Can LLMs Infer Political Parties’ Strategic Behavior? Evidence from Spatial Voting Models

Yulu Pi, Aojie Ju

Keywords: LLM validation, agent-based simulation, spatial voting, strategy inference, political competition

Abstract:

Before LLMs can credibly serve as political simulation agents to approximate political actors, they must first demonstrate an accurate
understanding of the strategic behaviors they are expected to emulate. We introduce a validation task that tests this prerequisite:
given only time-series data on party positions and vote shares from a Laver (2005) multiparty competition model, LLMs must infer
which of five adaptive strategies governs each party’s movement, without access to strategy labels. Across six model configurations
spanning three families, overall accuracy ranges from 9.4% to 39% , with substantial variation by strategy type and model choice. These
results reveal that current LLMs lack reliable comprehension of the competitive dynamics they would need to reproduce, raising a
foundational concern for their ready deployment as simulation agents for political dynamics

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 4