mCoT: Multilingual Instruction Tuning for Reasoning Consistency in Language Models

TL;DR

The paper investigates multilingual reasoning consistency in LLMs and shows that cross-language CoT reasoning can be improved via instruction tuning. It introduces mCoT-MATH, a large-scale multilingual dataset (~6.3M samples across 11 languages) built by translating English math-CoT data, and trains a 7B model, mCoT, that achieves strong cross-language consistency and competitive performance against larger models. Through MGSM and MSVAMP benchmarks, mCoT demonstrates substantial gains in both final-answer accuracy and reasoning alignment across languages, especially for underrepresented languages. The work provides practical resources and methodology for enhancing multilingual reasoning, reducing language-resource disparities, and guiding future cross-lingual reasoning research.

Abstract

Large language models (LLMs) with Chain-of-thought (CoT) have recently emerged as a powerful technique for eliciting reasoning to improve various downstream tasks. As most research mainly focuses on English, with few explorations in a multilingual context, the question of how reliable this reasoning capability is in different languages is still open. To address it directly, we study multilingual reasoning consistency across multiple languages, using popular open-source LLMs. First, we compile the first large-scale multilingual math reasoning dataset, mCoT-MATH, covering eleven diverse languages. Then, we introduce multilingual CoT instruction tuning to boost reasoning capability across languages, thereby improving model consistency. While existing LLMs show substantial variation across the languages we consider, and especially low performance for lesser resourced languages, our 7B parameter model mCoT achieves impressive consistency across languages, and superior or comparable performance to close- and open-source models even of much larger sizes.

Figures (6)

• Figure 1: Overview of multilingual reasoning; LLMs are expected to have consistent reasoning capabilities across different languages when given the same problem which has the same answer. Shown in picture are three example languages: English (EN), Swahili (SW), and Chinese (ZH). For EN, we show the problem formulation, and the Chain-of-Thought (CoT) reasoning.
• Figure 2: Accuracy (%) on MGSM of different models with the few-shot method. All machine-translated prompts are translated from English data using Google Translate.
• Figure 3: Multilingual reasoning consistency. The triangle above the marked diagonal shows the consistency of the models on the correct answers; the triangle below the diagonal contains the consistency between the language pairs where the final answer is the same but incorrect.
• Figure 4: Overview of multilingual CoT reasoning data. English data is first automatically translated into target languages, and then inserted into the templates to construct multilingual instruction data.
• Figure 5: Multilingual reasoning consistency of mCoT. The triangle above the marked diagonal shows the consistency of the models on the correct answers; the triangle below the diagonal contains the consistency between the language pairs where the final answer is the same but incorrect.