Simulating Tenant Responses to Energy Policy Interventions with Transaction-Cost-Aware LLM Agents

Researchers built a friction-aware persona framework for LLMs that includes perceived transaction costs such as effort, uncertainty, and coordination demands.

On Dutch survey data about energy-efficient renovation, adding these factors improved performance for both prompt-only and fine-tuned models.

The gains held across GPT-3.5-turbo and open-weight models such as Ministral-8B-Instruct and Llama-3.1-8B-Instruct, including SFT and GRPO training.

The study suggests transaction-cost-aware personas can simulate tenant responses to energy policy more accurately and in a more interpretable way.