Targeted LoRA Fine-Tuning Mitigates LLM Cultural Misalignment
September 7, 2026
Evaluation of Gemma3, Bielik, and Qwen3 reveals significant cultural misalignment with World Values Survey data. Targeted LoRA fine-tuning on just 1,200 pairs can reduce bias by 16.8%, though results suggest fine-tuning may redistribute rather than eliminate bias across demographics.
HOW THIS AFFECTS YOU
●
researcherYou can use these metrics to quantify the distributional misalignment of LLMs across cultures.
●
policyThis highlights the difficulty of removing bias through localized fine-tuning.