Be Less BIASed: With a Research-Inspired Multi-Model Pipeline Towards Less Biases in LLM Outputs

Within the BIAS project, we investigated AI biases in the labor market through a combination of ethnographic field studies and interdisciplinary research in the social sciences and humanities (SSH). In parallel, the technical part of the project dealing with NLP technologies developed a demonstrator aiming at supporting human resources (HR) practitioners in formulating prompts that lead to less biased answers from any third-party large language model (LLM). In this paper, we present the architecture of the proposed multi-model pipeline and report on an initial validation study, comparing expert-based human evaluation with a LLM-as-a-judge approach. Our results contribute both to research on bias reduction in AI-supported hiring processes and to methodological advances in the evaluation of such systems.