Political Ideology Shifts in Large Language Models — Data & Results

This repository contains the supplementary material for the paper Political Ideology Shifts in Large Language Models.
We provide two ZIP archives:

• data.zip — All necessary input files to reproduce the results in the paper.

• results.zip — Processed output files containing the results reported in the paper.

The associated codebase for running the experiments is available at:
https://github.com/d-lab/political-ideology-shifts-in-LLMs

Our study investigates how adopting synthetic personas influences ideological expression in seven large language models (7B–70B+ parameters), using the Political Compass Test as a standardized probe. We identify scale-dependent patterns in ideological malleability and reveal that both explicit ideological cues and thematic content in persona descriptions systematically steer model behavior.