Distinguishing Input Language from Reasoning Language: A Diagnostic Framework for Cross-Lingual Moral Alignment in Large Language Models

The paper proposes a diagnostic framework that separately manipulates the input language of moral dilemmas and the model's reasoning language, analyzing English-Chinese cross-lingual differences in moral judgment across 13 large language models. The framework extends standard evaluations, which test only matched conditions (e.g., an English dilemma with English reasoning), to mismatched conditions, and interprets changes in judgment content according to moral foundations theory. Results show that the variance caused by reasoning language is approximately twice that caused by input language, and that standard evaluations miss the context dependence of nearly half of the models. The authors accordingly provide a deployment-oriented diagnostic taxonomy and release code and datasets.