mAceReason-Math: Multilingual Math Dataset

• mAceReason-Math is a multilingual dataset of advanced math problems, curated from an English corpus and translated into 14 languages with rigorous cleaning.
• It employs a hybrid LLM and native-speaker translation pipeline, ensuring preservation of mathematical accuracy and invariant LaTeX formatting.
• The dataset supports RLVR experiments by providing scalable, verifiable problem instances through parallel and full-train splits optimized for policy training.

Searching arXiv for the specified paper and closely related work to ground the article in current literature. mAceReason-Math is a multilingual dataset of hard mathematics problems designed specifically to make multilingual research in reinforcement learning from verifiable rewards (RLVR) practical. It is presented as a multilingual extension of AceReason-Math rather than a new problem source: a carefully filtered, cleaned, and translated subset of an English corpus already curated for RL-based post-training on advanced mathematical reasoning. Its defining properties are the same ones emphasized in RLVR-oriented math training more broadly—difficulty calibrated for contemporary models, scale sufficient for policy optimization, and final answers that remain automatically verifiable across languages. The release covers 14 languages, provides over 140k translated problems, and includes both per-language full-train splits and a fully parallel subset for controlled multilingual experiments (Dobler et al., 11 Mar 2026).

1. Position within multilingual RLVR research

The immediate motivation for mAceReason-Math is the English-centric structure of current RLVR research. Recent mathematical-reasoning work using methods such as GRPO has relied heavily on English-only corpora, including AceReason-Math, DAPO-Math-17k, and POLARIS-53k, while existing multilingual math datasets have generally been either much easier or much smaller. The paper identifies an “English bottleneck”: multilingual mathematical reasoning may lag not because of insufficient model capacity, but because multilingual data with the right RLVR properties are missing (Dobler et al., 11 Mar 2026).

The dataset is explicitly aligned with a “goldilocks zone” of task difficulty. If problems are too easy, rewards saturate and training signal weakens; if they are too hard, rewards become too sparse. For this reason, the dataset inherits from AceReason-Math rather than from easier school-math corpora. This design is congruent with math-RL recipes that rely on binary verifier-derived rewards, strict final-answer extraction, and hard verifiable prompts rather than open-ended proof scoring (Chen et al., 22 May 2025).

Relative to prior multilingual resources, mAceReason-Math is differentiated by scale and intended use. The authors note that harder multilingual benchmarks such as PolyMath and MMATH have fewer than 600 samples per language, which makes them useful for evaluation but not for multilingual GRPO or RLVR training. By contrast, mAceReason-Math supplies more than 10,000 samples per language and a fully parallel subset, thereby targeting multilingual training and multilingual evaluation simultaneously (Dobler et al., 11 Mar 2026).

2. English-source curation and corruption handling

The source corpus is AceReason-Math, but the English base was not translated directly without intervention. The authors first performed a cleaning pass because they found a noticeable amount of corruption and noise in the original samples. They separate these issues into “critical issues,” which cause exclusion, and “salvageable issues,” which are repaired before translation. Critical issues affect roughly 4% of the considered source data (Dobler et al., 11 Mar 2026).

Critical issues include problems that require missing diagrams or external context; items containing URLs, image placeholders such as [IMAGE], or references such as “Figure” and “diagram” when the referenced content is unavailable; questions that already contain the solution; malformed open-ended prompts such as “How can you …” paired with a single short answer; non-English source items; and other severely corrupted or incoherent examples. Programmatically, custom regular expressions were used to remove image references, URLs, and answer-format artifacts such as occurrences of \boxed{[...]} that often remained from earlier formatting instructions. Less structured errors were handled through an LLM-based cleaning stage that classified corruption and, when possible, returned cleaned versions (Dobler et al., 11 Mar 2026).

Salvageable issues were more common, at about 11% of the original samples. These included task annotations such as “Task 5.4,” source labels such as olympiad headers, formatting debris, extraction and LaTeX errors that were obvious from context, and spurious LLM instructions accidentally included in the text. For these cases, the authors used a cleaning prompt with Claude Sonnet 4 to normalize the English text before translation, under a conservative rule that artifacts could be removed only if doing so would not alter the expected solution or mathematical context. Problems referring to diagrams encoded in [asy] were retained and translated in a separate split rather than discarded outright, because they remained solvable but required specialized ability to parse Asymptote-style diagram code (Dobler et al., 11 Mar 2026).

3. Translation pipeline and quality assurance

The translation pipeline is hybrid: LLM translation at scale combined with native-speaker validation. The initial translation model was Claude Sonnet 4. A pilot study translated 100 randomly sampled English problems and submitted them for native-speaker review in German, French, Italian, Spanish, Chinese, Korean, Japanese, Thai, Russian, and Brazilian Portuguese. Swahili, Telugu, and Bengali were excluded from human validation in this pilot because annotators were unavailable, although translations into those languages were still produced. In the pilot, 97% of translations were rated either “acceptable” or “excellent,” and the main issues were often LaTeX mismatches rather than gross semantic mistranslation (Dobler et al., 11 Mar 2026).

The full translation run used an iterative refinement loop, again with Claude Sonnet 4. The translation prompt imposed unusually strong invariance constraints: preservation of all mathematical concepts, variables, equations, and logical relations; prohibition on translating LaTeX inside $...$ or $...$ except natural language inside \text{...}; exact preservation of LaTeX notation style and even spacing; translation into natural target-language mathematical prose rather than literal word-for-word text; language-specific number formatting conventions; and Arabic numerals for answers in all languages. This produced a cross-lingual invariant in which the mathematical object stays fixed while the surrounding prose is localized. This suggests that the dataset is engineered to minimize reward distortions caused by translation artifacts rather than reasoning errors (Dobler et al., 11 Mar 2026).

After initial translation, each sample was automatically graded by another prompt modeled on the human annotation guidelines. The grading step checked preservation of mathematical accuracy, exact LaTeX matching, correct answer handling, fluency, and proper mathematical terminology. If problems were detected, the translation was revised with a separate improvement prompt. The loop ran for up to five iterations, and translations not accepted by then were discarded. The paper notes that most samples were accepted immediately or after one improvement round (Dobler et al., 11 Mar 2026).

The test set received a further native-speaker review pass with rendered mathematics via MathJax, so annotators did not need to inspect raw LaTeX. The authors filtered annotator suggestions selectively with Claude Sonnet 4 and manually checked all items marked problematic. After this review, fewer than 0.1% of individual translations in the validated test set were still rated problematic—two translations total, one in French and one in Brazilian Portuguese. In addition, stylistic or naturalness-only post-edits were applied to 36% of samples (Dobler et al., 11 Mar 2026).

4. Dataset composition and split structure

mAceReason-Math covers 14 languages: English, German, French, Spanish, Chinese, Russian, Japanese, Swahili, Telugu, Bengali, Thai, Portuguese, Italian, and Korean. The final compilation has three principal split types. The fully parallel train split contains 7,620 samples per language, with every problem present in all supported languages. The human-validated parallel test set contains 190 samples per language and is removed from training. The larger non-parallel full-train splits contain all accepted translations for each language and range from 10,270 to 12,245 samples per language (Dobler et al., 11 Mar 2026).

The exact full-train totals are as follows.

Language	Full train total
English	12,245
German	11,151
French	11,007
Spanish	11,346
Chinese	10,470
Russian	11,237
Japanese	10,376
Swahili	11,124
Telugu	10,964
Bengali	11,082
Thai	11,104
Portuguese	10,632
Italian	10,646
Korean	10,270

The paper describes the resource at a high level as “over 140k high-quality translations” and “10k+ translated problems across 13 languages,” plus cleaned English originals. Because the English source is also included in cleaned form, the dataset can be used both monolingually and multilingualy. The paper does not provide a fine-grained internal taxonomy of mathematical subdomains or difficulty labels, nor an explicit answer-type ontology. It instead defines structure through split membership and implicit parallel alignment, while examples and the verification setup indicate predominantly short-form verifiable answers, often numeric or symbolic expressions suitable for automated checking (Dobler et al., 11 Mar 2026).

5. RLVR readiness and verifier interface

The dataset’s claim to be “ready for RLVR” is procedural rather than algorithmic. The paper introduces no new reward function, RL objective, or verifier formula. Its contribution is infrastructural: the problems are difficult enough to provide useful reward signals for modern models, and their final answers can be checked automatically. During evaluation, models are instructed to provide the final answer inside <answer>...</answer> tags, with this instruction translated into the prompt language. If those tags are absent, the system attempts fallback extraction from \boxed{...}. The extracted answer is then checked against the original ground truth using the math-verify toolkit from Hugging Face, and string equality against the translated answer is also checked to accommodate language-specific formatting differences such as decimal commas (Dobler et al., 11 Mar 2026).

This answer-extraction-and-verification logic is narrower than the post-training and reward-modeling systems developed in adjacent math-reasoning work. AceMath, for example, studies outcome reward models, Bradley-Terry ranking losses, and rm@8 reranking over multiple candidate solutions, whereas mAceReason-Math focuses on supplying multilingual, hard, answer-verifiable problem instances rather than a new verifier or reranker (Liu et al., 2024). The two are complementary: one contributes multilingual problem infrastructure; the other contributes reward modeling and inference-time selection.

The appendix makes the verifier assumptions more concrete. The translation prompt standardizes answers to Arabic numerals, preserves formulas and notation exactly, and localizes non-mathematical text without changing problem logic. The grading prompt enforces exact LaTeX consistency and answer handling. In effect, the dataset aims to preserve a cross-lingual invariant under which the reward reflects reasoning ability rather than translation noise. A plausible implication is that multilingual RLVR experiments on this corpus can isolate language effects more cleanly than experiments on loosely translated or nonparallel benchmarks (Dobler et al., 11 Mar 2026).

6. Benchmarking results and cross-lingual behavior

The experiments use mAceReason-Math primarily as a benchmark on the 190-sample human-validated parallel test set. Closed models are evaluated with pass@1 from a single response, while open-weight models are evaluated with avg@8, the average accuracy over eight sampled responses per prompt. This asymmetry is intended to reduce variance under stochastic decoding for open models (Dobler et al., 11 Mar 2026).

Among closed models in the main table, Gemini 2.5 Flash is strongest overall with an average of 81.4% ± 2.2 across languages and is described as notably strong in multilingual consistency. Claude Sonnet 4.5 averages 75.7% ± 2.7. Among open models in the main table, gpt-oss-20b is the most balanced with 76.3% ± 3.4 avg@8. DeepSeek-R1-Distill-7B reaches 83.3 on English and 80.4 on Chinese but falls off substantially in several other languages, averaging 64.3% ± 13.3, and smaller models show much larger cross-lingual degradation, especially on Bengali, Telugu, and Swahili (Dobler et al., 11 Mar 2026).

The appendix reports especially strong results for the Qwen3 family. Qwen3-14B achieves an average of 88.6% ± 5.2 across languages, with scores around or above 87–92 on nearly all languages except Swahili, where it reaches 69.7. Qwen3-8B and Qwen3-4B are also described as very strong. The authors speculate that these results may indicate the data were present in the Qwen3 reasoning models’ RLVR training mix. That caveat matters for benchmark interpretation, but it does not eliminate the dataset’s value as a controlled multilingual testbed (Dobler et al., 11 Mar 2026).

One of the paper’s central empirical implications is that strong models can transfer math-problem-solving ability across translated instances of the same underlying problems. The benchmark therefore permits a more precise decomposition of multilingual mathematical performance into a language component and a reasoning component. This suggests that some apparent multilingual deficits in math may reflect training-data asymmetry more than intrinsic inability to reason in non-English settings (Dobler et al., 11 Mar 2026).

7. Limitations, scientific uses, and relation to adjacent systems

The paper is explicit about limitations. Because translation and refinement are LLM-based, residual errors may remain. Human validation was unavailable for Swahili, Telugu, and Bengali, so translation quality may be more variable in those languages. Only the test set received exhaustive final native-speaker review; the larger train splits did not. Automated answer verification can also be brittle, particularly for more complex expressions or formatting variations, which may generate false negatives. The authors also note that exceptionally high benchmark performance by some models may reflect contamination or inclusion of the data in prior training (Dobler et al., 11 Mar 2026).

Within the broader ecosystem, mAceReason-Math is best understood as a dataset-layer contribution rather than a training-method or architecture paper. AceReason-Nemotron studies strict on-policy GRPO, progressive response-length curricula, difficulty filtering, and verifier-based math RL on a large hard-verifiable math corpus (Chen et al., 22 May 2025). mAceReason-Math extends the multilingual data substrate needed for comparable experiments across languages, but does not specify a separate “mAceReason-Math” model line, RL schedule, or reward equation. In that sense, it is enabling infrastructure for multilingual RLVR rather than an end-to-end multilingual reasoning system.

The dataset is usable in several distinct experimental regimes. For RLVR training, the parallel split supports matched multilingual instances for language-balanced policy optimization and transfer studies. For monolingual or mixed-language fine-tuning, the larger per-language full-train splits provide substantially more scale than prior multilingual math benchmarks. For evaluation, the 190-sample human-validated parallel test set offers a controlled robustness benchmark in which mathematical content is held constant across languages. Taken together, these properties make mAceReason-Math a mechanism for replacing the English-only default in verifier-based math post-training with genuinely multilingual experimentation (Dobler et al., 11 Mar 2026).