This study examines bias patterns across 13 different Large Language Models, revealing that bias similarities persist even across different model families and after fine-tuning.

• Models from the same family show highly similar bias patterns (up to 0.99 correlation)
• Even models from different families show significant bias correlations
• Fine-tuning has minimal impact on reducing underlying bias distribution
• Models with similar architectures exhibit similar bias patterns regardless of training data

From a security perspective, these findings suggest that bias transfer between models creates persistent vulnerabilities that current debiasing techniques fail to address effectively.

Bias Similarity Across Large Language Models