Multilingual Political Views of Large Language Models: Identification and Steering

TL;DR

The paper investigates political biases in modern open-source instruction-tuned LLMs across architectures and languages, addressing gaps in generalizability and controllability. It conducts a large-scale multilingual evaluation using the Political Compass Test (PCT) across seven models in 14 languages with 11 paraphrase prompts, and introduces a center-of-mass inference-time steering method to manipulate ideological outputs. Key findings show that larger models exhibit a shift toward libertarian-left positions with language- and model-specific variations, and that political orientations can be steered through lightweight attention-head interventions across languages. The work provides a replicable methodology, datasets, and code, underscoring implications for bias mitigation, model alignment, and responsible deployment in multilingual, user-facing AI systems.

Abstract

Large language models (LLMs) are increasingly used in everyday tools and applications, raising concerns about their potential influence on political views. While prior research has shown that LLMs often exhibit measurable political biases--frequently skewing toward liberal or progressive positions--key gaps remain. Most existing studies evaluate only a narrow set of models and languages, leaving open questions about the generalizability of political biases across architectures, scales, and multilingual settings. Moreover, few works examine whether these biases can be actively controlled. In this work, we address these gaps through a large-scale study of political orientation in modern open-source instruction-tuned LLMs. We evaluate seven models, including LLaMA-3.1, Qwen-3, and Aya-Expanse, across 14 languages using the Political Compass Test with 11 semantically equivalent paraphrases per statement to ensure robust measurement. Our results reveal that larger models consistently shift toward libertarian-left positions, with significant variations across languages and model families. To test the manipulability of political stances, we utilize a simple center-of-mass activation intervention technique and show that it reliably steers model responses toward alternative ideological positions across multiple languages. Our code is publicly available at https://github.com/d-gurgurov/Political-Ideologies-LLMs.

Figures (11)

• Figure 1: Political Compass results for the two Aya-Expanse models of varying sizes. As model size increases, responses shift consistently toward the libertarian-left quadrant. Results for the other evaluated models are provided in Appendix \ref{['app:results']}.
• Figure 2: Probes trained for LLaMA-3.1-8B.
• Figure 3: Intervention results for LLaMA-3.1-8B for $K$=512 and $\alpha$=20. Top: direction towards liberal. Bottom: direction towards conservative.
• Figure 4: Political compass results across the rest of the models of various sizes. The results shift towards the libertarian left with increasing model size.
• Figure 5: Political compass intervention results on 256 heads for two different intervention strengths for both directions on the PCT test in English. The plots on the right demonstrate steering towards politically right responses, and the plots on the left--towards politically left responses.