New Framework for Evaluating LLM Bias Across Use Cases

A new decision framework has been developed by researchers to aid in the selection of bias and fairness metrics for Large Language Models (LLMs), tailored to specific deployment contexts. This framework aligns LLM applications, characterized by the model and prompt demographics, with pertinent metrics that take into account the type of task, mentions of protected attributes, and the priorities of stakeholders. It tackles issues such as toxicity, stereotyping, counterfactual unfairness, and allocational harms, introducing innovative metrics that utilize stereotype classifiers and counterfactual text similarity. Additionally, an open-source Python library named langfair has been launched for practical implementation. Experiments involving five LLMs and five prompt populations reveal that benchmark performance alone is insufficient for accurately assessing fairness risks.