Measuring Behavior Portability in Large Language Models

The article introduces a formal framework for measuring behavioral portability in large language models (LLMs), addressing how well a model's learned behavior in one decision environment transfers to another with the same incentive structure. The authors develop a protocol that benchmarks predictive performance against an oracle trained on target data, revealing significant portability losses across seven canonical economic decision problems. This research highlights the limitations of using LLMs as autonomous decision-makers across varied environments, emphasizing the need for careful evaluation of their behavioral mappings.