Quantifying Bias in Agentic Large Language Models: A Benchmarking Approach

The rapid adoption of large language models (LLMs) as agents raises concerns about potential biases in their decision-making processes. While previous work has explored bias mitigation in open text generation, the analysis of bias in LLM-based agents with constrained choices is under-explored. This paper introduces a new benchmark for evaluating bias in such agents, utilizing a question-answering framework across simulated real-life scenarios in healthcare, criminal justice, and business. We analyze potential biases related to race, gender, age, political affiliation, and socioeconomic status. Our novel question-answering bias distribution diversity metric quantifies the LLM’s decision-making tendencies. We find that pre-trained models exhibit varying degrees of bias across domains and categories, offering insights for future bias mitigation strategies.