Aligning Large Language Model Agents with Rational and Moral Preferences: A Supervised Fine-Tuning Approach

Researchers developed a supervised fine-tuning approach to align large language model agents with specific economic preferences, addressing systematic deviations from rational behavior in strategic environments. The study demonstrates how LLM agents can be trained to follow either self-interested or morally-guided strategies, producing distinct outcomes in economic games and pricing scenarios.

• →Off-the-shelf LLM agents show excessive cooperation and limited responsiveness to incentives in economic games.
• →Supervised fine-tuning can align AI agents with explicit economic preferences like self-interest maximization or Kantian moral principles.
• →Fine-tuning on small, theory-driven synthetic datasets produces persistent behavioral changes in strategic decision-making.
• →Different preference alignments lead to systematically distinct equilibrium outcomes in moral dilemmas and pricing scenarios.
• →The research frames AI alignment in multi-agent settings as an objective-design problem guided by economic theory.