Interpretable Multi-Agent Debate for Political Opinion Simulation

We present a multi-agent debate framework for simulating political party identification from demographic and attitudinal profiles. Our system employs two advocate agents arguing for opposing party affiliations, with a judge agent evaluating their arguments and producing probabilistic predictions. Using data from the 2024 American National Election Studies (ANES), we evaluate our approach across six demographically diverse subgroups. While simple baselines achieve superior distributional matching by construction, our debate system achieves competitive distributional fidelity while providing interpretable reasoning traces that explain how demographic characteristics and policy attitudes interact to predict party identification. We argue that for political opinion simulation, interpretability is a crucial dimension alongside distributional fidelity, as understanding why predictions are made enables validation, debugging, and insight generation that opaque methods cannot provide.