TPvG Framework Evaluates LLM Moral Decisions with Consequence Feedback.

This paper introduces TPvG (Text-based Pain-versus-Gain), a novel framework designed to assess the moral decision-making capabilities of large language models (LLMs). Unlike traditional evaluations that present isolated moral vignettes for single-shot decisions, TPvG incorporates consequence feedback, a critical factor known to influence human moral behavior. The framework consists of five moral decision tasks, progressing from minimal-context, one-shot choices to sequential decisions where explicit feedback on consequences is provided. These tasks embed an everyday moral dilemma: avoiding harm to others versus maximizing self-gain. The study's results indicate that LLM moral decisions are significantly influenced by the decision format, showing different responses in one-shot versus sequential scenarios. Furthermore, explicit receiver feedback produced varied effects across different LLMs, and crucially, LLM responses to this feedback often converged from observed human patterns. These findings suggest that LLMs may employ different decision processes compared to humans and underscore the importance of evaluating LLM moral behavior in dynamic, high-stakes interactive environments to ensure stability and alignment with human values.