BiasLab 多语言双框架用于 LLM 偏见测量：面向职场与人力资源场景

Jul 16, 2026

Upload by Anda

Large language models (LLMs) exhibit systematic biases with high impact in workplace and HR settings, yet existing evaluation methods are fragmented and limit practitioners’ ability to assess deployment risk. The study proposes BiasLab, a multilingual dual-framing framework that quantifies and compares directional, output-level bias across six HR-relevant topics. Using mirrored prompt pairs, perturbation wrappers, fixed-choice constraints, and polarity-aligned scoring, 10 LLMs generate 43,200 responses across 12 languages.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages26

Reading Time1 h 6 min

Abstract

Background and Motivation

Introduction

Ethical Approval and Declarations

AI Use Disclosure

Funding Statement

Data Availability and Live Demo

Methods

Results

Conclusions

BiasLab: A Multilingual Dual-Framing Framework for LLM Bias Measurement, Applied to Workplace and HR Contexts William Guey¹, Wei Zhang¹, Pei-Luen Patrick Rau¹, Pierrick Bougault¹, Vitor D. de Moura², Bertan Ucar¹and José O. Gomes³ ¹Department of Industrial Engineering, Tsinghua University, Beijing, China ²School of Social Sciences, Tsinghua University, Beijing, China ³Department of Industrial Engineering, Federal University of Rio de Janeiro, Brazil Corresponding author William Guey Email: [guijt24@mails.tsinghua.edu.cn](guijt24@mails.tsinghua.edu.cn) Address: Shunde Building, Tsinghua University, Beijing, China Author Contributions William Guey: Conceptualization, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Visualization, Project Administration. Wei Zhang: Conceptualization, Supervision, Writing – Review & Editing. Pei-Luen Patrick Rau: Conceptualization, Methodology, Supervision, Writing – Review & Editing. Pierrick Bougault: Software, Writing – Review & Editing. Vitor D. de Moura: Conceptualization, Writing – Review & Editing. Bertan Ucar: Software, Formal Analysis, Writing – Review & Editing. José Orlando Gomes: Conceptualization, Writing – Review & Editing. Acknowledgements The authors thank the Department of Industrial Engineering at Tsinghua University for their support. Statements and Declarations Ethical approval This study did not involve human participants, human data, or human tissue. No ethical approval was required. Consent to participate Not applicable. This study did not involve human participants. AI Use Disclosure The BiasLab evaluation system was developed with AI-assisted programming support. Large language model tools were used in the writing and debugging of code used to conduct the automated multilingual bias evaluation, data collection, and visualization described in this study. All code was reviewed and validated by the authors. No AI tools were used in the interpretation of results or the writing of this manuscript. Informed consent Not applicable. This study does not contain data from any individual person. Conflict of interest The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Funding statement This work was supported by a full doctoral scholarship awarded by Tsinghua University to the first author. Data Availability and Live Demo The full BiasLab implementation, including multilingual probe generation, robustness evaluation, scoring, and visualization modules, is publicly available via: (i) the project GitHub repository ([https://github.com/williamguey/LLMbiaslab](https://github.com/williamguey/LLMbiaslab)); (ii) a mirrored release on Hugging Face ([https://huggingface.co/spaces/Realmente/biaslab](https://huggingface.co/spaces/Realmente/biaslab)); and (iii) the project website [llmbias.org](llmbias.org), which provides access to a public testing interface, documentation, and example outputs. The repository additionally includes the raw Excel output files for all six evaluated topics and the original analysis charts generated during the study. Abstract Background: Large language models (LLMs) harbor systematic biases that are particularly consequential in workplace and HR contexts, where their outputs increasingly influence hiring, job design, and organizational decisions. Yet existing bias-evaluation approaches remain methodologically fragmented, limiting practitioners' ability to assess deployment risks. Objective: This study introduces BiasLab, a multilingual dual-framing framework to quantify and compare directional output-level bias in LLMs, demonstrated across six workplace and HRrelevant topics. Methods: BiasLab combines mirrored affirmative and reverse prompt pairs, randomized wrapper perturbations, fixed-choice response constraints, and polarity-aligned scoring. Ten LLMs were evaluated acros

FAQs

What problem does BiasLab address in evaluating LLM bias for HR use?

It addresses the fragmented nature of existing bias-evaluation approaches, which makes it hard for practitioners to assess deployment risks where LLM outputs influence hiring and organizational decisions.

How does BiasLab measure directional bias in LLM outputs?

BiasLab uses mirrored affirmative and reverse prompt pairs, randomized wrapper perturbations, fixed-choice response constraints, and polarity-aligned scoring to quantify preferences at the output level.

What key bias pattern did the evaluated models show across workplace and HR topics?

All ten models showed consistent directional preferences, and they exhibited an asymmetric pattern: they rejected disfavored claims more strongly than they endorsed their opposites, which single-frame designs could miss.

Jul 16, 2026

Upload by Anda