The proposed alignment framework treats LLMs as proxy agents to enhance social science survey methodologies. By formulating alignment as a two-stage problem, the authors develop a system named P2P, which enables LLMs to reflect realistic survey responses without dependency on demographic information. This approach promises to increase data efficiency by leveraging aggregated survey results rather than focusing heavily on personalization.
The authors validate their method using real-world opinion survey datasets, demonstrating that the aligned LLM agents can effectively reproduce aggregate response patterns while maintaining diversity in their outputs. This capability is particularly significant given the rising challenges of survey deployment costs and demographic imbalances in response data. As such, the proposed framework not only has implications for data collection practices but also offers a valuable platform for further advancements in pluralistic alignment studies.
👉 Pročitaj original: arXiv AI Papers
