Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMs

2025-05-21

AI Summary

Implicit conceptual biases in LLMs—such as asymmetric associations between target concepts and sentiment-polarity reference concepts—undermine model fairness and reliability. Existing bias detection methods, which rely on labeled test sets and social-group contrasts, suffer from high annotation costs and narrow coverage. This paper introduces BiasLens, the first test-set-agnostic framework for large-model bias analysis. It jointly models an unsupervised concept space using Concept Activation Vectors (CAVs) and sparse autoencoders (SAEs), enabling quantification of representational shifts of target concepts along reference directions (e.g., positive/negative sentiment). BiasLens innovatively uncovers structural biases induced by non-social attributes—such as insurance status—in high-stakes domains like clinical decision support. Without requiring labeled data, it achieves strong agreement with human evaluations (Spearman *r* > 0.85), supporting interpretable, scalable, and real-time bias diagnostics.

Technology Category

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

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Bias in Large Language Models (LLMs) significantly undermines their reliability and fairness. We focus on a common form of bias: when two reference concepts in the model's concept space, such as sentiment polarities (e.g.,"positive"and"negative"), are asymmetrically correlated with a third, target concept, such as a reviewing aspect, the model exhibits unintended bias. For instance, the understanding of"food"should not skew toward any particular sentiment. Existing bias evaluation methods assess behavioral differences of LLMs by constructing labeled data for different social groups and measuring model responses across them, a process that requires substantial human effort and captures only a limited set of social concepts. To overcome these limitations, we propose BiasLens, a test-set-free bias analysis framework based on the structure of the model's vector space. BiasLens combines Concept Activation Vectors (CAVs) with Sparse Autoencoders (SAEs) to extract interpretable concept representations, and quantifies bias by measuring the variation in representational similarity between the target concept and each of the reference concepts. Even without labeled data, BiasLens shows strong agreement with traditional bias evaluation metrics (Spearman correlation r>0.85). Moreover, BiasLens reveals forms of bias that are difficult to detect using existing methods. For example, in simulated clinical scenarios, a patient's insurance status can cause the LLM to produce biased diagnostic assessments. Overall, BiasLens offers a scalable, interpretable, and efficient paradigm for bias discovery, paving the way for improving fairness and transparency in LLMs.

Problem

• Detects bias in LLMs without manual test sets
• Measures asymmetric concept correlations causing bias
• Reveals hidden biases in clinical and social contexts

Innovation

• Uses Concept Activation Vectors for bias analysis
• Employs Sparse Autoencoders to extract concepts
• Quantifies bias via representational similarity variation