Skeleton language choice affects multilingual math reasoning in large models
Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning
Computation and LanguageArtificial Intelligence
Summary
Language models solve math problems by following step-by-step guides called skeletons. This study looked at whether using skeletons in English or other languages helps these models solve math in different languages. They found English skeletons tend to work a little better on average, especially for smaller models and less common languages, but it is not always the best choice. The best skeleton language depends on factors like the model, the math problem, and how the answers are checked. So, choosing a skeleton language is not a one-size-fits-all decision.
What this means in practice
- •For multilingual ai developers: Adjust skeleton-language choices to improve reasoning performance in multilingual math applications for varied model sizes and languages.
- •For machine learning engineers: Test different skeleton-language prompts in model evaluation to find more effective reasoning strategies across benchmarks and languages.
Authors
HyeonSeok Lim, SeungWoo Song, Inho Won, Hoyun Song, Jihyo Kim, KyungTae Lim
Abstract
Skeleton-based reasoning prompting is a promising training-free approach for structuring LLM reasoning, but prior work largely assumes an English-centric setting. We propose the Language-Aware Skeleton Exploration Framework (LASEF) to study skeleton-language choice in multilingual mathematical reasoning. Across math benchmarks, model scales, and languages, we show that English skeletons yield a small positive tendency on average, most visible for smaller models and low-resource languages. However, few language-level gains remain significant after correction, and English is not universally optimal. Combining greedy decoding, multi-rollout evaluation, translation ablation, and cross-benchmark validation, we further find three patterns of skeleton-language effects: directionally consistent, evaluation- and benchmark-dependent, and asymmetric negative. These effects cannot be fully explained by generation quality alone. Overall, skeleton language is a context-dependent design variable that requires multi-level exploration. All resources are released at https://github.com/lhsstn/LASEF.