FairMonitor: A Dual-framework for Detecting Stereotypes and Biases in Large Language Models

TL;DR

The paper tackles the challenge of detecting stereotypes and biases in large-language models by arguing that traditional embedding- or probability-based methods fail to capture implicit biases in realistic scenarios. It introduces FairMonitor, a dual static–dynamic framework that combines a static three-stage bias test (direct inquiry, implicit association, unknown situation) with a dynamic multi-agent education-focused simulation, and releases Edu-FairMonitor, a benchmark of 10,262 open-ended questions across 9 factors and 26 educational scenarios. The dynamic component is formalized through a role-based agent system whose interactions are governed by the relation $\mathcal{E} = f(\mathcal{A}, \mathcal{M}, \mathcal{P}, \mathcal{I})$, enabling ~600 simulated dialogues and diverse interaction modes. Experimental results across five LLMs show that the combined static/dynamic approach detects more stereotypes and biases than static methods alone, revealing explicit and implicit biases across gender, race, age, SES, and learning profiles in educational contexts, with significant implications for fairness in real-world applications.

Abstract

Detecting stereotypes and biases in Large Language Models (LLMs) is crucial for enhancing fairness and reducing adverse impacts on individuals or groups when these models are applied. Traditional methods, which rely on embedding spaces or are based on probability metrics, fall short in revealing the nuanced and implicit biases present in various contexts. To address this challenge, we propose the FairMonitor framework and adopt a static-dynamic detection method for a comprehensive evaluation of stereotypes and biases in LLMs. The static component consists of a direct inquiry test, an implicit association test, and an unknown situation test, including 10,262 open-ended questions with 9 sensitive factors and 26 educational scenarios. And it is effective for evaluating both explicit and implicit biases. Moreover, we utilize the multi-agent system to construst the dynamic scenarios for detecting subtle biases in more complex and realistic setting. This component detects the biases based on the interaction behaviors of LLMs across 600 varied educational scenarios. The experimental results show that the cooperation of static and dynamic methods can detect more stereotypes and biased in LLMs.

Figures (12)

• Figure 1: The motivation of this work.
• Figure 2: The framework for the FairMoniter.
• Figure 3: Static Detection: Overall Performance Analysis
• Figure 4: Static Detection: Performance Analysis on three stages.
• Figure 5: Static Detection: Performance Analysis on Nine Sensitive Factors.