Abstract
Large Language Models (LLMs) are increasingly used to access and interpret information, raising concerns about their social bias and influence on public discourse. Despite growing interest in this issue, political bias in LLMs remains underexplored, particularly in low-resource languages. In this work, we investigate political bias in Polish LLMs and examine whether ideological tendencies can be steered using parliamentary speeches. We construct the Polish Parliamentary Speeches Corpus (PPSC) and fine-tune Bielik-11B-v2.2 and PLLuM-12B-instruct to produce left- and right-leaning variants. To evaluate ideological behavior, we introduce a Political Views benchmark with 266 policy statements across five domains. Our experiments show that both base models exhibit left-leaning tendencies under a basic prompt, that fine-tuning shifts ideological behavior asymmetrically and in a model-dependent way, and that system prompts can influence ideological responses more strongly than fine-tuning.
Paper Type
Short Paper
DOI
10.62036/ISD.2026.48
Asymmetric Ideological LLM Fine-Tuning in the Polish Political Domain
Large Language Models (LLMs) are increasingly used to access and interpret information, raising concerns about their social bias and influence on public discourse. Despite growing interest in this issue, political bias in LLMs remains underexplored, particularly in low-resource languages. In this work, we investigate political bias in Polish LLMs and examine whether ideological tendencies can be steered using parliamentary speeches. We construct the Polish Parliamentary Speeches Corpus (PPSC) and fine-tune Bielik-11B-v2.2 and PLLuM-12B-instruct to produce left- and right-leaning variants. To evaluate ideological behavior, we introduce a Political Views benchmark with 266 policy statements across five domains. Our experiments show that both base models exhibit left-leaning tendencies under a basic prompt, that fine-tuning shifts ideological behavior asymmetrically and in a model-dependent way, and that system prompts can influence ideological responses more strongly than fine-tuning.
Recommended Citation
Jankowski, D., Gajewski, R., Barczewski, J., Sikora, M. & Majkutewicz, J.(2026). Asymmetric Ideological LLM Fine-Tuning in the Polish Political Domain. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.48