This research introduces steering vectors that modify model activations during inference to mitigate biases in large language models.

• Achieves 12.2%, 4.7%, and 3.2% bias reduction in Mistral, Llama, and Qwen models respectively
• Uses Bayesian optimization to identify effective contrastive datasets across nine bias dimensions
• Demonstrates that ensembles of steering vectors can address multiple bias types simultaneously
• Provides a computationally efficient approach requiring no model fine-tuning

This research significantly advances AI security by addressing harmful biases that could lead to discrimination or unfair treatment when LLMs are deployed in real-world applications.

Shifting Perspectives: Steering Vector Ensembles for Robust Bias Mitigation in LLMs