Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

2026-07-30

AI Summary

This work addresses the pervasive issue of demographic biases implicitly encoded in large language models and proposes Fairness Pruning, a precise intervention mechanism that identifies and selectively nullifies neurons disproportionately responsive to demographic attributes. By leveraging minimal contrastive prompts and activation analysis during inference, the method pinpoints bias-sensitive neurons within the down_proj input layer of GLU-based MLPs. Remarkably, ablating as few as 40 neurons—less than 0.031% of the MLP width—significantly attenuates biased model responses while preserving 99.49% of general reasoning and commonsense capabilities. This study provides the first empirical evidence that biased and general-purpose functionalities are neurologically separable, establishing both theoretical grounding and practical feasibility for targeted behavioral modulation over blunt, performance-degrading interventions.

Abstract

This work presents Fairness Pruning, a lightweight structural intervention method designed for the management and future mitigation of demographic bias in large language models (LLMs). As a foundational empirical validation of this method, this work focuses on causal bias localization. Using minimally contrastive prompt pairs and inference-time activation capture, the method identifies neurons that react differentially when processing demographic attributes in GLU architectures, evaluating the signal at the down_proj input. Empirical evaluation was conducted on models of up to 3 billion parameters (Llama-3.2 family and Salamandra-2B), combining standardized benchmark evaluation with qualitative text generation experiments. Results demonstrate that zeroing the identified neurons alters how the model responds to associated demographic variables. However, rather than producing flat mitigation, the intervention causes bidirectional bias destabilization: because BiasScore is unsigned, candidate sets mix neurons that push toward and against the stereotype, and the net effect on aggregate bias depends on which sign dominates. The intervention is extremely surgical: zeroing at most 40 neurons in Llama-3.2-1B (less than 0.031% of total MLP width) achieves a mean retention of 99.49% in reasoning and general knowledge capabilities. These findings empirically confirm that demographic bias processing and model capabilities operate on dissociable circuits, establishing the methodological foundations for transitioning from blind zeroing toward directional behavior modulation.

Problem

• demographic bias
• large language models
• bias localization
• GLU-MLP
• fairness

Innovation

• Fairness Pruning
• GLU-MLP
• differential activation
• bias localization
• structural intervention