Developing A Framework to Support Human Evaluation of Bias in Generated Free Response Text

TL;DR

The paper tackles the challenge of evaluating bias in large language models within ecologically valid, free-text question-answering contexts. It adapts the Bias Benchmark for Question Answering (BBQ) to elicit open-ended responses, enabling a semi-automated, human-in-the-loop evaluation pipeline. A key contribution is an operational bias definition based on equivalence under name reversal, which supports automatic pruning of strictly unbiased cases and focused categorization of residual bias. The work also identifies problematic templates and offers guidance for constructing robust, ecologically valid bias benchmarks that capture how LLMs behave in practice.

Abstract

LLM evaluation is challenging even the case of base models. In real world deployments, evaluation is further complicated by the interplay of task specific prompts and experiential context. At scale, bias evaluation is often based on short context, fixed choice benchmarks that can be rapidly evaluated, however, these can lose validity when the LLMs' deployed context differs. Large scale human evaluation is often seen as too intractable and costly. Here we present our journey towards developing a semi-automated bias evaluation framework for free text responses that has human insights at its core. We discuss how we developed an operational definition of bias that helped us automate our pipeline and a methodology for classifying bias beyond multiple choice. We additionally comment on how human evaluation helped us uncover problematic templates in a bias benchmark.