Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics?

2025-12-23

AI Summary

This work investigates the decoupling of demographic bias mechanisms from foundational identification capabilities in language models: specifically, whether racial/gender stereotypes can be mitigated without degrading accuracy in recognizing names, occupations, or education levels. We propose a multi-task evaluation framework and sparse autoencoder-based interpretability analysis, revealing—for the first time—that bias stems from task-specific internal mechanisms rather than global demographic token representations. Building on this insight, we design dual-path intervention strategies: attribution-driven and correlation-driven feature localization and editing. Applying surgical inference-time interventions on Gemma-2-9B yields three key results: (1) occupational stereotyping is significantly reduced while name recognition accuracy remains intact; (2) educational bias mitigation is more effectively achieved via correlation-driven methods; and (3) attribution-based ablation induces prior collapse and paradoxical bias amplification, empirically validating both the necessity and efficacy of mechanistic decoupling.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Computer Vision: Bias, Fairness & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

We investigate how independent demographic bias mechanisms are from general demographic recognition in language models. Using a multi-task evaluation setup where demographics are associated with names, professions, and education levels, we measure whether models can be debiased while preserving demographic detection capabilities. We compare attribution-based and correlation-based methods for locating bias features. We find that targeted sparse autoencoder feature ablations in Gemma-2-9B reduce bias without degrading recognition performance: attribution-based ablations mitigate race and gender profession stereotypes while preserving name recognition accuracy, whereas correlation-based ablations are more effective for education bias. Qualitative analysis further reveals that removing attribution features in education tasks induces ``prior collapse'', thus increasing overall bias. This highlights the need for dimension-specific interventions. Overall, our results show that demographic bias arises from task-specific mechanisms rather than absolute demographic markers, and that mechanistic inference-time interventions can enable surgical debiasing without compromising core model capabilities.

Problem

• Investigates whether language models can remove demographic bias while preserving demographic recognition capabilities
• Compares attribution-based and correlation-based methods for locating bias features in models
• Examines if bias arises from task-specific mechanisms rather than absolute demographic markers

Innovation

• Targeted sparse autoencoder feature ablations reduce bias
• Attribution-based ablations mitigate stereotypes while preserving accuracy
• Correlation-based ablations effectively address education bias