PaperlayerView exact revision on arXiv

Source-linked AI summary

Crosslingual Reasoning through Test-Time Scaling

Zheng-Xin Yong, M. Farid Adilazuarda, Jonibek Mansurov, Ruochen Zhang, Niklas Muennighoff, Carsten Eickhoff, Genta Indra Winata, Julia Kreutzer, Stephen H. Bach, Alham Fikri Aji

arXiv:2505.05408v1cs.CLcs.AIcs.LG

TL;DR

The paper asks whether English-centric reasoning finetuning transfers across languages and studies test-time scaling and reasoning-language control. It finds gains for multilingual math reasoning, but limited generalization to out-of-domain tasks and low-resource languages.

  • Problem

    Reasoning and test-time scaling have been studied mainly in English, leaving open whether English-centric RLMs built from multilingual models can reason across languages.

    Source: §1 · p. 1

  • Method

    The paper evaluates English-centric RLMs under scaled inference budgets, examines their language mixing, and tests control of the reasoning language.

    Source: §1 · p. 2

  • Results

    Test-time scaling improves multilingual math reasoning, while scaling does not consistently help non-STEM tasks and can hurt cultural-commonsense performance.

    Source: §1 · p. 2, §9 · p. 13

  • Takeaways & Limitations

    The authors recommend high-resource reasoning languages for English-centric RLMs and call for further work on low-resource languages and out-of-domain tasks.

    Source: §1 · p. 2, §9 · p. 13

    Reasoning remains poor in low-resource languages, where tokenization disparities also increase inference costs.

    Source: §8 · p. 13

Abstract

from arXiv · show

Reasoning capabilities of large language models are primarily studied for English, even when pretrained models are multilingual. In this work, we investigate to what extent English reasoning finetuning with long chain-of-thoughts (CoTs) can generalize across languages. First, we find that scaling up inference compute for English-centric reasoning language models (RLMs) improves multilingual mathematical reasoning across many languages including low-resource languages, to an extent where they outperform models twice their size. Second, we reveal that while English-centric RLM's CoTs are naturally predominantly English, they consistently follow a quote-and-think pattern to reason about quoted non-English inputs. Third, we discover an effective strategy to control the language of long CoT reasoning, and we observe that models reason better and more efficiently in high-resource languages. Finally, we observe poor out-of-domain reasoning generalization, in particular from STEM to cultural commonsense knowledge, even for English. Overall, we demonstrate the potentials, study the mechanisms and outline the limitations of crosslingual generalization of English reasoning test-time scaling. We conclude that practitioners should let English-centric RLMs reason in high-resource languages, while further work is needed to improve reasoning in low-resource languages and out-of-domain contexts.

1 Introduction

The paper asks whether English-centric reasoning finetuning and test-time scaling transfer to multilingual reasoning, how models mix languages, and whether transfer extends beyond STEM. It reports benefits across languages and reasoning-language choices, but limited out-of-domain gains.

Contributions

  • Scaling test-time compute improves multilingual math reasoning, with benefits for high- and low-resource languages and performance exceeding models twice the size.The authors recommend data-efficient English reasoning finetuning to preserve multilingual capability and high-resource reasoning languages for performance and compute efficiency.Source: §1 · p. 2
  • English-centric RLMs often quote non-English phrases and reason about them, a pattern the authors call “quote-and-think.”They suggest that multilingual parsing and understanding support this crosslingual generalization.Source: §1 · p. 2
  • Forcing reasoning in high-resource languages yields better performance than in low-resource languages, with more token-efficient long CoTs.The paper identifies language control as an effective strategy.Source: §1 · p. 2
  • Scaling thinking tokens brings no consistent benefit on non-STEM tasks, and can hurt performance on tasks requiring cultural commonsense knowledge.The paper calls for further work on out-of-domain tasks and low-resource-language reasoning.Source: §1 · p. 2

2 Background and Related Work

The paper situates its work in long-CoT reasoning and test-time scaling, then focuses on how inference length and reasoning language affect multilingual reasoning.

Reasoning language models (RLMs)

  • RLMs use intermediate chain-of-thought reasoning, and longer chains can support backtracking and self-correction.Long-CoT RLMs are often made by distilling English reasoning chains into multilingual pretrained models.Source: §2 · p. 2

Test-time scaling and s1

  • Test-time scaling allocates more computation at inference; the s1 recipe combines small-scale reasoning finetuning with a larger inference budget.Prior test-time scaling work has mostly evaluated English math benchmarks.Source: §2 · p. 3

Multilingual reasoning

  • Prior multilingual reasoning work explores English as a pivot, query translation, representation or output alignment, and broader reasoning-data coverage.This paper studies how long-CoT length and reasoning language at test time affect multilingual reasoning; it contrasts its focus with negative results from 1.5B models [17].Source: §2 · p. 3

3 Experimental Setup

The experiments use English-finetuned s1 models, budget-forcing strategies, multilingual math and cross-domain benchmarks, and Qwen-based baselines.

Models

  • The study uses multilingual Qwen2.5-Instruct s1.1 models finetuned on 1k English-only reasoning samples, at sizes from 1.5B to 32B parameters.The models and training data are open-sourced, and the study selects them for their English math performance and accessibility.Source: §3 · p. 3

Budget forcing

  • Budget forcing controls long-CoT inference budgets through truncation or extrapolation, which appends “Wait” to prompt continued reasoning.The extrapolation setup adds “Wait” once because further lengthening did not yield significant gains.Source: §3 · p. 3

Evaluation data

  • MGSM provides 250 grade-school math problems manually translated into ten languages, with no observed 8-gram overlap with s1 training samples.It is used for RQ1–RQ3.Source: §3 · p. 3
  • For cross-domain RQ4, the study uses Global-MMLU, English-language FORK food-culture questions, and Indonesian COPAL-ID causal-reasoning questions.Global-MMLU spans multiple subjects and marks whether questions are culturally agnostic or specific.Source: §3 · pp. 3–4

Baselines

  • Qwen2.5-Instruct serves as the main baseline, with zero-shot and few-shot CoT comparisons on MGSM and zero-shot comparisons on cross-domain benchmarks.MGSM comparisons also include prior state-of-the-art models and R1-distilled-Qwen; few-shot CoT is prompted in English or the input language.Source: §3 · p. 4

4 Crosslingual Test-Time Scaling

On multilingual MGSM, the paper finds that test-time scaling helps sufficiently large s1 models and can outperform prompting and larger competing models. Its analyses also identify a model-capacity threshold and a 14B accuracy–compute trade-off.

4 Crosslingual Test-Time Scaling

  • The study examines zero-shot crosslingual scaling of English-centric models finetuned from multilingual pretrained models in the same domain.The figure compares model sizes under a maximum thinking-token limit and in an inference-FLOP Pareto analysis.Source: §4 · p. 4

4.1 Effectiveness of Crosslingual Test-Time Scaling

  • With high inference budgets, s1 beats Qwen few-shot prompting across non-English MGSM languages; scaling is effective at 3B parameters and above.At 14B, the accuracy gain is +∆9.4% from 0.5k to 8k maximum thinking tokens; 1.5B gains only +∆1.8% and peaks at 4,000 tokens.Source: §4.1 · pp. 4–5
  • On MGSM’s Pareto frontier, the 14B s1 model achieves above 80% accuracy with substantially lower inference FLOPs than 32B; 3B and 7B cannot match larger models’ accuracy ceilings, even at maximum performance.Source: §4.1 · p. 5

4.2 Performance Comparison on MGSM Benchmark

  • Crosslingual test-time scaling improves both high- and low-resource languages: French gains +∆23.1% relative accuracy and Swahili +∆41.6%.Truncation and extrapolation produce similar performance because s1 already generates extensive reasoning chains.Source: §4.2 · p. 5
  • Scaled s1 outperforms prior multilingual-finetuned state-of-the-art models on MGSM, including MetaMath, MAPO, and models cited as [52] [49] [35].The authors suggest these prior models’ Llama bases generate shorter reasoning traces and lack some verification and backtracking behaviors.Source: §4.2 · p. 6
  • 14B s1 can outperform the twice-larger R1-Distill-Qwen-32B and Gemma-3-27B-it on MGSM.The authors associate s1’s minimal forgetting and better crosslingual generalization with its finetuning on only 1k English samples for 5 epochs.Source: Table 1 · p. 5, §4.2 · p. 6
  • The reported two-times-larger-model comparison is scoped to MGSM; Qwen3 is reported as the most performant model, with the authors attributing this probably to long reasoning and extensive multilingual training data.Gemma-3 outperforms s1 on low-resource languages, while its gap on high-resource languages may reflect shorter reasoning time.Source: §4.2 · p. 6

5 Language-Mixing Behaviors

English-centric s1 shifts multilingual responses toward English, yet its reasoning usually quotes and interprets non-English input rather than switching languages wholesale. This pattern suggests that multilingual understanding supports crosslingual reasoning.

5.1 Dominant Language in Model Outputs

  • In multilingual math settings, s1 makes English the dominant response language, while Qwen generally responds in the query language.This shift appears even though s1 was finetuned on only 1k English reasoning samples.Source: §5.1 · p. 6, Figure 2 · p. 7

5.2 Language-Mixing Patterns During Reasoning

  • At least 92.5% of s1’s reasoning sentences are English-only; when languages mix, quote-and-think is the main pattern.The model quotes non-English phrases, often from the question, then interprets their meaning in its reasoning.Source: Figure 3 · p. 7, §5.2 · p. 7
  • Russian has the most sentence-level switching, while at least 70% of within-sentence mixing cases resemble quote-and-think.In many of those cases, non-English phrases are drawn from the question without quotation marks.Source: Figure 3 · p. 7, §5.2 · p. 7
  • Quote-and-think goes beyond translation: s1 uses Japanese syntax and semantics to reason about mathematical relationships in the question.The analysis suggests the base model’s multilingual language understanding is preserved.Source: §5.2 · p. 7

6 Language Forcing

The paper tests ways to force s1 to reason in specified languages and compares their accuracy, language compliance, and inference cost. High-resource languages generally preserve strong performance, whereas strict low-resource reasoning can underperform and language control can trade off against accuracy.

6.2 In-Language Forcing

  • High-resource reasoning languages perform near the English baseline, but strict in-language forcing underperforms for low-resource languages.Strategies that allow English and the target language to mix can outperform strict in-language approaches for low-resource languages.Source: Table 2 · p. 8, §6.2 · p. 9
  • Near-100% language compliance from the combined strategy comes with the worst accuracy for 14B- and 32B-sized models.For 7B models, the combined strategy is needed to achieve high language compliance.Source: Table 2 · p. 8, §6.2 · p. 9

6.1 Methodology: Language Forcing Techniques

  • Forcing combines a translated “Wait” after English reasoning, a translated prefix, a translated system prompt, or all three.The study measures task accuracy and the share of generated tokens in the intended reasoning language.Source: §6.1 · p. 8

6.3 Crosslingual Language Forcing

  • Across query–reasoning language pairs, English, French, and German reasoning score similarly, with English highest and French close behind.The best reasoning language is not necessarily the query language; some query languages can perform best with reasoning in another language.Source: Table 3 · p. 9, §6.3 · p. 10
  • Translating a Swahili question into French raises accuracy from 59.6 to 90.8, even when reasoning remains in French.Performance varies more with reasoning language for low-resource query languages than for high-resource ones.Source: Table 3 · p. 9, §6.3 · p. 11
  • Token count and math accuracy have a -0.811 correlation; Swahili reasoning uses about 3.5 times the compute of French.The authors consider low-resource-language tokenization disparity a likely reason for the cost gap.Source: Figure 5 · p. 10, §6.3 · p. 10

7 Cross-Domain Generalization

Cross-domain evaluation finds that STEM benefits from test-time scaling, while non-STEM and cultural-commonsense tasks show little or inconsistent benefit. Scaling can even reduce performance in medicine and on English cultural reasoning benchmarks.

7.1 In-domain generalization

  • In STEM subjects, scaling improves Global-MMLU performance; s1 gains +∆19.8% for Bengali and +∆13.1% for Swahili over Qwen2.5-32B-Instruct.The authors connect these results to s1 training data that includes OlympicArena STEM knowledge.Source: §7.1 · p. 11

7.2 Out-of-domain generalization

  • Medicine gains only +∆0.8% (73.0% →73.8%) at 4k tokens and falls −∆2.0% (73.0% →71.0%) at 8k; business gains +∆3.2% versus STEM’s +∆11.5%.Across non-STEM domains, business benefits most, but its gain remains far below STEM’s.Source: §7.2 · pp. 11–12
  • For cultural-commonsense benchmarks, scaling provides minimal benefit and more thinking tokens substantially worsen English FORK performance.The paper describes this worsening as overthinking, where long reasoning expends excessive compute and leads to worse performance.Source: §7.2 · p. 12
  • Global-MMLU shows no clear culture-specific versus culture-agnostic advantage from s1 training, and language forcing yields minimal, often inconsistent gains.Humanities performance drops mostly affect culture-specific questions, while medical-domain drops mostly affect culture-agnostic questions.Source: §7.2 · p. 12, Figure 7 · p. 12

8 Discussion and Future Work

The discussion weighs data-efficient English reasoning finetuning and broader multilingual training data against limits in low-resource languages and small reasoning models.

Data-efficient English reasoning finetuning

  • With 1k English samples over five epochs, s1 retains multilingual text generation with minimal catastrophic forgetting, unlike R1-Distill-Qwen’s marked low-resource-language forgetting.The authors advise data-efficient English-centric finetuning to preserve multilingual capability.Source: §8 · p. 12

Reasoning finetuning with multilingual data

  • Multilingual reasoning data with broad language and domain coverage is a possible response, but translation-based augmentation remains limited by cultural alignment, Western-centric bias, and low-resource translation quality.The authors call for systematic study of back-translation and synthetic-data generation [70] [71].Source: §8 · pp. 12–13

Reasoning in low-resource languages (LRLs)

  • Low-resource-language reasoning performs poorly, and tokenization disparities raise test-time inference costs enough to create accessibility barriers.The authors call for more equitable tokenization strategies for reasoning across languages.Source: §8 · p. 13

Crosslingual generalization for small RLMs

  • Crosslingual test-time scaling succeeds at the 3B-parameter threshold, while smaller RLMs struggle to learn from long CoTs and generalize multilingually.The authors recommend evaluating small RLMs on multilingual reasoning, not only English tasks [75] [17].Source: §8 · p. 13

9 Conclusion

The conclusion summarizes English-centric reasoning models’ multilingual test-time scaling, language-mixing patterns, and reasoning-language choices, while emphasizing limited cross-domain generalization. The appendices credit contributors and provide further MGSM comparisons across model sizes.

9 Conclusion

  • The study documents quote-and-think language mixing and examines reasoning-language choice, alongside crosslingual math test-time scaling and limited cross-domain generalization.The conclusion highlights the need for more robust, domain-agnostic reasoning strategies for multilingual applications.Source: §9 · p. 13

A Authors’ Contributions

  • Zheng-Xin Yong led the project and most experiments; other contributors handled language forcing, GlobalMMLU cross-domain work, model training, advising, and writing.The passage credits Muhammad Farid Adilazuarda and Jonibek Mansurov with language forcing, Ruochen Zhang with GlobalMMLU, and Niklas Muennighoff with training different model sizes.Source: A Authors’ Contributions · p. 22

B Further Details on Crosslingual Test-Time Scaling

  • Tables 4–6 compare MGSM performance for 32B-, 7B-, and 3B-sized s1 models, each with a maximum of 8k thinking tokens.The table captions specify the model size and maximum thinking-token budget.Source: Table 4 · p. 22, Table 5 · p. 22, Table 6 · p. 22

C.1 Methodology

The appendix describes how the authors identify dominant languages and annotate language mixing in model responses. It also presents dominant-language distributions for four queried languages without specifying the distribution outcomes in the supplied passages.

C.1 Methodology

  • To filter language-mixed sentences, the authors identify each response’s dominant, or matrix, language using lingua.They focus on Japanese, Russian, Thai, and Mandarin Chinese because stanza and lingua support them, avoiding Latin-script languages that may share vocabulary with English.Source: §C.1 · pp. 22–23
  • Table 7 compares MGSM performance for the 1.5B s1 model with an 8k-token maximum; extrapolation budget forcing was omitted because its CoTs were already extremely long.Source: Table 7 · p. 23

C.2 Annotation for Language-Mixing Patterns in s1 Reasoning

  • The annotation distinguishes quoted foreign words within a sentence as quote-and-think, mixed-language sentences without quotes as intrasentential, and fully non-English sentences as intersentential.The procedure assigns the highest-probability sentence language, checks word-level language labels, and uses quotation marks to distinguish quote-and-think from intrasentential mixing.Source: §C.1 · p. 23, §C.2 · p. 23

C.3 Dominant Language in 14B-Sized Model Outputs

  • Figure 8 presents dominant-language proportions for 14B models responding to Japanese, Russian, Thai, and Mandarin Chinese questions; “same” means the output matches the query language.The supplied text identifies the figure’s scope and encoding but does not state its distribution results.Source: Figure 8 · p. 23, §C.3 · p. 23

C.4 Quote-And-Think Pattern in s1’s Training Data

The appendix traces quote-and-think behavior to patterns in s1’s English training data and distinguishes it from other forms of intrasentential language mixing. A forced Japanese reasoning example retains quote-and-think, but quotes English phrases instead.

C.4 Quote-And-Think Pattern in s1’s Training Data

  • 68.3% of s1’s English training samples exhibit quote-and-think, and at least half of those copy phrases directly from question prompts.This pattern is consistent with transfer of the model’s learned prompt-quoting behavior into long CoTs.Source: §C.4 · pp. 23–24

C.5 Fine-Grained Analysis of s1-32B’s Intrasentential Language Mixing

  • Extract-and-explain is the dominant annotated mixing pattern, alongside insertional and clause-level code-switching.Extract-and-explain reuses prompt phrases without quotation marks; insertional switching places non-English words in an English sentence, while clause-level switching changes languages between clauses.Source: §C.5 · p. 24, Figure 9 · p. 24

D.1 Example of CoTs after Language Forcing

  • Under combined language forcing, a Japanese-dominant CoT still quotes English phrases, including a translated question prompt.This contrasts with natural CoTs, which the appendix presents as the comparison case.Source: §D.1 · p. 24

E.1 Evaluation Details

For GlobalMMLU, the authors test reasoning in the question’s language and find inconsistent effects: high-resource Latin-script languages show marginal gains, while lower-resource languages decline. They also describe how they parse generated answers for evaluation.

E.1 Evaluation Details

  • For GlobalMMLU, GPT-4o-mini extracts final answer choices from variable-format reasoning traces using a two-token output limit.The extraction checks for one of the four answer options.Source: §E.1 · p. 25

E.2 Language Forcing for GlobalMMLU

  • The GlobalMMLU experiment applies combined language forcing to test whether reasoning in the question’s language helps performance.The strategy was selected because it yielded the highest language compliance in multilingual generations.Source: §E.2 · p. 26
  • About 6% average accuracy degradation across domains accompanies language forcing for Bengali and Swahili, while high-resource Latin-script languages show marginal gains.Effects are inconsistent even within German: business improves by around 3%, whereas humanities falls by nearly 5%.Source: §E.2 · p. 26, Table 13 · p. 29
Loading 2505.05408v1…