Generative Value Conflicts Reveal LLM Priorities

Researchers introduced ConflictScope, an automated pipeline that evaluates how large language models prioritize competing values when faced with ethical dilemmas. The study found that LLMs shift away from protective values like harmlessness toward personal values like user autonomy in open-ended scenarios, though system prompting can improve alignment by 14%.

• →ConflictScope automatically generates scenarios where LLMs must choose between conflicting values to evaluate their priorities.
• →Models favor personal values over protective values in open-ended responses compared to multiple-choice evaluations.
• →Detailed value orderings in system prompts can improve alignment with target rankings by 14%.
• →Existing AI alignment datasets lack sufficient value conflict scenarios for proper evaluation.
• →The research provides a foundation for evaluating and improving value prioritization in AI systems.