Moral Reasoning Across Languages: The Critical Role of Low-Resource Languages in LLMs

2025-04-28

AI Summary

This work evaluates the moral reasoning capabilities of large language models (LLMs) across multilingual settings—particularly for low-resource languages such as Vietnamese—across three levels of contextual complexity: sentence-, paragraph-, and document-level. To this end, we introduce MMRB, a typologically diverse, cross-lingual, multi-level moral reasoning benchmark covering five languages. Our study is the first to empirically demonstrate that low-resource languages exert a dominant influence on multilingual alignment, challenging the “high-resource centrism” assumption. Using LLaMA-3-8B, we conduct monolingual fine-tuning integrating moral alignment with controlled toxicity injection, enabling fine-grained cross-lingual analysis. Experiments reveal a pronounced decline in moral reasoning performance with increasing contextual complexity—most severe for low-resource languages. Notably, Vietnamese accuracy improves by 17.2% post-fine-tuning, underscoring the critical role of low-resource language modeling in enhancing global alignment robustness.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Knowledge Representation and Reasoning: Knowledge Representation Languages

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

In this paper, we introduce the Multilingual Moral Reasoning Benchmark (MMRB) to evaluate the moral reasoning abilities of large language models (LLMs) across five typologically diverse languages and three levels of contextual complexity: sentence, paragraph, and document. Our results show moral reasoning performance degrades with increasing context complexity, particularly for low-resource languages such as Vietnamese. We further fine-tune the open-source LLaMA-3-8B model using curated monolingual data for alignment and poisoning. Surprisingly, low-resource languages have a stronger impact on multilingual reasoning than high-resource ones, highlighting their critical role in multilingual NLP.

Problem

• Evaluating LLM moral reasoning across diverse languages and complexities
• Assessing performance degradation in low-resource languages like Vietnamese
• Exploring low-resource languages' impact on multilingual reasoning in NLP

Innovation

• Multilingual Moral Reasoning Benchmark (MMRB) for evaluation
• Fine-tuned LLaMA-3-8B with monolingual alignment data
• Low-resource languages boost multilingual reasoning performance