Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bias Beyond Borders: Political Ideology Evaluation and Steering in Multilingual LLMs

Afrozah Nadeem, Agrima, Mehwish Nasim, Usman Naseem

TL;DR

This work investigates political bias in multilingual LLMs across 50 countries and 33 languages using the Political Compass Test (PCT) to quantify economic and social dimensions. It introduces Cross-Lingual Alignment Steering (CLAS), an activation-level, post-hoc mitigation framework that first aligns language-specific ideological representations into a shared subspace and then applies uncertainty-aware steering to reduce bias without harming utility. The study demonstrates substantial bias reduction across languages and axes, with CLAS achieving more consistent cross-lingual mitigation than traditional monolingual steering approaches, particularly in low-resource languages. Overall, the approach provides a scalable, interpretable paradigm for fair, culturally aware multilingual LLM governance that preserves linguistic and semantic nuance while improving cross-language neutrality.

Abstract

Large Language Models (LLMs) increasingly shape global discourse, making fairness and ideological neutrality essential for responsible AI deployment. Despite growing attention to political bias in LLMs, prior work largely focuses on high-resource, Western languages or narrow multilingual settings, leaving cross-lingual consistency and safe post-hoc mitigation underexplored. To address this gap, we present a large-scale multilingual evaluation of political bias spanning 50 countries and 33 languages. We introduce a complementary post-hoc mitigation framework, Cross-Lingual Alignment Steering (CLAS), designed to augment existing steering methods by aligning ideological representations across languages and dynamically regulating intervention strength. This method aligns latent ideological representations induced by political prompts into a shared ideological subspace, ensuring cross lingual consistency, with the adaptive mechanism prevents over correction and preserves coherence. Experiments demonstrate substantial bias reduction along both economic and social axes with minimal degradation in response quality. The proposed framework establishes a scalable and interpretable paradigm for fairness-aware multilingual LLM governance, balancing ideological neutrality with linguistic and cultural diversity.

Bias Beyond Borders: Political Ideology Evaluation and Steering in Multilingual LLMs

TL;DR

This work investigates political bias in multilingual LLMs across 50 countries and 33 languages using the Political Compass Test (PCT) to quantify economic and social dimensions. It introduces Cross-Lingual Alignment Steering (CLAS), an activation-level, post-hoc mitigation framework that first aligns language-specific ideological representations into a shared subspace and then applies uncertainty-aware steering to reduce bias without harming utility. The study demonstrates substantial bias reduction across languages and axes, with CLAS achieving more consistent cross-lingual mitigation than traditional monolingual steering approaches, particularly in low-resource languages. Overall, the approach provides a scalable, interpretable paradigm for fair, culturally aware multilingual LLM governance that preserves linguistic and semantic nuance while improving cross-language neutrality.

Abstract

Large Language Models (LLMs) increasingly shape global discourse, making fairness and ideological neutrality essential for responsible AI deployment. Despite growing attention to political bias in LLMs, prior work largely focuses on high-resource, Western languages or narrow multilingual settings, leaving cross-lingual consistency and safe post-hoc mitigation underexplored. To address this gap, we present a large-scale multilingual evaluation of political bias spanning 50 countries and 33 languages. We introduce a complementary post-hoc mitigation framework, Cross-Lingual Alignment Steering (CLAS), designed to augment existing steering methods by aligning ideological representations across languages and dynamically regulating intervention strength. This method aligns latent ideological representations induced by political prompts into a shared ideological subspace, ensuring cross lingual consistency, with the adaptive mechanism prevents over correction and preserves coherence. Experiments demonstrate substantial bias reduction along both economic and social axes with minimal degradation in response quality. The proposed framework establishes a scalable and interpretable paradigm for fairness-aware multilingual LLM governance, balancing ideological neutrality with linguistic and cultural diversity.
Paper Structure (28 sections, 4 equations, 12 figures, 3 tables)

This paper contains 28 sections, 4 equations, 12 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Ideological bias evaluation with PCT
  4. Why political bias matters?
  5. Multilingual and cross-lingual fairness in LLMs
  6. Methodology
  7. Bias Evaluation
  8. Bias Mitigation
  9. Steering Vector Computation
  10. Cross-Lingual Alignment Steering
  11. Bias Measurement
  12. Prompt specification and Reproducibility
  13. Results
  14. Bias Evaluation
  15. Political Inclination Across Languages:
  16. ...and 13 more sections

Figures (12)

  • Figure 1: Multilingual stance distribution across 30+ languages for PCT statements in the deepseek model. The top image visualizes aggregated agreement patterns, while the bottom image describes the proportion of responses ranging from Strongly Disagree to Strongly Agree. Variations across languages reveal significant cross-linguistic differences in ideological alignment, emphasizing how translation and cultural framing affect stance interpretation in multilingual LLM evaluations.
  • Figure 2: Multilingual political stance evaluation. Mean stance scores with 95% bootstrap confidence intervals (left) and stance score distributions (right) across languages, highlighting cross-lingual variation and consistency without normative interpretation.
  • Figure 3: Cross-linguistic variation in model bias across social and economic axes for five European languages.
  • Figure 4: After steering mean stance per language with 95% bootstrap confidence intervals (shaded), result shows the bias reduction in ideological magnitude and cross-lingual variance, bringing most languages close to neutral.
  • Figure 5: Cross-linguistic variation in model bias across social and economic axes for DeepSeek model.
  • ...and 7 more figures