Crosslingual Reasoning through Test-Time Scaling

TL;DR

This work investigates whether English-centric reasoning finetuning with long chain-of-thoughts can generalize to multilingual reasoning via test-time scaling. Using s1 models trained on English reasoning data, the authors show that larger models (3B+) substantially benefit multilingual math reasoning, with 14B achieving notable gains and even outperforming models twice its size under certain conditions. They reveal a dominant quote-and-think language-mixing pattern, demonstrate that forcing reasoning in high-resource languages improves performance while forcing into low-resource languages often hurts, and show limited cross-domain generalization beyond STEM domains. The study provides practical guidance: rely on English-centric RLMs reasoning in high-resource languages for efficiency and reliability, pursue multilingual data and better tokenization to improve LRLs, and recognize the constraints of cross-domain transfer. Overall, the work establishes crosslingual test-time scaling as a strong multilingual baseline while outlining key mechanisms and limitations for future enhancement in multilingual reasoning systems.

Abstract

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.

Figures (9)

• Figure 1: Crosslingual test-time scaling of s1 and Qwen models on the MGSM benchmark (excluding English) across different model sizes. In subfigure (a) we enforce a hard limit of maximum thinking token, and in (b) we measure their inference FLOP compute for a Pareto frontier analysis. $\Delta$ measures the absolute difference between average accuracy scores at 0.5k and 8k maximum thinking tokens. Dash lines indicate the best few-shot prompting baseline performance of Qwen.
• Figure 2: Proportion of dominant languages in models' entire responses when queried with multilingual math questions. "same" indicates that the response language is the same as query language.
• Figure 3: Breakdown of language-mixing patterns in s1's reasoning. Percentage indicates the probability of a sentence being English only, quoting non-English phrases (quote-and-think), entirely being in a different language (intersentential), or mixing different languages within the same sentence (intrasentential).
• Figure 4: Language and domain breakdown for Global-MMLU benchmark. Dashed lines indicate the performance of zero-shot prompting of Qwen-32B-Instruct models.
• Figure 5: MGSM accuracy against number of thinking tokens in s1 models' outputs in different reasoning languages.