Mitigating Judgment Preference Bias in Large Language Models through Group-Based Polling

TL;DR

The paper addresses judgment preference bias in LLM-based evaluators and proposes Genii, an unsupervised, group-based polling framework that organizes multiple judgment models into a rotating client–server multi-agent system. Through group-consistency scoring and Direct Preference Optimization (DPO), Genii aligns individual models with the group's collective preferences without requiring labeled data. Empirical results show Genii outperforms supervised, annotation-heavy baselines and remains effective even when weaker models serve as servers, illustrating robust bias mitigation and cross-model generalization. This approach offers a scalable, annotation-free pathway to more reliable LLM evaluations and model alignment in practical settings.

Abstract

Large Language Models (LLMs) as automatic evaluators, commonly referred to as LLM-as-a-Judge, have also attracted growing attention. This approach plays a vital role in aligning LLMs with human judgments, providing accurate and reliable assessments. However, LLM-based judgment models often exhibit judgment preference bias during the evaluation phase, tending to favor responses generated by themselves, undermining the reliability of their judgments. This paper introduces the Group-Based Polling Optimization (Genii), an unsupervised multi-agent collaborative optimization framework that mitigates the inherent judgment preference bias of judgment models. Specifically, Genii integrates various LLM-based judgment models into a multi-agent system and simulates the interactive client-server polling mechanism to optimize each client agent unsupervisedly. Our experiments demonstrate that Genii outperforms supervised models trained on annotated judgment data, while requiring no human-labeled annotations. Genii consistently improves performance across different client agents during the polling, even when weaker models act as server agents. Further analysis reveals that Genii effectively mitigates judgment preference bias of LLM-based judgment models, demonstrating its effectiveness. All codes are available at https://github.com/NEUIR/Genii.

Figures (10)

• Figure 1: The Framework of Genii. It alleviates judgment preference of LLMs during the evaluation process.
• Figure 2: The Framework of Our Genii Method.
• Figure 3: Correlation between Judge Accuracy and Task Accuracy. Each circle represents one judgment model, with the color denoting model family.
• Figure 4: Perplexity Scores of Vanilla LLMs and Genii on the Incorrect Answers Generated by Vanilla LLMs.
• Figure 5: Embedding Visualization of Judgments. We randomly select a sample from the evaluation datasets and then use t-SNE to visualize the embeddings of judgments generated by the Vanilla LLM and Genii.