B-score: Detecting biases in large language models using response history

TL;DR

This paper introduces B-score, a new metric for detecting biases in large language models by analyzing their responses over multi-turn conversations, improving bias detection and answer verification accuracy.

Contribution

The paper proposes B-score, a novel bias detection metric that outperforms existing confidence measures in multi-turn LLM interactions across various question types.

Findings

LLMs can reduce bias in multi-turn conversations for certain question types.

B-score effectively detects biases in subjective, random, and objective questions.

Using B-score improves answer verification accuracy on multiple benchmarks.

Abstract

Large language models (LLMs) often exhibit strong biases, e.g, against women or in favor of the number 7. We investigate whether LLMs would be able to output less biased answers when allowed to observe their prior answers to the same question in a multi-turn conversation. To understand which types of questions invite more biased answers, we test LLMs on our proposed set of questions that span 9 topics and belong to three types: (1) Subjective; (2) Random; and (3) Objective. Interestingly, LLMs are able to "de-bias" themselves in a multi-turn conversation in response to questions that seek an Random, unbiased answer. Furthermore, we propose B-score, a novel metric that is effective in detecting biases to Subjective, Random, Easy, and Hard questions. On MMLU, HLE, and CSQA, leveraging B-score substantially improves the verification accuracy of LLM answers (i.e, accepting LLM correct answers and rejecting incorrect ones) compared to using verbalized confidence scores or the frequency of single-turn answers alone. Code and data are available at: https://b-score.github.io.