Mitigating Selection Bias with Node Pruning and Auxiliary Options

ACL ARR 2025 February Submission1606 Authors

Abstract: Large language models (LLMs) often show a preference for certain choice options when responding to multiple-choice questions. This behavior, called selection bias, makes a model’s answers less accurate and helpful. Previous solutions to this problem have used debiasing methods to adjust the model’s inputs or outputs. Our work, in contrast, looks inside the model to understand and remove the sources of selection bias. We present two solutions: Bias Node Pruning (BNP), which removes parts of the model that cause selection bias, and Auxiliary Option Injection (AOI), which adds an extra answer choice to reduce bias. We also introduce a new measure of selection bias, Choice Kullback-Leibler Divergence (CKLD), which addresses the insensitivity of other metrics to imbalance in choice labels. Tests with three LLMs show that our methods work well with diverse questions and datasets.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: model bias, fairness evaluation, model bias/unfairness mitigation, large language models, selection bias

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 1606