FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs

TL;DR

<3-5 sentence high-level summary> FairMT-Bench introduces a dedicated benchmark for evaluating fairness in multi-turn dialogues of LLMs, addressing bias accumulation and contextual interference that single-turn benchmarks miss. It defines a three-stage task taxonomy (context understanding, interaction fairness, fairness trade-off) and builds the FairMT-10K dataset from stereotype and toxicity sources, plus a more challenging FairMT-1K for rigorous testing. Using GPT-4 as the primary judge and Llama-Guard-3 as an auxiliary detector, the study evaluates multiple models across turns, bias types, and attributes, revealing systematic bias growth and significant model-task variation. The work underscores the need for robust multi-turn fairness alignment and provides a valuable dataset and framework to drive future improvements in realistic conversational AI fairness.

Abstract

The growing use of large language model (LLM)-based chatbots has raised concerns about fairness. Fairness issues in LLMs can lead to severe consequences, such as bias amplification, discrimination, and harm to marginalized communities. While existing fairness benchmarks mainly focus on single-turn dialogues, multi-turn scenarios, which in fact better reflect real-world conversations, present greater challenges due to conversational complexity and potential bias accumulation. In this paper, we propose a comprehensive fairness benchmark for LLMs in multi-turn dialogue scenarios, \textbf{FairMT-Bench}. Specifically, we formulate a task taxonomy targeting LLM fairness capabilities across three stages: context understanding, user interaction, and instruction trade-offs, with each stage comprising two tasks. To ensure coverage of diverse bias types and attributes, we draw from existing fairness datasets and employ our template to construct a multi-turn dialogue dataset, \texttt{FairMT-10K}. For evaluation, GPT-4 is applied, alongside bias classifiers including Llama-Guard-3 and human validation to ensure robustness. Experiments and analyses on \texttt{FairMT-10K} reveal that in multi-turn dialogue scenarios, current LLMs are more likely to generate biased responses, and there is significant variation in performance across different tasks and models. Based on this, we curate a challenging dataset, \texttt{FairMT-1K}, and test 15 current state-of-the-art (SOTA) LLMs on this dataset. The results show the current state of fairness in LLMs and showcase the utility of this novel approach for assessing fairness in more realistic multi-turn dialogue contexts, calling for future work to focus on LLM fairness improvement and the adoption of \texttt{FairMT-1K} in such efforts.

Figures (23)

• Figure 1: An illustration of the challenges in multi-turn dialogues. When biased content is conveyed using pronouns in multi-turn dialogues, LLMs that appear fair in single-turn dialogues may fail to understand the context of the bias, thus continuing to generate biased content.
• Figure 2: An overview of our Fair-MT Bench. We first formulate a task taxonomy targeting LLM fairness capabilities across three stages: context understanding, user interaction, and instruction trade-offs, with each stage comprising two tasks. Based on this, we collect datasets encompassing two major bias types (stereotype, toxicity) and six bias attributes (gender, race, religion, etc.), covering nearly all bias types and attributes commonly addressed in fairness evaluation.
• Figure 3: Bias ratio of different LLMs on FairMT-10K evaluated by Llama-Guard-3-8B. We use abbreviations instead of task names, SQ stands for Scattered Questions, AE stands for Anaphora Ellipsis, JT stands for Jailbreak Tips, IM stands for Interference from Misinformation, NF stands for Negative Feedback, FF stands for Fixed Format.
• Figure 4: Comparison of bias ratio in single versus multi-turn dialogues in terms of LLMs.
• Figure 5: Comparison of bias ratio in single versus multi-turn dialogues in terms of tasks.