Do Morals Guide How LLMs Think? The Role of Ethical Perspectives in General Problem Solving

This study investigates how different moral conditions influence the general problem-solving capabilities of Large Language Models (LLMs). We examine whether the role of morality in human decision-making can also serve as a useful lens for analyzing variation in LLM behavior. Specifically, we define distinct moral conditions based on Kohlberg's theory of moral development and design prompts intended to elicit model outputs aligned with each condition. The validity of this alignment is assessed using the Defining Issues Test, a human evaluation tool. We then evaluate models under each condition on the MMLU benchmark, which measures general problem-solving ability across diverse domains. Experimental results show that different moral perspectives correspond to differences in model behavior during general reasoning, as reflected in both responses and internal representations. In particular, more advanced moral conditions tend to elicit more reflective reasoning patterns, which are often linked to improved performance. Our study broadens the scope of LLM morality, which has traditionally been examined mainly in ethical judgment settings. More broadly, it suggests that morality may function not only as a mechanism for safety alignment but also as a factor that shapes model behavior during reasoning. Code and data are available at https: //github.com/ISEOKIM/llm-morality