Cultural Misalignment Patterns Contradict Expectations
According to arxiv.org research published in September 2026, large language models exhibit unexpected cultural bias patterns that contradict common assumptions about home-country favoritism. The study evaluated three open-weight LLMs—Gemma3-12B (USA), Bielik-11B-v3 (Poland), and Qwen3-4B (China)—against World Values Survey Wave 7 data across 63 demographic personas in three countries.
Contrary to expectations, no model favored its home country. The Chinese-built Qwen3-4B “performs worst on its own Chinese population (W1 = 0.436, the highest misalignment in the entire model x country matrix),” according to the study.
Targeted Fine-Tuning Shows Mixed Results
The research demonstrated that targeted LoRA fine-tuning on the five worst-case personas—requiring fewer than 1,200 training pairs and under 15 minutes on a single GPU—reduced bias by 16.8% for Bielik-11B (p_Bonf = 0.002, d = -4.4) with all five targets improving, according to arxiv.org.
However, the study revealed that “fine-tuning redistributes rather than removes bias: Bielik’s worst-case personas swap entirely from American to Chinese elderly, with zero overlap between pre- and post-correction sets.”
Efficient Personalization Framework Proposed
In parallel research, arxiv.org reported on PLUME (Personalized Low-Rank Adaptation through User Modulation and Shared Subspace), which “reduces per-user parameters by over 95%” while achieving comparable or superior performance to baselines on personalized text generation benchmarks.