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

TL;DR

Objective: assess moral reasoning in LLMs across five typologically diverse languages and three context levels. Method: introduce MMRB with 2,170 scenarios spanning ETHICSBASE, ETHICSPRO, and ETHICSMAX, translated into English, Chinese, Russian, Vietnamese, and Indonesian; evaluate five LLMs with greedy decoding; perform fine-tuning of LLaMA-3-8B under Monolingual Alignment and Monolingual Poisoning. Contributions: (1) MMRB dataset and cross-language results showing inconsistencies; (2) analysis of monolingual alignment and poisoning effects, with Indonesian alignment yielding strongest cross-language gains and Vietnamese poisoning causing largest degradation; (3) insights into the role of low-resource languages in multilingual NLP. Significance: underscores data quality, fairness, and the need for language-informed safeguards in multilingual AI deployments.

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.

Figures (2)

• Figure 1: The performance of three levels datasets across five LLMs in a multilingual setting.
• Figure 2: The win-rate between five languages in ETHICSPRO dataset based on Llama-3-8B.