English benchmarks. In some cases, the benchmarks were designed to evaluate only English and translated later, such as translated MMLU (e.g. Li et al., 2024; Chen et al., 2023; OpenAI, 2025; Singh et al., 2024) or MMLU-ProX (Xuan et al., 2025), MGSM (Shi et al., 2023) or MLAMA (Kassner et al., 2021). In other cases, the benchmark was multilingual at the time of its creation, but means of creation of the non-English data was through translating English sourced data, such as Belebele Bandarkar et al. (2024), Mintaka (Sen et al., 2022), or X-FACTR (Jiang et al., 2020). Taken together, translated benchmarks span quite a range of tasks, such as question answering (Artetxe et al., 2020; Lewis et al., 2020; Qi et al., 2023; Ohmer et al., 2023), natural language inference (Conneau et al., 2018), paraphrase detection (Zhang et al., 2019), general linguistic competence (Jumelet et al., 2025), reading comprehension (Artetxe et al., 2020; Bandarkar et al., 2024) and commonsense reasoning (Ponti et al., 2020), and even instruction following (He et al., 2024). With the exception of question answering and of course instruction following, however, many of these tasks have gone (somewhat) out of fashion for LLM evaluation, a trend which is mirrored also in the usage of their multilingual counterparts. As mentioned before, translated benchmarks have the advantage of containing parallel data, allowing for some form of comparability across languages, but are English-centric in content and may suffer from translationese (see e.g. Romanou et al., 2024; Chen et al., 2024, for a recent discussion of this)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.87,
+ 0.825,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "Multilingual benchmarks sourced from scratch Though much rarer, there are also benchmarks that are created independently for each language they include. Clark et al. (2020) release a question answering dataset separately sourced for 11 different languages, with a protocol relatively similar"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.936,
+ 0.509,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "26"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.092,
+ 0.827,
+ 0.244
+ ],
+ "angle": 0,
+ "content": "to ours. In a different category, Hardalov et al. (2020), Zhang et al. (2023) and Romanou et al. (2024) and Sánchez et al. (2024) do not create benchmark data, but instead collect existing exam or competition questions from official human exams. In case of Zhang et al. (2023), the exams are graduation exams of primary, middle and high school; Hardalov et al. (2020) includes official state exams taken by graduating high school students, which may contain parallel pairs in case countries allow examinations to be taken in multiple languages; Romanou et al. (2024), cover academic exams at middle and high school and university level, professional certifications and licenses, and exams to obtain regional licenses. Sánchez et al. (2024) instead focus on questions from the International Linguistic Olympiad corpus. Lastly, as part of their study Ohmer et al. (2023) create a dataset called SIMPLE FACTS, containing factual questions created through a shared template filled in with language specific factual data."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.259,
+ 0.828,
+ 0.398
+ ],
+ "angle": 0,
+ "content": "Consistency evaluation A rather different approach to assess multilinguality in LLMs is to focus not on accuracy across different languages, but to consider whether predictions are consistent across languages. This tests knowledge and skill transfer between languages more explicitly. Two recent examples of studies incorporating consistency-based evaluations on factual knowledge questions are Qi et al. (2023) and Ohmer et al. (2023). Qi et al. (2023) focuseses specifically on sample-level consistency of answers across different languages, requiring existing parallel benchmarks. Ohmer et al. (2023), instead, ask models to translate benchmark questions themselves before answering them again. This can, with some caveats, be applied to any existing monolingual benchmark, but – requiring multiple steps – it is more involved as a paradigm, and is somewhat bottlenecked by the translation ability of the model to be evaluated."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.412,
+ 0.825,
+ 0.455
+ ],
+ "angle": 0,
+ "content": "Translation as a proxy for multilinguality Another, more implicit method to assess multilinguality in LLMs is to evaluate their ability to translate from one language to another. This approach was famously used by Brown et al. (2020), but has not been common since."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.17,
+ 0.469,
+ 0.828,
+ 0.622
+ ],
+ "angle": 0,
+ "content": "Monolingual non-English evaluation In our discussion, we have focussed on multilingual evaluation options that cover multiple other languages. After all, a benchmark to evaluate models on Bengali (e.g. Shafayat et al., 2024) or Arabic (e.g. Alwajih et al., 2024) can contribute to multilingual evaluation when combined with other benchmarks, but does not so on its own. Because such benchmarks are usually created by language experts for the respective languages, they usually target locally relevant skills and knowledge and are likely of higher quality than benchmarks created for many languages simultaneously (either through translation or from scratch). Yet, composing a suite including many languages that allows direct comparisons between languages remains challenging. We believe such benchmarks can be important for multilingual evaluation in LLMs, but will not further discuss benchmarks focussing on individual languages or very small sets of languages within one family here."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.936,
+ 0.509,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "27"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10356/15a8c18d-57d9-46bd-bbe1-f5ca7eeeb023_origin.pdf b/data/2025/2504_10xxx/2504.10356/15a8c18d-57d9-46bd-bbe1-f5ca7eeeb023_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0c185764119b73a9a50504f981750bf53e29ead8
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/15a8c18d-57d9-46bd-bbe1-f5ca7eeeb023_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c52fc7232e34f1a190815c01170c1446895dacc5eeecb0e633287149a1cadafc
+size 570490
diff --git a/data/2025/2504_10xxx/2504.10356/full.md b/data/2025/2504_10xxx/2504.10356/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..18fb2d1a4278c260e6f8d2a5084aa18c0bf4b547
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/full.md
@@ -0,0 +1,444 @@
+# MultiLoKo: a multilingual local knowledge benchmark for LLMs spanning 31 languages
+
+Dieuwke Hupkes* Nikolay Bogoychev*
+
+Meta
+
+{dieuwkehupkes,nbogoych}@meta.com
+
+# Abstract
+
+We present MultiLoKo, a new benchmark for evaluating multilinguality in LLMs covering 31 languages. MultiLoKo consists of three partitions: a main partition consisting of 500 questions per language, separately sourced to be locally relevant to the specific language, and two translated partitions, containing human-authored translations from 30 non-English languages to English and vice versa. For comparison, we also release corresponding machine-authored translations. The data is equally distributed over two splits: a dev split and a blind, out-of-distribution test split. MultiLoKo can be used to study a variety of questions regarding the multilinguality of LLMs as well as meta-questions about multilingual benchmark creation. We compute MultiLoKo scores for 11 base and chat models marketed to be multilingual and study their average performance, their performance parity across languages, how much their ability to answer questions depends on the question language, and which languages are most difficult. None of the models we studied performs well on MultiLoKo, as indicated by low average scores as well as large differences between the best and worst scoring languages. Furthermore, we find a substantial effect of the question language, indicating suboptimal knowledge transfer between languages. Lastly, we find that using local vs English-translated data can result in differences more than 20 points for the best performing models, drastically change the estimated difficulty of some languages. For using machines instead of human translations, we find a weaker effect on ordering of language difficulty, a larger difference in model rankings, and a substantial drop in estimated performance for all models.
+
+# 1 Introduction
+
+With the growing presence and deployment of LLMs across the world, evaluating their abilities in languages other than English becomes more and more eminent. Yet, studying and evaluating multilinguality in LLMs remains a challenging enterprise, and it is hardly exaggerated to call the current state of multilingual evaluation in LLMs insufficient. Older multilingual benchmarks such as PAWS-X (Zhang et al., 2019), XNLI (Conneau et al., 2018) or XCOPA (Ponti et al., 2020) often do not fit the demands for evaluating auto-regressive models and are rarely used to evaluate recent models. Furthermore, their coverage of languages is relatively small compared to the number of languages in which LLMs are intended to be proficient.
+
+More often used for LLM evaluation are benchmarks translated from English, such as MGSM (translated GSM8K, Shi et al., 2023), MMMLU (tranlated MMLU, OpenAI, 2025) or (less frequently) Belebele (Bandarkar et al., 2024). These benchmarks provide good coverage over many languages, but using translated data comes with its own set of issues. One such issue is that even when human-rather than machine-authored translations are used, translated data is known to differ from native text in several ways (Clark et al., 2020). Furthermore, using translated benchmarks imposes a strong
+
+English-centric bias: translated data may be multilingual on the surface, it is not in its content. The benchmarks MLQA (Lewis et al., 2020) and TidyQA (Clark et al., 2020) to some extent address the issue by sourcing data separately for different languages. Even in their sourcing protocols, however, there is no explicit focus on selecting locally relevant content for the chosen languages. In addition to that, their coverage is again small compared to the above mentioned translated benchmarks.
+
+In response to these issues, we introduce a wide-coverage multilingual benchmark with locally-sourced questions for 31 different languages. Because the benchmark targets multilingual local knowledge, we dub it MultiLoKo. The release of MultiLoKo serves two interconnected goals:
+
+1) Provide a better means to evaluate multilinguality in LLMs;
+2) Provide data to study the effect of various design choices in multilingual evaluation.
+
+To address our first goal, we create 500 questions per language, written from scratch for each language, using a sourcing protocol specifically designed to ensure local relevance of the question topics. To also reap the benefits of parallel data, we commissioned both human and machine-authored translations for all non-English questions into English and vice versa, providing a total of 15500 parallel questions, sourced across the 31 languages in the benchmark. The translated data facilitates the study of transfer between languages and also serves our second goal. By comparing the English-translated data with the locally sourced data, we can explicitly compare the adequacy of using translated benchmarks; by comparing human- with machine-authored translations, we can better estimate the potential issues of the latter. To prevent quick overfitting and inadvertent contamination, we release a development set of the benchmark, while test scores can only be obtained through an external provider.| Average EM | The first main metric we use to quantify performance for MultiLoKo is the average Exact Match score across languages, which expresses how many of the answers match one of the gold standard answers verbatim (after post-processing the answers). |
| Gap | The second main metric is the gap between a model's best and worst performing language. We gap to quantify the extent to which a model has achieved parity across languages. Because a small gap can be achieved both through parity on high scores as parity on low scores, it is most informative in combination with average benchmark performance. |
| Mother tongue effect (MTE) | MTE expresses the impact of asking questions in a language in which the requested information is locally salient, compared to asking it in English. A positive MTE indicates information is more readily available in the language it was (likely) present in the training data, whereas a negative mother tongue effect indicates the information is more easily accessible in English. |
| Locality effect (LE) | LE quantifies the effect of using locally sourced vs translated data. It is measured by computing the difference between scores for locally sourced data and translated English-sourced data. A positive LE implies that using translated English data underestimates performance on a language, a negative LE that using translated English data overestimates performance. |
+
+Table 1: MultiLoKo metric cheatsheet. We use several metrics to quantify model performance using MultiLoKo. This table provides a cheatsheet for their meaning.
+
+LE The locality effect (LE) is a measure of how much performance on knowledge tasks is over- or underestimated through the use of using translated English data, as opposed to locally relevant data. We quantify the locality effect as the difference in EM for English translated data and locally sourced data. If for a language the English translated data has as a higher EM, the LE is positive, indicating that using English translated data likely overestimating a model's ability on providing knowledge for that language. If the LE is negative the English translated data may provide an underestimation of the score for that language. Note that because we often observe both positive and negative LEs for the 30 non-English languages in MultiLoKo, the average LE across languages may be small, even if the differences for individual languages may be large.
+
+# 4 Experimental setup
+
+We test and showcase our benchmark by running experiments with 11 different models of varying sizes, that were all marketed to have multilingual abilities.
+
+# 4.1 Models
+
+To test the extent to which MultiLoKo provides useful signal across training stages, we consider both base and chat models. The base models we include in our experiments are Llama 3.1 70B and 405B (Dubey et al., 2024), Mixtral 8x22B (team, 2024), and Qwen 2.5 72B (Qwen et al., 2025), the seven chat models are Gemini 2.0 Flash (Google DeepMind, 2024), GPT4-o (OpenAI et al., 2024), Claude 3.5 Sonnet (Anthropic, 2025), Llama 3.1 70B and 405B Chat, Mixtral 8x22B-it, and Qwen 2.5 72B instruct. As mentioned before, we run chat and base models with separate prompts.
+
+# 4.2 Experimental setup
+
+We run all of our experiments with the generation temperature set to 0. To facilitate automatic evaluation, we include an instruction to answer questions curtly and precisely, producing only a number/name/location/etc. Full template information can be found in our github repository.
+
+Few-shot prompting For base models we use a 5-shot prompt. For chat models, we use a 0-shot prompt, as this is the most likely use mode by chat model users.
+
+Post-processing Because base models are good at following the instructions, minimal postprocessing is needed: we only lowercase the output and strip punctuation. Chat models often deviate from the required format, especially in English, in various ways that we discuss in Appendix B. To evaluate such models beyond their instruction-following issues, we perform more complex post-processing, aiming to remove any words resembling "answer" from the LLM output, as well as several special cases for English and Japanese. We provide full details about post-processing in Appendix C.
+
+ | Description | Example |
| 1. | Extremely local and relatively obscure. Content that is of interest only to a small, localized group, such as a specific town, region, or community. These topics are typically obscure and not widely known beyond their immediate area. | Local radio stations, small town historical events, regional businesses, or niche local cultural practices. |
| 2. | Regional interest. Topics that have some relevance beyond a specific locality but are still primarily of interest within a particular region or country. | State or provincial politicians, regional cuisine, local sports teams, or medium-sized companies with regional influence. |
| 3. | National Significance. Content that is widely recognized within a single country, but relatively un-known internationally. | National politicians (not internationally known), popular national media figures, major corporations within a country, or significant national historical events. |
| 4. | International recognition. Topics that are recognized and have relevance in multiple countries but may not be universally known across the globe. These topics often have international influence and are likely to be covered in international media, though their impact may vary by region. | International brands which may be recognized in more than one country, celebrities with some international reach, significant cultural movements, or political conflicts with some awareness on the international stage. |
| 5. | Global prominence. Content that is widely recognized and relevant across a large number of countries around the world. These topics have a global impact or appeal and are likely to be well-represented in media across diverse cultures and regions. | Globally famous celebrities (e.g., Cristiano Ronaldo), multinational corporations (e.g., Apple), major world events, or universally recognized cultural icons. |
+
+Table 3: Rubric for locality rating task. In the locality rating task, we ask the annotators to rate paragraphs with respect to how locally relevant the topic is to the locale.
+
+# D.2 Question generation
+
+The second and main annotation step in our pipeline is the step in which we ask annotators to generate questions about sampled paragraphs. We ask annotators to generate a challenging question with a short answer. The answer should be easy to evaluate with string-matching metrics, the questions should not be open-ended or have many possible correct answers, be ambiguous or subjective, and the expected short answer should be concise. To ensure difficulty, we ask that answering the question
+
+requires combining information from different parts in the accompanying text; It should not be answerable by mere regurgitation of a single sentence. We furthermore ask that the question is formulated such that its answer will not change over time (e.g. not 'How many medals has Sifan Hassan won', but 'How many medals has Sifan Hassan won between 2018 and 2022 (including)'), and that the question is answerable also without the article (e.g. not 'How many tv shows did the person in this article produce?'). To facilitate validation checks in the next round, we also ask that the question authors write a longer answer to explain how they arrived at the short answer. We also ask the question authors to annotate what is the type of the correct answer (e.g. number, name, date, etc) In the pilot, we observed that – for some languages – the vast majority of questions were questions that required some form of numerical reasoning. Because the intention of the benchmark is to address knowledge more than reasoning, we afterwards restricted the number of numerical questions to $10\%$ . Similarly, we asked question authors to avoid yes/no questions.
+
+# D.3 Question review
+
+In the first round of question review, we asked annotators from a different provider to judge whether the questions abide by the rules provided to the question authors. All question reviewers are native speakers. Specifically, we ask them to check if:
+
+- The question pertains to a locally relevant topic
+- The question is clear and understandable, and not subjective
+- The question has a clear and concise answer
+- If there are multiple possible variations of the answer possible (e.g. 'Dick Schoof' / 'Minister Dick Schoof' / 'Prime Minister Dick Schoof' / etc), all versions of the answer are provided.
+- The question and answer are in the correct language
+- The question is understandable without the article
+- That the answer to the question will not likely change in the near future
+
+When a question can be fixed with a minor change (e.g. add a time indication to make sure an answer will not change in the near future, or add an extra answer version), we ask the question reviewers to implement this fix and describe it. In the pilot round, we use the annotator feedback to finetune our annotation protocol and provide feedback to the question-authors. During the rest of the data collection, we simply disregard questions that are not useable as is or can be corrected with minor changes.
+
+# D.4 Validation through question answering
+
+In the last stage of our question generation pipeline, we have additional annotators answer the sourced and reviewed question. The goal of this validation task is to confirm that the questions are answerable, correct, non-ambiguous when read by individuals other than the original question author, and that all possible versions of the answers are included. For each question, we ask two additional annotators to first answer the question, using the snippets the questions were sourced from for context. After they have answered the question, they are shown the list of reference answers written by the original author of the question as well as the rational they provided, and we ask them to reflect upon the answer they gave themselves. If their answer did not match any answer in the original reference list, we ask them to either add their answer to the list if it is semantically equivalent to their own answer or indicate which answer they believe to be correct, their own or the original answer. We disregard all questions where at least one annotator disagrees with the original question author.
+
+# E Related work
+
+In this paper, we introduce a new multilingual benchmark for LLMs, that we believe addresses gaps and pitfalls in existing benchmarks. We (concisely) outlined those gaps and pitfalls and mentioned several other works related to ours in the introduction of those paper. Here, we discuss multilingual evaluation of LLMs in more detail. Specifically, we discuss what datasets recent LLM releases have used for multilingual evaluation (Appendix E.1) and what other datasets and approaches they could have used but did not (Appendix E.2).
+
+| Claude 3.5 Sonnet | MGSM (Shi et al., 2023) |
| Gemini 2.0 Flash | Mentions multilingual audio, no multilingual benchmarks scores reported. |
| GPT4-o | ARC-Easy and TruthfulQA translated into five African languages (internal benchmark), Uhura-Eval (internal benchmark). |
| Llama 3.1 | MGSM (Shi et al., 2023), Multilingual MMLU (internal benchmark) |
| Mixtral 8x22B | translated ARC-C, HellaSwag and MMLU (internal benchmarks) |
| Qwen2.5 72B | M3Exam (Zhang et al., 2023), IndoMMLU (Koto et al., 2023), ruMMLU (Fenogenova et al., 2024), translated MMLU (Chen et al., 2023), Belebele (Bandarkar et al., 2024), XCOPA (Ponti et al., 2020), XWinograd (Muennighoff et al., 2023), XStoryClose (Lin et al., 2022), PAWS-X (Zhang et al., 2019), MGSM (Shi et al., 2023), Flores-101 (Goyal et al., 2022) |
+
+Table 4: Multilingual evaluation of recent LLM releases, overview. We provide an overview table of the benchmark for which scores are reported in the release papers or notes of the LLMs we evaluated in this paper. Models are sorted alphabetically.
+
+# E.1 Multilingual evaluation of LLMs in practice
+
+While multilinguality is something frequently mentioned in the release papers or posts of recent LLM releases, the datasets for which they actual report scores is in most cases quite limited. Of the models that we evaluated for this paper, Gemini 2.0 Flash reported no multilingual scores at all; GPT4-o and Mixtral 8x22B report scores only on internally translated but not publicly available English benchmarks; Claude 3.5 Sonnet reports scores for only one benchmark - MGSM. MGSM is also the only publicly available benchmark for which Llama 3.1 reports scores, along with - also - an internally translated version of MMLU that is not publicly available. The only model that extensively reports multilingual benchmark values, on more than 10 benchmarks, is Qwen2.5 72B. We provide an overview of the multilingual benchmarks for which scores are reported for these models in Table 4.
+
+# E.2 Multilingual evaluation options for LLMs
+
+While, as we discuss below, there are gaps and challenges with multilingual evaluation for LLMs, there are in fact many more options than is suggested by what is reported in recent releases. Below, we discuss other options for multilingual LLM evaluation.
+
+Translated English benchmarks As mentioned earlier on, benchmarks used for LLM evaluation are often translated English benchmarks. In some cases, the benchmarks were designed to evaluate only English and translated later, such as translated MMLU (e.g. Li et al., 2024; Chen et al., 2023; OpenAI, 2025; Singh et al., 2024) or MMLU-ProX (Xuan et al., 2025), MGSM (Shi et al., 2023) or MLAMA (Kassner et al., 2021). In other cases, the benchmark was multilingual at the time of its creation, but means of creation of the non-English data was through translating English sourced data, such as Belebele Bandarkar et al. (2024), Mintaka (Sen et al., 2022), or X-FACTR (Jiang et al., 2020). Taken together, translated benchmarks span quite a range of tasks, such as question answering (Artetxe et al., 2020; Lewis et al., 2020; Qi et al., 2023; Ohmer et al., 2023), natural language inference (Conneau et al., 2018), paraphrase detection (Zhang et al., 2019), general linguistic competence (Jumelet et al., 2025), reading comprehension (Artetxe et al., 2020; Bandarkar et al., 2024) and commonsense reasoning (Ponti et al., 2020), and even instruction following (He et al., 2024). With the exception of question answering and of course instruction following, however, many of these tasks have gone (somewhat) out of fashion for LLM evaluation, a trend which is mirrored also in the usage of their multilingual counterparts. As mentioned before, translated benchmarks have the advantage of containing parallel data, allowing for some form of comparability across languages, but are English-centric in content and may suffer from translationese (see e.g. Romanou et al., 2024; Chen et al., 2024, for a recent discussion of this).
+
+Multilingual benchmarks sourced from scratch Though much rarer, there are also benchmarks that are created independently for each language they include. Clark et al. (2020) release a question answering dataset separately sourced for 11 different languages, with a protocol relatively similar
+
+to ours. In a different category, Hardalov et al. (2020), Zhang et al. (2023) and Romanou et al. (2024) and Sánchez et al. (2024) do not create benchmark data, but instead collect existing exam or competition questions from official human exams. In case of Zhang et al. (2023), the exams are graduation exams of primary, middle and high school; Hardalov et al. (2020) includes official state exams taken by graduating high school students, which may contain parallel pairs in case countries allow examinations to be taken in multiple languages; Romanou et al. (2024), cover academic exams at middle and high school and university level, professional certifications and licenses, and exams to obtain regional licenses. Sánchez et al. (2024) instead focus on questions from the International Linguistic Olympiad corpus. Lastly, as part of their study Ohmer et al. (2023) create a dataset called SIMPLE FACTS, containing factual questions created through a shared template filled in with language specific factual data.
+
+Consistency evaluation A rather different approach to assess multilinguality in LLMs is to focus not on accuracy across different languages, but to consider whether predictions are consistent across languages. This tests knowledge and skill transfer between languages more explicitly. Two recent examples of studies incorporating consistency-based evaluations on factual knowledge questions are Qi et al. (2023) and Ohmer et al. (2023). Qi et al. (2023) focuseses specifically on sample-level consistency of answers across different languages, requiring existing parallel benchmarks. Ohmer et al. (2023), instead, ask models to translate benchmark questions themselves before answering them again. This can, with some caveats, be applied to any existing monolingual benchmark, but – requiring multiple steps – it is more involved as a paradigm, and is somewhat bottlenecked by the translation ability of the model to be evaluated.
+
+Translation as a proxy for multilinguality Another, more implicit method to assess multilinguality in LLMs is to evaluate their ability to translate from one language to another. This approach was famously used by Brown et al. (2020), but has not been common since.
+
+Monolingual non-English evaluation In our discussion, we have focussed on multilingual evaluation options that cover multiple other languages. After all, a benchmark to evaluate models on Bengali (e.g. Shafayat et al., 2024) or Arabic (e.g. Alwajih et al., 2024) can contribute to multilingual evaluation when combined with other benchmarks, but does not so on its own. Because such benchmarks are usually created by language experts for the respective languages, they usually target locally relevant skills and knowledge and are likely of higher quality than benchmarks created for many languages simultaneously (either through translation or from scratch). Yet, composing a suite including many languages that allows direct comparisons between languages remains challenging. We believe such benchmarks can be important for multilingual evaluation in LLMs, but will not further discuss benchmarks focussing on individual languages or very small sets of languages within one family here.
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10356/images/02b5749be8a0b0cad9681d40427e51c88c0bc0e81cda587783d142cb140bf240.jpg b/data/2025/2504_10xxx/2504.10356/images/02b5749be8a0b0cad9681d40427e51c88c0bc0e81cda587783d142cb140bf240.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..64b86ded115538dc988afc46a7fb7210a9cb0a69
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/02b5749be8a0b0cad9681d40427e51c88c0bc0e81cda587783d142cb140bf240.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee0de4a6ca9eba3dd096870f9635617cf0668d1cf5a3542eb7381b5ce186596a
+size 17028
diff --git a/data/2025/2504_10xxx/2504.10356/images/07d531440daa7d359f135908c66b8f86aff2111a8ae67e0aa29fd27c7f2f0eeb.jpg b/data/2025/2504_10xxx/2504.10356/images/07d531440daa7d359f135908c66b8f86aff2111a8ae67e0aa29fd27c7f2f0eeb.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a96c6de64dde78704850c44ed903a12f58232606
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/07d531440daa7d359f135908c66b8f86aff2111a8ae67e0aa29fd27c7f2f0eeb.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfbe00e56140143d9df2fa0896c9d0f5ac9ff0000234aa7a7752d7b84dadaad2
+size 25184
diff --git a/data/2025/2504_10xxx/2504.10356/images/0ee44851e6330b28bb51edddb359d3b46ffd35f998c59f2c4c70256478702034.jpg b/data/2025/2504_10xxx/2504.10356/images/0ee44851e6330b28bb51edddb359d3b46ffd35f998c59f2c4c70256478702034.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c0d374cb19265603723834ad02de935d5898b9cb
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/0ee44851e6330b28bb51edddb359d3b46ffd35f998c59f2c4c70256478702034.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d15e9ad57e97ea67daf22f7016475c98892493131f81238898c2879c03a9a76e
+size 68868
diff --git a/data/2025/2504_10xxx/2504.10356/images/187a2566712d7e7b5269313bad6b1bd1cade176fdb265f30224cbd55ff0b8bdc.jpg b/data/2025/2504_10xxx/2504.10356/images/187a2566712d7e7b5269313bad6b1bd1cade176fdb265f30224cbd55ff0b8bdc.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a2aa7d45bb6094ebf8160973bcfc8f2ee1909012
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/187a2566712d7e7b5269313bad6b1bd1cade176fdb265f30224cbd55ff0b8bdc.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf63c54de6950fac9e3364f897ff7c2aca562a9fe7aaa019a8ae3620667d7ed3
+size 93610
diff --git a/data/2025/2504_10xxx/2504.10356/images/4292a85211fc2019ae5f9eb0d70959896e6e3c18041a561b5072bafe1218ace8.jpg b/data/2025/2504_10xxx/2504.10356/images/4292a85211fc2019ae5f9eb0d70959896e6e3c18041a561b5072bafe1218ace8.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..254d038c6d1d72e9c39190f7c71b217e870f2bf6
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/4292a85211fc2019ae5f9eb0d70959896e6e3c18041a561b5072bafe1218ace8.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e55363b2a925455debdc7d447bbac620058df4c81905cc31f77e5363773f1555
+size 43734
diff --git a/data/2025/2504_10xxx/2504.10356/images/58b5ac41956d6eb877cd3c5fb7488a56cf7367f079130be27c34bff2cd0898d3.jpg b/data/2025/2504_10xxx/2504.10356/images/58b5ac41956d6eb877cd3c5fb7488a56cf7367f079130be27c34bff2cd0898d3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..21ea2fbdaafeea63a5c57a9a996dd2d91a4970b7
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/58b5ac41956d6eb877cd3c5fb7488a56cf7367f079130be27c34bff2cd0898d3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0931adb259fb00740495d7914fb69c94507d7d2f6ef76cf4deb7d4178feebbe
+size 27876
diff --git a/data/2025/2504_10xxx/2504.10356/images/5a773333e4ab35574de8f34ba1e9ec4e4d7fb66080d19020a6de0b49c9e62b74.jpg b/data/2025/2504_10xxx/2504.10356/images/5a773333e4ab35574de8f34ba1e9ec4e4d7fb66080d19020a6de0b49c9e62b74.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f735c8da46ed41fc8a2bd3a3e1944e132e8a86df
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/5a773333e4ab35574de8f34ba1e9ec4e4d7fb66080d19020a6de0b49c9e62b74.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5fc400a1b247916ebd0ab2fb650dea4c6ce7d88ab8a2c1821b4611dd38ae7c2
+size 97009
diff --git a/data/2025/2504_10xxx/2504.10356/images/5d86d006a27cdf31406eb72a1b910b6a9d13150c13b3ee9064c1db294d2d41f2.jpg b/data/2025/2504_10xxx/2504.10356/images/5d86d006a27cdf31406eb72a1b910b6a9d13150c13b3ee9064c1db294d2d41f2.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..dfbf8751ce059c649501ecd68992d5f75faed90f
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/5d86d006a27cdf31406eb72a1b910b6a9d13150c13b3ee9064c1db294d2d41f2.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e92949dc4c58037c8be7aafff70cb4db132c738538b2936b366b2a5f898d6ac8
+size 20417
diff --git a/data/2025/2504_10xxx/2504.10356/images/5e9488ee42bcffe9751990decbf76ca92024b4aaa74c5ac9b05c6f80f75e0642.jpg b/data/2025/2504_10xxx/2504.10356/images/5e9488ee42bcffe9751990decbf76ca92024b4aaa74c5ac9b05c6f80f75e0642.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ff16c1d3dcf918bbbe087a9aeb45d66bd1362f68
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/5e9488ee42bcffe9751990decbf76ca92024b4aaa74c5ac9b05c6f80f75e0642.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1de452b73f377b255d6fc3e880519b324579c39b4227603da5ea8974f27fe79b
+size 38108
diff --git a/data/2025/2504_10xxx/2504.10356/images/6097bbba21f178dc4be15a6d96ad58e0b1c3c7cc17e299a76a31147c2d105aca.jpg b/data/2025/2504_10xxx/2504.10356/images/6097bbba21f178dc4be15a6d96ad58e0b1c3c7cc17e299a76a31147c2d105aca.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e9eb842173a09d145598da75a9650f84ad81f385
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/6097bbba21f178dc4be15a6d96ad58e0b1c3c7cc17e299a76a31147c2d105aca.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07661fe59807570b4cb70b79dcba9faaa448c861baf4e73fb033571d3533763b
+size 108233
diff --git a/data/2025/2504_10xxx/2504.10356/images/8d54153ccb122aaea2874537a950dad799895f33acc86ca682873fd95442f949.jpg b/data/2025/2504_10xxx/2504.10356/images/8d54153ccb122aaea2874537a950dad799895f33acc86ca682873fd95442f949.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e75d1fdd818a5da2fcc42c9e09d821be1691e970
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/8d54153ccb122aaea2874537a950dad799895f33acc86ca682873fd95442f949.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d03352abd589d3b99f421214de0e49a30c3ea93baec4127b99d5db923bbf8034
+size 136700
diff --git a/data/2025/2504_10xxx/2504.10356/images/b0918ad445a1cd7b34ff620e97c14ae4d7e0eab4778d0e71993ce777ae50542b.jpg b/data/2025/2504_10xxx/2504.10356/images/b0918ad445a1cd7b34ff620e97c14ae4d7e0eab4778d0e71993ce777ae50542b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ab8dd528c6598791467cef8939fd1d36731f3f81
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/b0918ad445a1cd7b34ff620e97c14ae4d7e0eab4778d0e71993ce777ae50542b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f147aeb00912166bbafd228fb039bd90fc3cc12ac141279f14cd1caa2a1551f
+size 29465
diff --git a/data/2025/2504_10xxx/2504.10356/images/b801df6ef4552fe7a4acd00f70e7b0057dc62f87938a607d60d51710da9f7fbd.jpg b/data/2025/2504_10xxx/2504.10356/images/b801df6ef4552fe7a4acd00f70e7b0057dc62f87938a607d60d51710da9f7fbd.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3a29c2feebbd8610e1d2eb2d058b4938d50b0c13
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/b801df6ef4552fe7a4acd00f70e7b0057dc62f87938a607d60d51710da9f7fbd.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4b0a9cb74b81063ab76ab659961be317add5ff279877aa0c13876a90f96c573
+size 60673
diff --git a/data/2025/2504_10xxx/2504.10356/images/ba1c71502e3d1b62d80b80df10464eeef839823bfd41082afe99302cc0c1fa68.jpg b/data/2025/2504_10xxx/2504.10356/images/ba1c71502e3d1b62d80b80df10464eeef839823bfd41082afe99302cc0c1fa68.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..46a9aef631aee0fece4423c8349c50c558ada19d
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/ba1c71502e3d1b62d80b80df10464eeef839823bfd41082afe99302cc0c1fa68.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:666d1da09545991e270d3e3ff60eb6522b198d0888302d20e0a5c23e406cf399
+size 29225
diff --git a/data/2025/2504_10xxx/2504.10356/images/bd414948b1fc64eac369d890930e1fd6fdc455be1d18c2df3bdd0cd54887fd29.jpg b/data/2025/2504_10xxx/2504.10356/images/bd414948b1fc64eac369d890930e1fd6fdc455be1d18c2df3bdd0cd54887fd29.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..8387c297a88eeceb3439afb11543b1ca083feb1d
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/bd414948b1fc64eac369d890930e1fd6fdc455be1d18c2df3bdd0cd54887fd29.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf5da394aef8c242e2ce849b2134a0b9af694922be4795b9fa19d5e3e08f5e0b
+size 83910
diff --git a/data/2025/2504_10xxx/2504.10356/images/cc7d7f7d7a0ea73f1bd888448d909a38821f0cdcc30c4807be351475caeba83c.jpg b/data/2025/2504_10xxx/2504.10356/images/cc7d7f7d7a0ea73f1bd888448d909a38821f0cdcc30c4807be351475caeba83c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fe563773ca41df3019a7c12d702d7a4868b634c4
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/cc7d7f7d7a0ea73f1bd888448d909a38821f0cdcc30c4807be351475caeba83c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daf02bd96a665a8fc471063972b8ea560e5eeacc2d5b395f4331919ce50cd753
+size 34509
diff --git a/data/2025/2504_10xxx/2504.10356/images/e5f1a6301f9eacb88d277334073c9d3bc06a6bceaf152447d2047710024fbe06.jpg b/data/2025/2504_10xxx/2504.10356/images/e5f1a6301f9eacb88d277334073c9d3bc06a6bceaf152447d2047710024fbe06.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..7748989afa14afdae7b89f0f8f3534424b426fd1
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/e5f1a6301f9eacb88d277334073c9d3bc06a6bceaf152447d2047710024fbe06.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dda71e0b598701aae1ff58219a8d96bf51bac172d43a1562a4e8fce74a51bf2
+size 33291
diff --git a/data/2025/2504_10xxx/2504.10356/images/fa2dfae75eef0a7bb838f0dfc25bd66e8d7acffe982633b0c359ac0cb43632f3.jpg b/data/2025/2504_10xxx/2504.10356/images/fa2dfae75eef0a7bb838f0dfc25bd66e8d7acffe982633b0c359ac0cb43632f3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2f1e07716f531abe0e97581a5a9f218f9fb25dc7
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/images/fa2dfae75eef0a7bb838f0dfc25bd66e8d7acffe982633b0c359ac0cb43632f3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a707af3f34a9b12893a26ed1a140fdf262be0b944eb705b5dd201e741f06196
+size 228257
diff --git a/data/2025/2504_10xxx/2504.10356/layout.json b/data/2025/2504_10xxx/2504.10356/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..df28a9f9924091445a92e8b15b09496e4382a322
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10356/layout.json
@@ -0,0 +1,10796 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 139,
+ 97,
+ 471,
+ 138
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 97,
+ 471,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 97,
+ 471,
+ 138
+ ],
+ "type": "text",
+ "content": "MultiLoKo: a multilingual local knowledge benchmark for LLMs spanning 31 languages"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 219,
+ 178,
+ 390,
+ 190
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 178,
+ 390,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 178,
+ 390,
+ 190
+ ],
+ "type": "text",
+ "content": "Dieuwke Hupkes* Nikolay Bogoychev*"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 294,
+ 191,
+ 317,
+ 201
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 191,
+ 317,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 191,
+ 317,
+ 201
+ ],
+ "type": "text",
+ "content": "Meta"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 223,
+ 201,
+ 388,
+ 213
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 201,
+ 388,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 201,
+ 388,
+ 213
+ ],
+ "type": "text",
+ "content": "{dieuwkehupkes,nbogoych}@meta.com"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 281,
+ 241,
+ 329,
+ 253
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 241,
+ 329,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 241,
+ 329,
+ 253
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 140,
+ 265,
+ 470,
+ 506
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 265,
+ 470,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 265,
+ 470,
+ 506
+ ],
+ "type": "text",
+ "content": "We present MultiLoKo, a new benchmark for evaluating multilinguality in LLMs covering 31 languages. MultiLoKo consists of three partitions: a main partition consisting of 500 questions per language, separately sourced to be locally relevant to the specific language, and two translated partitions, containing human-authored translations from 30 non-English languages to English and vice versa. For comparison, we also release corresponding machine-authored translations. The data is equally distributed over two splits: a dev split and a blind, out-of-distribution test split. MultiLoKo can be used to study a variety of questions regarding the multilinguality of LLMs as well as meta-questions about multilingual benchmark creation. We compute MultiLoKo scores for 11 base and chat models marketed to be multilingual and study their average performance, their performance parity across languages, how much their ability to answer questions depends on the question language, and which languages are most difficult. None of the models we studied performs well on MultiLoKo, as indicated by low average scores as well as large differences between the best and worst scoring languages. Furthermore, we find a substantial effect of the question language, indicating suboptimal knowledge transfer between languages. Lastly, we find that using local vs English-translated data can result in differences more than 20 points for the best performing models, drastically change the estimated difficulty of some languages. For using machines instead of human translations, we find a weaker effect on ordering of language difficulty, a larger difference in model rankings, and a substantial drop in estimated performance for all models."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 190,
+ 537
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 190,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 190,
+ 537
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 104,
+ 548,
+ 504,
+ 638
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 548,
+ 504,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 548,
+ 504,
+ 638
+ ],
+ "type": "text",
+ "content": "With the growing presence and deployment of LLMs across the world, evaluating their abilities in languages other than English becomes more and more eminent. Yet, studying and evaluating multilinguality in LLMs remains a challenging enterprise, and it is hardly exaggerated to call the current state of multilingual evaluation in LLMs insufficient. Older multilingual benchmarks such as PAWS-X (Zhang et al., 2019), XNLI (Conneau et al., 2018) or XCOPA (Ponti et al., 2020) often do not fit the demands for evaluating auto-regressive models and are rarely used to evaluate recent models. Furthermore, their coverage of languages is relatively small compared to the number of languages in which LLMs are intended to be proficient."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 639,
+ 506,
+ 706
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 639,
+ 506,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 639,
+ 506,
+ 706
+ ],
+ "type": "text",
+ "content": "More often used for LLM evaluation are benchmarks translated from English, such as MGSM (translated GSM8K, Shi et al., 2023), MMMLU (tranlated MMLU, OpenAI, 2025) or (less frequently) Belebele (Bandarkar et al., 2024). These benchmarks provide good coverage over many languages, but using translated data comes with its own set of issues. One such issue is that even when human-rather than machine-authored translations are used, translated data is known to differ from native text in several ways (Clark et al., 2020). Furthermore, using translated benchmarks imposes a strong"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 14,
+ 209,
+ 37,
+ 560
+ ],
+ "type": "aside_text",
+ "angle": 270,
+ "lines": [
+ {
+ "bbox": [
+ 14,
+ 209,
+ 37,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 14,
+ 209,
+ 37,
+ 560
+ ],
+ "type": "text",
+ "content": "arXiv:2504.10356v2 [cs.CL] 15 Apr 2025"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 117,
+ 712,
+ 195,
+ 723
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 712,
+ 195,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 712,
+ 195,
+ 723
+ ],
+ "type": "text",
+ "content": "*Equal contributions"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 504,
+ 128
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 504,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 504,
+ 128
+ ],
+ "type": "text",
+ "content": "English-centric bias: translated data may be multilingual on the surface, it is not in its content. The benchmarks MLQA (Lewis et al., 2020) and TidyQA (Clark et al., 2020) to some extent address the issue by sourcing data separately for different languages. Even in their sourcing protocols, however, there is no explicit focus on selecting locally relevant content for the chosen languages. In addition to that, their coverage is again small compared to the above mentioned translated benchmarks."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 129,
+ 506,
+ 163
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 129,
+ 506,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 129,
+ 506,
+ 163
+ ],
+ "type": "text",
+ "content": "In response to these issues, we introduce a wide-coverage multilingual benchmark with locally-sourced questions for 31 different languages. Because the benchmark targets multilingual local knowledge, we dub it MultiLoKo. The release of MultiLoKo serves two interconnected goals:"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 127,
+ 165,
+ 478,
+ 189
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 4,
+ "blocks": [
+ {
+ "bbox": [
+ 127,
+ 165,
+ 382,
+ 175
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 127,
+ 165,
+ 382,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 165,
+ 382,
+ 175
+ ],
+ "type": "text",
+ "content": "1) Provide a better means to evaluate multilinguality in LLMs;"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 127,
+ 178,
+ 478,
+ 189
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 127,
+ 178,
+ 478,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 178,
+ 478,
+ 189
+ ],
+ "type": "text",
+ "content": "2) Provide data to study the effect of various design choices in multilingual evaluation."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 104,
+ 191,
+ 506,
+ 301
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 191,
+ 506,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 191,
+ 506,
+ 301
+ ],
+ "type": "text",
+ "content": "To address our first goal, we create 500 questions per language, written from scratch for each language, using a sourcing protocol specifically designed to ensure local relevance of the question topics. To also reap the benefits of parallel data, we commissioned both human and machine-authored translations for all non-English questions into English and vice versa, providing a total of 15500 parallel questions, sourced across the 31 languages in the benchmark. The translated data facilitates the study of transfer between languages and also serves our second goal. By comparing the English-translated data with the locally sourced data, we can explicitly compare the adequacy of using translated benchmarks; by comparing human- with machine-authored translations, we can better estimate the potential issues of the latter. To prevent quick overfitting and inadvertent contamination, we release a development set of the benchmark, while test scores can only be obtained through an external provider.| Average EM | The first main metric we use to quantify performance for MultiLoKo is the average Exact Match score across languages, which expresses how many of the answers match one of the gold standard answers verbatim (after post-processing the answers). |
| Gap | The second main metric is the gap between a model's best and worst performing language. We gap to quantify the extent to which a model has achieved parity across languages. Because a small gap can be achieved both through parity on high scores as parity on low scores, it is most informative in combination with average benchmark performance. |
| Mother tongue effect (MTE) | MTE expresses the impact of asking questions in a language in which the requested information is locally salient, compared to asking it in English. A positive MTE indicates information is more readily available in the language it was (likely) present in the training data, whereas a negative mother tongue effect indicates the information is more easily accessible in English. |
| Locality effect (LE) | LE quantifies the effect of using locally sourced vs translated data. It is measured by computing the difference between scores for locally sourced data and translated English-sourced data. A positive LE implies that using translated English data underestimates performance on a language, a negative LE that using translated English data overestimates performance. |
",
+ "image_path": "8d54153ccb122aaea2874537a950dad799895f33acc86ca682873fd95442f949.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 239,
+ 504,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 239,
+ 504,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 239,
+ 504,
+ 264
+ ],
+ "type": "text",
+ "content": "Table 1: MultiLoKo metric cheatsheet. We use several metrics to quantify model performance using MultiLoKo. This table provides a cheatsheet for their meaning."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 104,
+ 293,
+ 506,
+ 392
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 293,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 293,
+ 506,
+ 392
+ ],
+ "type": "text",
+ "content": "LE The locality effect (LE) is a measure of how much performance on knowledge tasks is over- or underestimated through the use of using translated English data, as opposed to locally relevant data. We quantify the locality effect as the difference in EM for English translated data and locally sourced data. If for a language the English translated data has as a higher EM, the LE is positive, indicating that using English translated data likely overestimating a model's ability on providing knowledge for that language. If the LE is negative the English translated data may provide an underestimation of the score for that language. Note that because we often observe both positive and negative LEs for the 30 non-English languages in MultiLoKo, the average LE across languages may be small, even if the differences for individual languages may be large."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 228,
+ 424
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 228,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 228,
+ 424
+ ],
+ "type": "text",
+ "content": "4 Experimental setup"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 434,
+ 504,
+ 458
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 434,
+ 504,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 434,
+ 504,
+ 458
+ ],
+ "type": "text",
+ "content": "We test and showcase our benchmark by running experiments with 11 different models of varying sizes, that were all marketed to have multilingual abilities."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 163,
+ 479
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 163,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 163,
+ 479
+ ],
+ "type": "text",
+ "content": "4.1 Models"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 487,
+ 504,
+ 555
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 487,
+ 504,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 487,
+ 504,
+ 555
+ ],
+ "type": "text",
+ "content": "To test the extent to which MultiLoKo provides useful signal across training stages, we consider both base and chat models. The base models we include in our experiments are Llama 3.1 70B and 405B (Dubey et al., 2024), Mixtral 8x22B (team, 2024), and Qwen 2.5 72B (Qwen et al., 2025), the seven chat models are Gemini 2.0 Flash (Google DeepMind, 2024), GPT4-o (OpenAI et al., 2024), Claude 3.5 Sonnet (Anthropic, 2025), Llama 3.1 70B and 405B Chat, Mixtral 8x22B-it, and Qwen 2.5 72B instruct. As mentioned before, we run chat and base models with separate prompts."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 215,
+ 582
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 215,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 215,
+ 582
+ ],
+ "type": "text",
+ "content": "4.2 Experimental setup"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 104,
+ 590,
+ 504,
+ 624
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 590,
+ 504,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 590,
+ 504,
+ 624
+ ],
+ "type": "text",
+ "content": "We run all of our experiments with the generation temperature set to 0. To facilitate automatic evaluation, we include an instruction to answer questions curtly and precisely, producing only a number/name/location/etc. Full template information can be found in our github repository."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 628,
+ 504,
+ 651
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 628,
+ 504,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 628,
+ 504,
+ 651
+ ],
+ "type": "text",
+ "content": "Few-shot prompting For base models we use a 5-shot prompt. For chat models, we use a 0-shot prompt, as this is the most likely use mode by chat model users."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 656,
+ 506,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 656,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 656,
+ 506,
+ 723
+ ],
+ "type": "text",
+ "content": "Post-processing Because base models are good at following the instructions, minimal postprocessing is needed: we only lowercase the output and strip punctuation. Chat models often deviate from the required format, especially in English, in various ways that we discuss in Appendix B. To evaluate such models beyond their instruction-following issues, we perform more complex post-processing, aiming to remove any words resembling \"answer\" from the LLM output, as well as several special cases for English and Japanese. We provide full details about post-processing in Appendix C."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "5"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 119,
+ 70,
+ 486,
+ 215
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 119,
+ 70,
+ 486,
+ 215
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 70,
+ 486,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 70,
+ 486,
+ 215
+ ],
+ "type": "table",
+ "html": " | Description | Example |
| 1. | Extremely local and relatively obscure. Content that is of interest only to a small, localized group, such as a specific town, region, or community. These topics are typically obscure and not widely known beyond their immediate area. | Local radio stations, small town historical events, regional businesses, or niche local cultural practices. |
| 2. | Regional interest. Topics that have some relevance beyond a specific locality but are still primarily of interest within a particular region or country. | State or provincial politicians, regional cuisine, local sports teams, or medium-sized companies with regional influence. |
| 3. | National Significance. Content that is widely recognized within a single country, but relatively un-known internationally. | National politicians (not internationally known), popular national media figures, major corporations within a country, or significant national historical events. |
| 4. | International recognition. Topics that are recognized and have relevance in multiple countries but may not be universally known across the globe. These topics often have international influence and are likely to be covered in international media, though their impact may vary by region. | International brands which may be recognized in more than one country, celebrities with some international reach, significant cultural movements, or political conflicts with some awareness on the international stage. |
| 5. | Global prominence. Content that is widely recognized and relevant across a large number of countries around the world. These topics have a global impact or appeal and are likely to be well-represented in media across diverse cultures and regions. | Globally famous celebrities (e.g., Cristiano Ronaldo), multinational corporations (e.g., Apple), major world events, or universally recognized cultural icons. |
",
+ "image_path": "fa2dfae75eef0a7bb838f0dfc25bd66e8d7acffe982633b0c359ac0cb43632f3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 594,
+ 504,
+ 617
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 594,
+ 504,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 594,
+ 504,
+ 617
+ ],
+ "type": "text",
+ "content": "Table 3: Rubric for locality rating task. In the locality rating task, we ask the annotators to rate paragraphs with respect to how locally relevant the topic is to the locale."
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 105,
+ 647,
+ 219,
+ 658
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 219,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 219,
+ 658
+ ],
+ "type": "text",
+ "content": "D.2 Question generation"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 667,
+ 505,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 667,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 667,
+ 505,
+ 723
+ ],
+ "type": "text",
+ "content": "The second and main annotation step in our pipeline is the step in which we ask annotators to generate questions about sampled paragraphs. We ask annotators to generate a challenging question with a short answer. The answer should be easy to evaluate with string-matching metrics, the questions should not be open-ended or have many possible correct answers, be ambiguous or subjective, and the expected short answer should be concise. To ensure difficulty, we ask that answering the question"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "24"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 23
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 204
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 204
+ ],
+ "type": "text",
+ "content": "requires combining information from different parts in the accompanying text; It should not be answerable by mere regurgitation of a single sentence. We furthermore ask that the question is formulated such that its answer will not change over time (e.g. not 'How many medals has Sifan Hassan won', but 'How many medals has Sifan Hassan won between 2018 and 2022 (including)'), and that the question is answerable also without the article (e.g. not 'How many tv shows did the person in this article produce?'). To facilitate validation checks in the next round, we also ask that the question authors write a longer answer to explain how they arrived at the short answer. We also ask the question authors to annotate what is the type of the correct answer (e.g. number, name, date, etc) In the pilot, we observed that – for some languages – the vast majority of questions were questions that required some form of numerical reasoning. Because the intention of the benchmark is to address knowledge more than reasoning, we afterwards restricted the number of numerical questions to "
+ },
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 204
+ ],
+ "type": "inline_equation",
+ "content": "10\\%"
+ },
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 204
+ ],
+ "type": "text",
+ "content": ". Similarly, we asked question authors to avoid yes/no questions."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 217,
+ 202,
+ 228
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 202,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 202,
+ 228
+ ],
+ "type": "text",
+ "content": "D.3 Question review"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 236,
+ 504,
+ 270
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 236,
+ 504,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 236,
+ 504,
+ 270
+ ],
+ "type": "text",
+ "content": "In the first round of question review, we asked annotators from a different provider to judge whether the questions abide by the rules provided to the question authors. All question reviewers are native speakers. Specifically, we ask them to check if:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 132,
+ 278,
+ 504,
+ 388
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 10,
+ "blocks": [
+ {
+ "bbox": [
+ 132,
+ 278,
+ 334,
+ 289
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 278,
+ 334,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 278,
+ 334,
+ 289
+ ],
+ "type": "text",
+ "content": "- The question pertains to a locally relevant topic"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 132,
+ 293,
+ 380,
+ 304
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 293,
+ 380,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 293,
+ 380,
+ 304
+ ],
+ "type": "text",
+ "content": "- The question is clear and understandable, and not subjective"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 132,
+ 308,
+ 321,
+ 319
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 308,
+ 321,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 308,
+ 321,
+ 319
+ ],
+ "type": "text",
+ "content": "- The question has a clear and concise answer"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 132,
+ 322,
+ 504,
+ 344
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 322,
+ 504,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 322,
+ 504,
+ 344
+ ],
+ "type": "text",
+ "content": "- If there are multiple possible variations of the answer possible (e.g. 'Dick Schoof' / 'Minister Dick Schoof' / 'Prime Minister Dick Schoof' / etc), all versions of the answer are provided."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 132,
+ 348,
+ 351,
+ 359
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 348,
+ 351,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 348,
+ 351,
+ 359
+ ],
+ "type": "text",
+ "content": "- The question and answer are in the correct language"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 132,
+ 362,
+ 343,
+ 373
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 362,
+ 343,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 362,
+ 343,
+ 373
+ ],
+ "type": "text",
+ "content": "- The question is understandable without the article"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 132,
+ 376,
+ 429,
+ 388
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 376,
+ 429,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 376,
+ 429,
+ 388
+ ],
+ "type": "text",
+ "content": "- That the answer to the question will not likely change in the near future"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 104,
+ 397,
+ 504,
+ 463
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 397,
+ 504,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 397,
+ 504,
+ 463
+ ],
+ "type": "text",
+ "content": "When a question can be fixed with a minor change (e.g. add a time indication to make sure an answer will not change in the near future, or add an extra answer version), we ask the question reviewers to implement this fix and describe it. In the pilot round, we use the annotator feedback to finetune our annotation protocol and provide feedback to the question-authors. During the rest of the data collection, we simply disregard questions that are not useable as is or can be corrected with minor changes."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 298,
+ 488
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 298,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 298,
+ 488
+ ],
+ "type": "text",
+ "content": "D.4 Validation through question answering"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 617
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 496,
+ 506,
+ 617
+ ],
+ "type": "text",
+ "content": "In the last stage of our question generation pipeline, we have additional annotators answer the sourced and reviewed question. The goal of this validation task is to confirm that the questions are answerable, correct, non-ambiguous when read by individuals other than the original question author, and that all possible versions of the answers are included. For each question, we ask two additional annotators to first answer the question, using the snippets the questions were sourced from for context. After they have answered the question, they are shown the list of reference answers written by the original author of the question as well as the rational they provided, and we ask them to reflect upon the answer they gave themselves. If their answer did not match any answer in the original reference list, we ask them to either add their answer to the list if it is semantically equivalent to their own answer or indicate which answer they believe to be correct, their own or the original answer. We disregard all questions where at least one annotator disagrees with the original question author."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 197,
+ 644
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 197,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 197,
+ 644
+ ],
+ "type": "text",
+ "content": "E Related work"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 656,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 656,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 656,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "content": "In this paper, we introduce a new multilingual benchmark for LLMs, that we believe addresses gaps and pitfalls in existing benchmarks. We (concisely) outlined those gaps and pitfalls and mentioned several other works related to ours in the introduction of those paper. Here, we discuss multilingual evaluation of LLMs in more detail. Specifically, we discuss what datasets recent LLM releases have used for multilingual evaluation (Appendix E.1) and what other datasets and approaches they could have used but did not (Appendix E.2)."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "25"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 24
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 70,
+ 504,
+ 218
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 108,
+ 70,
+ 504,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 70,
+ 504,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 70,
+ 504,
+ 218
+ ],
+ "type": "table",
+ "html": "| Claude 3.5 Sonnet | MGSM (Shi et al., 2023) |
| Gemini 2.0 Flash | Mentions multilingual audio, no multilingual benchmarks scores reported. |
| GPT4-o | ARC-Easy and TruthfulQA translated into five African languages (internal benchmark), Uhura-Eval (internal benchmark). |
| Llama 3.1 | MGSM (Shi et al., 2023), Multilingual MMLU (internal benchmark) |
| Mixtral 8x22B | translated ARC-C, HellaSwag and MMLU (internal benchmarks) |
| Qwen2.5 72B | M3Exam (Zhang et al., 2023), IndoMMLU (Koto et al., 2023), ruMMLU (Fenogenova et al., 2024), translated MMLU (Chen et al., 2023), Belebele (Bandarkar et al., 2024), XCOPA (Ponti et al., 2020), XWinograd (Muennighoff et al., 2023), XStoryClose (Lin et al., 2022), PAWS-X (Zhang et al., 2019), MGSM (Shi et al., 2023), Flores-101 (Goyal et al., 2022) |
",
+ "image_path": "6097bbba21f178dc4be15a6d96ad58e0b1c3c7cc17e299a76a31147c2d105aca.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 221,
+ 504,
+ 257
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 221,
+ 504,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 221,
+ 504,
+ 257
+ ],
+ "type": "text",
+ "content": "Table 4: Multilingual evaluation of recent LLM releases, overview. We provide an overview table of the benchmark for which scores are reported in the release papers or notes of the LLMs we evaluated in this paper. Models are sorted alphabetically."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 320,
+ 296
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 320,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 320,
+ 296
+ ],
+ "type": "text",
+ "content": "E.1 Multilingual evaluation of LLMs in practice"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 304,
+ 504,
+ 403
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 304,
+ 504,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 304,
+ 504,
+ 403
+ ],
+ "type": "text",
+ "content": "While multilinguality is something frequently mentioned in the release papers or posts of recent LLM releases, the datasets for which they actual report scores is in most cases quite limited. Of the models that we evaluated for this paper, Gemini 2.0 Flash reported no multilingual scores at all; GPT4-o and Mixtral 8x22B report scores only on internally translated but not publicly available English benchmarks; Claude 3.5 Sonnet reports scores for only one benchmark - MGSM. MGSM is also the only publicly available benchmark for which Llama 3.1 reports scores, along with - also - an internally translated version of MMLU that is not publicly available. The only model that extensively reports multilingual benchmark values, on more than 10 benchmarks, is Qwen2.5 72B. We provide an overview of the multilingual benchmarks for which scores are reported for these models in Table 4."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 415,
+ 309,
+ 427
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 309,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 415,
+ 309,
+ 427
+ ],
+ "type": "text",
+ "content": "E.2 Multilingual evaluation options for LLMs"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 435,
+ 506,
+ 469
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 435,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 435,
+ 506,
+ 469
+ ],
+ "type": "text",
+ "content": "While, as we discuss below, there are gaps and challenges with multilingual evaluation for LLMs, there are in fact many more options than is suggested by what is reported in recent releases. Below, we discuss other options for multilingual LLM evaluation."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 677
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 479,
+ 506,
+ 677
+ ],
+ "type": "text",
+ "content": "Translated English benchmarks As mentioned earlier on, benchmarks used for LLM evaluation are often translated English benchmarks. In some cases, the benchmarks were designed to evaluate only English and translated later, such as translated MMLU (e.g. Li et al., 2024; Chen et al., 2023; OpenAI, 2025; Singh et al., 2024) or MMLU-ProX (Xuan et al., 2025), MGSM (Shi et al., 2023) or MLAMA (Kassner et al., 2021). In other cases, the benchmark was multilingual at the time of its creation, but means of creation of the non-English data was through translating English sourced data, such as Belebele Bandarkar et al. (2024), Mintaka (Sen et al., 2022), or X-FACTR (Jiang et al., 2020). Taken together, translated benchmarks span quite a range of tasks, such as question answering (Artetxe et al., 2020; Lewis et al., 2020; Qi et al., 2023; Ohmer et al., 2023), natural language inference (Conneau et al., 2018), paraphrase detection (Zhang et al., 2019), general linguistic competence (Jumelet et al., 2025), reading comprehension (Artetxe et al., 2020; Bandarkar et al., 2024) and commonsense reasoning (Ponti et al., 2020), and even instruction following (He et al., 2024). With the exception of question answering and of course instruction following, however, many of these tasks have gone (somewhat) out of fashion for LLM evaluation, a trend which is mirrored also in the usage of their multilingual counterparts. As mentioned before, translated benchmarks have the advantage of containing parallel data, allowing for some form of comparability across languages, but are English-centric in content and may suffer from translationese (see e.g. Romanou et al., 2024; Chen et al., 2024, for a recent discussion of this)."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 689,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 689,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 689,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "content": "Multilingual benchmarks sourced from scratch Though much rarer, there are also benchmarks that are created independently for each language they include. Clark et al. (2020) release a question answering dataset separately sourced for 11 different languages, with a protocol relatively similar"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "26"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 25
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 193
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 72,
+ 506,
+ 193
+ ],
+ "type": "text",
+ "content": "to ours. In a different category, Hardalov et al. (2020), Zhang et al. (2023) and Romanou et al. (2024) and Sánchez et al. (2024) do not create benchmark data, but instead collect existing exam or competition questions from official human exams. In case of Zhang et al. (2023), the exams are graduation exams of primary, middle and high school; Hardalov et al. (2020) includes official state exams taken by graduating high school students, which may contain parallel pairs in case countries allow examinations to be taken in multiple languages; Romanou et al. (2024), cover academic exams at middle and high school and university level, professional certifications and licenses, and exams to obtain regional licenses. Sánchez et al. (2024) instead focus on questions from the International Linguistic Olympiad corpus. Lastly, as part of their study Ohmer et al. (2023) create a dataset called SIMPLE FACTS, containing factual questions created through a shared template filled in with language specific factual data."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 205,
+ 506,
+ 315
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 205,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 205,
+ 506,
+ 315
+ ],
+ "type": "text",
+ "content": "Consistency evaluation A rather different approach to assess multilinguality in LLMs is to focus not on accuracy across different languages, but to consider whether predictions are consistent across languages. This tests knowledge and skill transfer between languages more explicitly. Two recent examples of studies incorporating consistency-based evaluations on factual knowledge questions are Qi et al. (2023) and Ohmer et al. (2023). Qi et al. (2023) focuseses specifically on sample-level consistency of answers across different languages, requiring existing parallel benchmarks. Ohmer et al. (2023), instead, ask models to translate benchmark questions themselves before answering them again. This can, with some caveats, be applied to any existing monolingual benchmark, but – requiring multiple steps – it is more involved as a paradigm, and is somewhat bottlenecked by the translation ability of the model to be evaluated."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 326,
+ 504,
+ 360
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 326,
+ 504,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 326,
+ 504,
+ 360
+ ],
+ "type": "text",
+ "content": "Translation as a proxy for multilinguality Another, more implicit method to assess multilinguality in LLMs is to evaluate their ability to translate from one language to another. This approach was famously used by Brown et al. (2020), but has not been common since."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 492
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 506,
+ 492
+ ],
+ "type": "text",
+ "content": "Monolingual non-English evaluation In our discussion, we have focussed on multilingual evaluation options that cover multiple other languages. After all, a benchmark to evaluate models on Bengali (e.g. Shafayat et al., 2024) or Arabic (e.g. Alwajih et al., 2024) can contribute to multilingual evaluation when combined with other benchmarks, but does not so on its own. Because such benchmarks are usually created by language experts for the respective languages, they usually target locally relevant skills and knowledge and are likely of higher quality than benchmarks created for many languages simultaneously (either through translation or from scratch). Yet, composing a suite including many languages that allows direct comparisons between languages remains challenging. We believe such benchmarks can be important for multilingual evaluation in LLMs, but will not further discuss benchmarks focussing on individual languages or very small sets of languages within one family here."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "27"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 26
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10368/75b91eea-3940-4d8f-9204-ff0cff897b91_content_list.json b/data/2025/2504_10xxx/2504.10368/75b91eea-3940-4d8f-9204-ff0cff897b91_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..9028bceb2fecd2d05bc5a62a02e42cdb8084d352
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10368/75b91eea-3940-4d8f-9204-ff0cff897b91_content_list.json
@@ -0,0 +1,5860 @@
+[
+ {
+ "type": "text",
+ "text": "S1-Bench: A Simple Benchmark for Evaluating System 1 Thinking Capability of Large Reasoning Models",
+ "text_level": 1,
+ "bbox": [
+ 152,
+ 89,
+ 843,
+ 130
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Wenyuan Zhang*, Shuaiyi Nie*, Xinghua Zhang, Zefeng Zhang, Tingwen Liu†",
+ "bbox": [
+ 164,
+ 151,
+ 835,
+ 168
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Institute of Information Engineering, Chinese Academy of Sciences",
+ "bbox": [
+ 220,
+ 168,
+ 773,
+ 185
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "School of Cyber Security, University of Chinese Academy of Sciences",
+ "bbox": [
+ 213,
+ 186,
+ 784,
+ 202
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "{zhangwenyuan,nieshuaiyi,liutingwen}@iei.ac.cn",
+ "bbox": [
+ 267,
+ 203,
+ 731,
+ 219
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 260,
+ 261,
+ 339,
+ 275
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "We introduce S1-Bench, a novel benchmark designed to evaluate the performance of Large Reasoning Models (LRMs) on simple tasks that favor intuitive system 1 thinking rather than deliberative system 2 reasoning. While LRMs have achieved significant breakthroughs in complex reasoning tasks through explicit chains of thought, their heavy reliance on system 2 thinking may limit their system 1 thinking capabilities. However, there is a lack of an appropriate benchmark for evaluating LRM's system 1 thinking capabilities. To fill this gap, S1-Bench introduces a suite of simple, diverse, and natural questions across multiple domains and languages, specifically designed to assess LRMs' performance on questions more suitable for system 1. We conduct extensive evaluations across 28 LRMs, revealing their inefficiency, inadequate accuracy, and limited robustness when handling simple questions. Additionally, we observe a gap between their difficulty perception and generation length. Overall, this work paves the way toward dual-system compatibility in the development of LRMs| Paper Abbreviation | MATH | GSM8K | AIME | GPQA | AMC | MMLU | Olympiad-Bench | SVAMP | LiveCode-Bench | Common-SenseQA |
| Codi (Shen et al., 2025b) | | ✓ | | | | | | ✓ | | |
| CISC (Taubenfeld et al., 2025) | ✓ | ✓ | | | | ✓ | | | | |
| CoT-Valve (Ma et al., 2025b) | | ✓ | ✓ | | | | | | | |
| Dast (Shen et al., 2025a) | ✓ | | ✓ | | | | | | | |
| ATM (Chen et al., 2024a) | | ✓ | | | | | | ✓ | | ✓ |
| DEER (Yang et al., 2025b) | ✓ | | ✓ | ✓ | ✓ | | | | | |
| DPTS (Ding et al., 2025) | ✓ | ✓ | | | | | | | | |
| Dynasor (Fu et al., 2024) | | ✓ | | | | | | | ✓ | |
| ESC (Li et al., 2024) | ✓ | ✓ | | | | | | ✓ | | ✓ |
| Hawkeye (She et al., 2025) | ✓ | ✓ | ✓ | | | | | | | |
| token complexity (Lee et al., 2025) | ✓ | ✓ | | | | ✓ | | | | |
| INFTYTHINK (Yan et al., 2025) | ✓ | | ✓ | ✓ | | | | | | |
| KIMI K1.5 (Team et al., 2025a) | ✓ | | ✓ | | | ✓ | | | ✓ | |
| L1 (Aggarwal and Welleck, 2025) | ✓ | | ✓ | ✓ | | ✓ | ✓ | | | |
| LightThinker (Zhang et al., 2025a) | | ✓ | | ✓ | | ✓ | | | | |
| LS-Mixture SFT (Yu et al., 2025a) | ✓ | | ✓ | ✓ | | | | | | |
| DSC (Wang et al., 2024) | ✓ | ✓ | | | | | | | | |
| O1-Pruner (Luo et al., 2025) | ✓ | ✓ | | | | | | | | |
| MRT (Qu et al., 2025b) | ✓ | | ✓ | ✓ | | | | | | |
| Self-Doubt (Fu et al., 2025) | ✓ | | ✓ | | ✓ | | | | | |
| RASC (Wan et al., 2024) | | ✓ | | | | | | ✓ | | ✓ |
| NoThinking (Ma et al., 2025a) | | | ✓ | | ✓ | | ✓ | | ✓ | |
| Retro-Search (Lu et al., 2025) | ✓ | ✓ | ✓ | | ✓ | | ✓ | | | |
| RSD (Liao et al., 2025) | ✓ | ✓ | | ✓ | | ✓ | ✓ | | | |
| ST-BoN (Wang et al., 2025b) | ✓ | | | ✓ | | | | | | |
| Elastic Reasoning (Xu et al., 2025b) | ✓ | | ✓ | | ✓ | | ✓ | | ✓ | |
| FS-BoN (Munkhbat et al., 2025) | ✓ | ✓ | | | | | | | | |
| SoT (Aytes et al., 2025) | | ✓ | | | | | | ✓ | | ✓ |
| SpecReason (Pan et al., 2025) | ✓ | | ✓ | ✓ | | | | | | |
| Speculative Thinking (Yang et al., 2025d) | ✓ | | ✓ | ✓ | ✓ | | | | | |
| SPIRIT (Cui et al., 2025) | | ✓ | | | | | | | | |
| ITC Analysis (Wang et al., 2025a) | ✓ | | ✓ | ✓ | | | | | | |
| Think when needed (Yang et al., 2025c) | | | ✓ | | | | | | | |
| THINKPRUNE (Hou et al., 2025) | ✓ | | ✓ | | ✓ | | ✓ | | | |
| TALE (Han et al., 2024) | | ✓ | | | | | | | | |
| TokenSkip (Xia et al., 2025) | ✓ | ✓ | | | | | | | | |
| TOPS (Yang et al., 2025e) | ✓ | ✓ | ✓ | | | | | | | |
| efficient reasoning (Arora and Zanette, 2025) | ✓ | ✓ | ✓ | | | | | | | ✓ |
| TWT (Xu et al., 2025a) | | | | | | | | | | |
| Z1 (Yu et al., 2025b) | ✓ | | | ✓ | | | | | ✓ | |
| Count | 28 | 24 | 20 | 11 | 8 | 6 | 6 | 5 | 5 | 5 |
",
+ "bbox": [
+ 117,
+ 218,
+ 878,
+ 740
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Table 7: A total of 40 studies on LRM efficiency before May 2025 were included. Benchmarks that appeared more than four times are listed.",
+ "bbox": [
+ 112,
+ 749,
+ 882,
+ 778
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "16",
+ "bbox": [
+ 489,
+ 941,
+ 509,
+ 954
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "table",
+ "img_path": "images/832fc96c01a1dfa5c068c72fb411254118d8eaa8000d44b0e31f4b129b28d3d1.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| cate. | subcategories | Explanation and cases |
| reasoning question | numerical reasoning | Questions that require performing basic mathematical operations or solving simple algebraic equations to arrive at a numerical answer.\nCase: What's two plus three? |
| code reasoning | Questions that require tracing through and executing simple code snippets to determine their output or behavior when run in a specific programming environment.\nCase: What is the output of the following code when run in Python 3 environment: word = "hello"\\nprint(len(word)) |
| set reasoning | Questions that require applying simple syllogistic reasoning to determine whether elements belong to sets based on clearly stated relationships.\nCase: All squares are quadrilaterals. A shape is a square, is it a quadrilateral? |
| temporal reasoning | Questions that require calculating time durations, ages, or future dates by applying simple arithmetic operations to temporal information.\nCase: How many minutes equal 120 seconds? |
| spatial reasoning | Questions that require determining relative positions, directions, or orientations of objects in space based on simple spatial relationships.\nCase: If a bird is flying above a tree, where is the tree in relation to the bird? |
| causal reasoning | Questions that require determining outcomes by applying simple cause-and-effect relationships based on given conditional statements.\nCase: If ferromagnetic material is placed in a magnetic field, it will become magnetized. An iron nail was placed next to a strong magnet for some time. Has the nail been magnetized? |
| natural law reasoning | Questions that require applying basic knowledge of physical laws and natural phenomena to predict simple observable outcomes in everyday scenarios.\nCase: Which is faster, an airplane or the propagation of light? |
| knowledge question | geometry facts | Questions that require recalling simple and fundamental geometric properties about shapes, angles, and basic geometric figures.\nCase: How many angles does a trapezoid have? |
| geographic facts | Questions that require recalling simple factual information about locations, landmarks, political divisions, celestial bodies, and other basic geographic knowledge.\nCase: Which is the largest continent on Earth? |
| historical facts | Questions that require recalling basic facts about historical events.\nCase: Which country first invented paper? |
| biographical facts | Questions that require recalling basic facts about the identities, achievements, and characteristics of historical figures.\nCase: Who proposed the theory of universal gravitation? |
| measurement units | Questions that require recalling simple conversion relationships between standard units of measurement.\nCase: How many centimeters equal 1 meter? |
| scientific notation | Questions that require recalling basic scientific symbols, formulas, and standard units used in scientific communication.\nCase: What is the chemical symbol for oxygen? |
| creative authorship | Questions that require recalling the creators or originators of notable artistic, literary, musical, and cultural works.\nCase: Who is the author of Hamlet? |
",
+ "bbox": [
+ 129,
+ 171,
+ 863,
+ 803
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Table 8: The subcategory descriptions and cases of reasoning questions and knowledge questions.",
+ "bbox": [
+ 166,
+ 813,
+ 826,
+ 828
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "page_number",
+ "text": "17",
+ "bbox": [
+ 489,
+ 941,
+ 509,
+ 953
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "table",
+ "img_path": "images/1ca33c1ae16b6590af6a367ff5a0eae2b58a329182ae7f749aecd0ce9bf5cb6f.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| cate. | subcatrgories | Explanation and cases |
| instruction following | repetition constraints | Questions that require outputting specified characters, words, or phrases a specific number of times according to simple formatting instructions.\nCase: Output the number "7" four times, without using separators. |
| length constraints | Questions that require generating outputs of a specific length or with a specific number of components based on simple counting constraints.\nCase: Output a four-digit number. |
| character constraints | Questions that require generating words or numbers that conform to simple specified character patterns or formatting rules.\nCase: Output a number that begins with 8. |
| counting constraints | Questions that require counting specific characters or elements within a given text or sequence.\nCase: Output the number of letter "y" in the word "yes". |
| transformation constraints | Questions that require modifying text or numbers according to simple formatting or character substitution rules to produce a transformed output.\nCase: Output the word "good" with all letters capitalized directly. |
| sentence constraints | Questions that require generating sentences that conform to simple specified content or structural requirements.\nCase: Give a sentence that contains the phrase "have lunch" directly. |
| analysis question | sentiment classification | Questions that require determining whether simple statements express positive or negative emotions based on the tone and word choice.\nCase: Does the sentence "I hate rainy days." express a positive or negative emotion? |
| named entity recognition | Questions that require identifying the correct category of named entities (such as people, places, organizations, or time expressions) within simple sentences.\nCase: In the sentence "Napoleon died in 1821", is "1821" a time or a place name? |
| language classification | Questions that require identifying the language of origin for simple words or phrases based on their characteristic writing systems or common vocabulary.\nCase: Is the word "hello" English or Japanese? |
| topic classification | Questions that require identifying the primary subject matter or thematic category of simple sentences based on their content and context clues.\nCase: Is the topic of the sentence "The stock market rose 2% today" finance or technology? |
| intent recognition | Questions that require determining the communicative purpose behind simple utterances or statements based on their phrasing and context.\nCase: Is the intention of the sentence "I'm sorry I'm late." to apologize or to blame? |
| syntax classification | Questions that require identifying the correct grammatical structure or sentence type of simple expressions based on their form, punctuation, and communicative function.\nCase: Is "Close the door!" an imperative sentence or an interrogative sentence? |
| grammar classification | Questions that require identifying simple grammatical properties (like tense, voice, or polarity) of sentences based on their structure and verb forms.\nCase: Is "The apple was eaten." in active voice or passive voice? |
| coreference resolution | Questions that require identifying which entity a pronoun or reference term refers to in simple sentences by tracking relationships between words in the text.\nCase: In "My computer is broken, and I need to fix it." What does "it" refer to? |
",
+ "bbox": [
+ 115,
+ 177,
+ 878,
+ 797
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "text",
+ "text": "Table 9: The subcategory descriptions and cases of instruction following questions and analysis questions.",
+ "bbox": [
+ 139,
+ 807,
+ 855,
+ 822
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "page_number",
+ "text": "18",
+ "bbox": [
+ 489,
+ 941,
+ 509,
+ 954
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "text",
+ "text": "Prompt for construction workflow for S1-Bench",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 162,
+ 410,
+ 175
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Data Generation Prompt",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 179,
+ 265,
+ 192
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Generate 50 pairs of questions and answers in both Chinese and English based on the category's name, definition, and specific simplicity criteria. The following conditions must be satisfied:",
+ "bbox": [
+ 115,
+ 195,
+ 882,
+ 218
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "1. Questions must be naturally and clearly expressed, unambiguous, and free of intentional traps.",
+ "2. Answers must be unique or easily falsifiable, with no possibility of multiple correct answers.",
+ "3. Make the questions as diverse as possible."
+ ],
+ "bbox": [
+ 115,
+ 217,
+ 670,
+ 249
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Category Name and Definition: {name_and_defined}",
+ "bbox": [
+ 115,
+ 253,
+ 309,
+ 274
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Specific Simplicity Criteria: {criteria}",
+ "bbox": [
+ 115,
+ 279,
+ 289,
+ 300
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Cases: \n## English question: {question_en} \n## English Answer: {answer_en}",
+ "bbox": [
+ 115,
+ 305,
+ 235,
+ 357
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "```c\n## Chinese question:\n{question_zh}\n## Chinese Answer:\n{answer_zh}",
+ "bbox": [
+ 115,
+ 361,
+ 236,
+ 403
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Please generate 50 pairs of Chinese and English questions and answers in the following format: [question]English-question[answer]English-answer[question]Chinese-question[answer]Chinese-answer...",
+ "bbox": [
+ 115,
+ 407,
+ 717,
+ 429
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Start generating:",
+ "bbox": [
+ 115,
+ 432,
+ 213,
+ 445
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Quality Discrimination Prompt",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 450,
+ 302,
+ 462
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "Given a question, its answer, and its category, please analyze from the following perspectives as comprehensively as possible: 1. Whether the question belongs to the specified category and meet the Specific Simplicity Criteria.",
+ "2. Whether the question is easy, clear, unambiguous, and has an absolutely unique answer.",
+ "3. Whether the answer is absolutely correct; if not, what the correct answer should be.",
+ "4. Whether the question is similar to other given questions, and if similar, whether more diverse questions can be generated."
+ ],
+ "bbox": [
+ 115,
+ 466,
+ 833,
+ 519
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Category Name and Definition: {name_and_defined}",
+ "bbox": [
+ 115,
+ 523,
+ 309,
+ 544
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Specific Simplicity Criteria: {criteria}",
+ "bbox": [
+ 115,
+ 549,
+ 289,
+ 571
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Question and Answer: {question_with_answer}",
+ "bbox": [
+ 115,
+ 575,
+ 258,
+ 596
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Other Questions: {questions_list}",
+ "bbox": [
+ 115,
+ 601,
+ 226,
+ 621
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Begin your analysis, aiming to be as detailed and comprehensive as possible:",
+ "bbox": [
+ 115,
+ 626,
+ 556,
+ 639
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Difficulty Reduction Prompt",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 643,
+ 285,
+ 656
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "Given a question and answer that are too complex for the model to answer correctly, you need to further reduce their difficulty while trying to:",
+ "- Ensure the question aligns with the Category Name and Definition.",
+ "- Ensure the question meets the Specific Simplicity Criteria.",
+ "Category Name and Definition: {name_and_defined}",
+ "Specific Simplicity Criteria: {criteria}",
+ "Question and Answer: {question_with_answer}",
+ "The new question and answer:"
+ ],
+ "bbox": [
+ 115,
+ 659,
+ 880,
+ 796
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "text",
+ "text": "Table 10: \"Category Name and Definition\" refers to the subcategory name and its definition, while Specific Simplicity Criteria refers to the simplicity requirements specific to the main category.",
+ "bbox": [
+ 112,
+ 810,
+ 880,
+ 839
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "page_number",
+ "text": "19",
+ "bbox": [
+ 489,
+ 941,
+ 509,
+ 954
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "table",
+ "img_path": "images/ce7830741f441d5ce070ca571bc7c8b76adf816c5b619ffa656b9c0daef6fea3.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Model | Model ID | URL |
| Qwen2.5-7B | Qwen2.5-7B-Instruct | https://huggingface.co/Qwen/Qwen2.5-7B-Instruct |
| Llama3.1-8B | Llama-3.1-8B-Instruct | https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct |
| Mistral-8B | Ministral-8B-Instruct-2410 | https://huggingface.co/mistralai/Ministral-8B-Instruct-2410 |
| Gemma2-9B | gemma-2-9b-it | https://huggingface.co/google/gemma-2-9b-it |
| Qwen2.5-14B | Qwen2.5-14B-Instruct | https://huggingface.co/Qwen/Qwen2.5-14B-Instruct |
| Qwen2.5-32B | Qwen2.5-32B-Instruct | https://huggingface.co/Qwen/Qwen2.5-32B-Instruct |
| Qwen2.5-72B | Qwen2.5-72B-Instruct | https://huggingface.co/Qwen/Qwen2.5-72B-Instruct |
| Llama3.3-70B | Llama-3.3-70B-Instruct | https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct |
| DeepSeek-v3 | DeepSeek-V3-0324 | https://huggingface.co/deepseek-ai/DeepSeek-V3-0324 |
",
+ "bbox": [
+ 178,
+ 140,
+ 821,
+ 266
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "table",
+ "img_path": "images/1a981bd1901d5cbde725a3f1aa0862ca5bcdcf4e957d236157126d10111daeb5.jpg",
+ "table_caption": [
+ "Table 11: Mapping of LLM abbreviations and IDs used in this paper, with their open-source URLs."
+ ],
+ "table_footnote": [],
+ "table_body": "| Model ID | Abbreviation | Base Model | Alg. | Size |
| DeepSeek |
| DeepSeek-R1-Distill-Qwen-1.5B | DS-R1-1.5B | Qwen2.5-Math-1.5B | SFT | 800K |
| DeepSeek-R1-Distill-Qwen-7B | DS-R1-7B | Qwen2.5-Math-7B | SFT | 800K |
| DeepSeek-R1-Distill-Llama-8B | DS-R1-8B | Llama-3.1-8B | SFT | 800K |
| DeepSeek-R1-Distill-Qwen-14B | DS-R1-14B | Qwen2.5-14B | SFT | 800K |
| DeepSeek-R1-Distill-Qwen-32B | DS-R1-32B | Qwen2.5-32B | SFT | 800K |
| DeepSeek-R1-Distill-Llama-70B | DS-R1-70B | Llama-3.3-70B-Instruct | SFT | 800K |
| DeepSeek-R1 | DS-R1 | DeepSeek-V3-0324 | SFT&RL | 800K&- |
| Qwen |
| QwQ-32B | QwQ-32B | Qwen2.5-32B | - | - |
| Qwen3-235B-A22B | Qwen3-A22B | Qwen3-235B-A22B-Base | SFT&RL | -&- |
| Qwen3-30B-A3B | Qwen3-A3B | Qwen3-30B-A3B-Base | SFT&RL | -&- |
| Qwen3-32B | Qwen3-32B | Qwen3-32B-Base | SFT&RL | -&- |
| Qwen3-14B | Qwen3-14B | Qwen3-14B-Base | SFT&RL | -&- |
| Qwen3-8B | Qwen3-8B | Qwen3-8B-Base | SFT&RL | -&- |
| Qwen3-1.7B | Qwen3-1.7B | Qwen3-1.7B-Base | SFT&RL | -&- |
| qihoo360 |
| Light-R1-7B-DS | L-R1-7B-DS | DeepSeek-R1-Distill-Qwen-7B | SFT | 3K |
| Light-R1-14B-DS | L-R1-14B-DS | DeepSeek-R1-Distill-Qwen-14B | SFT&RL | 3K&- |
| Light-R1-32B-DS | L-R1-32B-DS | DeepSeek-R1-Distill-Qwen-32B | SFT | 3K |
| Light-R1-32B | L-R1-32B | Qwen2.5-32B-Instruct | SFT&DPO | 73K&- |
| simplescaling |
| s1.1-7B | s1.1-7B | Qwen2.5-7B-Instruct | SFT | 1K |
| s1.1-14B | s1.1-14B | Qwen2.5-14B-Instruct | SFT | 1K |
| s1.1-32B | s1.1-32B | Qwen2.5-32B-Instruct | SFT | 1K |
| LG AI Research |
| EXAONE-Deep-2.4B | EXAONE-2.4B | EXAONE-3.5-2.4B-Instruct | SFT&DPO&RL | 1.6M&20K&10K |
| EXAONE-Deep-7.8B | EXAONE-7.8B | EXAONE-3.5-7.8B-Instruct | SFT&DPO&RL | 1.6M&20K&10K |
| EXAONE-Deep-32B | EXAONE-32B | EXAONE-3.5-32B-Instruct | SFT&DPO&RL | 1.6M&20K&10K |
| NVIDIA |
| Llama-3.1-Nemotron-Nano-8B-v1 | Nemotron-8B | Llama-3.1-8B-Instruct | SFT&RL | -&- |
| Llama-3.3-Nemotron-Super-49B-v1 | Nemotron-49B | Llama-3.3-70B-Instruct | SFT&RL | -&- |
| NovaSky |
| Sky-T1-32B-Flash | Sky-T1-32B | Qwen2.5-32B-Instruct | SFT&SimPO | 17K&10K |
",
+ "bbox": [
+ 132,
+ 416,
+ 863,
+ 834
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "text",
+ "text": "Table 12: The open-source LRMs details evaluated for S1-Bench.",
+ "bbox": [
+ 275,
+ 843,
+ 719,
+ 857
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "page_number",
+ "text": "20",
+ "bbox": [
+ 487,
+ 941,
+ 509,
+ 954
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "text",
+ "text": "Prompt for Correctness Evaluation",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 117,
+ 332,
+ 131
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "Evaluation on S1-Bench",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 140,
+ 263,
+ 152
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "**Question:**",
+ "bbox": [
+ 115,
+ 162,
+ 203,
+ 174
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "{question}",
+ "bbox": [
+ 115,
+ 177,
+ 179,
+ 191
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "**Model Answer:**",
+ "bbox": [
+ 115,
+ 202,
+ 236,
+ 212
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "{model_answer}",
+ "bbox": [
+ 115,
+ 217,
+ 213,
+ 229
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "**Ground Truth:**",
+ "bbox": [
+ 115,
+ 241,
+ 228,
+ 253
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "{groundtruth}",
+ "bbox": [
+ 115,
+ 256,
+ 205,
+ 269
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "Your task is to evaluate whether the model's answer is correct.",
+ "bbox": [
+ 115,
+ 281,
+ 473,
+ 293
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "An answer is considered correct as long as it contains the ground truth (regardless of how complex or detailed the description is).",
+ "bbox": [
+ 115,
+ 296,
+ 853,
+ 307
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "If there are parenthetical notes after the ground truth, then there may be multiple correct answers. In this case, the given answer is just one example, and any answer that meets the requirements specified in the notes can be considered correct.",
+ "bbox": [
+ 114,
+ 311,
+ 880,
+ 332
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "Additionally, some reasonably uncertain supplementary information is also considered appropriate, including more details, possibilities, and expanded discussion. You should focus more on whether the reply contains the correct answer.",
+ "bbox": [
+ 114,
+ 336,
+ 880,
+ 359
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "You need to output a standard JSON, providing your explanation of the evaluation in the \"explain\" field, and giving the evaluation result in the \"result\" field, where 1 means the answer is correct and 0 means it is incorrect.",
+ "bbox": [
+ 112,
+ 370,
+ 880,
+ 392
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "Your action should follow the given format: \"explain\": \"\", \"result\": 0/1",
+ "bbox": [
+ 115,
+ 395,
+ 519,
+ 407
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "table",
+ "img_path": "images/e102964c50a709822136bdbfe994f39d60087b6ab50cc89332a3509b36706e08.jpg",
+ "table_caption": [
+ "Table 13: Prompt for Correctness Evaluation."
+ ],
+ "table_footnote": [],
+ "table_body": "