Acmite Framework Reduces LLM Gender Bias via Concept-Guided Mutual Information
October 2, 2026
Acmite utilizes a lightweight LoRA adapter and token-level KL divergence to minimize the statistical dependence between model outputs and stereotype concepts. It uses maximal marginal relevance to select diverse semantic concepts, providing more targeted debiasing than simple word substitution.
HOW THIS AFFECTS YOU
●
researcherThis method moves debiasing from simple substitution to modeling statistical dependence via mutual information.
●
policyThis provides a more robust technical mechanism for implementing safety and fairness constraints in LLMs.