I CAN Debias: Uncovering and Enhancing Self-debiasing Mechanisms of Social Bias in LLMs

LLMs have built-in "self-debiasing" abilities: Large language models possess an intrinsic mechanism that helps them avoid generating biased outputs, though this capability was previously poorly understood and underutilized in bias-mitigation efforts.

A tiny fraction of neurons do the heavy lifting: The research identifies that less than 2% of a model's parameters are responsible for self-debiasing, and these neurons are "causally critical"—when deactivated, biased responses jump above 95%, demonstrating their essential role in fair outputs.

Targeted fine-tuning is more efficient than previous approaches: Rather than relying on carefully crafted prompts or broad retraining that risks "catastrophic forgetting" (where models lose previously learned capabilities), the I-CAN-Debias framework enables targeted adjustments to specific neurons that improve fairness while preserving the model's general abilities.

Self-debiasing neurons are strategically located: These critical neurons cluster in specific architectural locations (query and value projections in later layers), suggesting that bias mitigation isn't distributed randomly throughout the model but concentrated in identifiable regions.

Improvements generalize beyond training scenarios: Models fine-tuned using this approach show robust fairness improvements across diverse evaluations, including open-ended fairness tests and resistance to jailbreak attacks, indicating the approach produces genuine bias reduction rather than narrow fixes.