Challenge: Existing studies have identified that judge models are vulnerable to certain biases, such as favoring longer answers regardless of content.
Approach: They propose to use open-source LLMs to fine-tune judge models and propose a meta-evaluation collection of test cases for each bias type.
Outcome: The proposed meta-evaluation dataset shows that fine-tuning improves the robustness of judge models against biases and improves performance across most evaluation scenarios.

Similar Papers

An Empirical Study of LLM-as-a-Judge for LLM Evaluation: Fine-tuned Judge Model is not a General Substitute for GPT-4 (2025.findings-acl)

Copied to clipboard

Challenge: Recent studies have fine-tuned judge models based on open-source LLMs to evaluate the quality of other LLM.
Approach: They propose to use open-source LLMs to evaluate Large Language Models (LLMs) their empirical results show that the models underperform GPT-4 in several dimensions .
Outcome: The proposed models outperform GPT-4 on several dimensions including generalizability, fairness and adaptability.
OpinionGPT: Modelling Explicit Biases in Instruction-Tuned LLMs (2024.naacl-demo)

Copied to clipboard

Challenge: Current research seeks to de-bias such models, or suppress potentially biased answers.
Approach: They present a web demo to test the biases of instruction-tuned Large Language Models . they identify 11 different biase based on a corpus of data .
Outcome: The proposed demo shows that biases in instruction-tuning are explicit and transparent . the demo shows how the model was trained and showcases the web application .
Open-DeBias: Toward Mitigating Open-Set Bias in Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing approaches to addressing harmful biases in LLMs are limited to predefined categories . a novel, data-efficient, and parameter-efficient debiasing method is proposed to mitigate existing social and stereotypical biase .
Approach: They propose an open-set bias detection and mitigation method to address harmful biases in text-based QA.
Outcome: The proposed method improves QA accuracy on Korean BBQ dataset by nearly 48% on ambiguous subsets and 6% on disambiguated ones.
Don’t Judge Code by Its Cover: Exploring Biases in LLM Judges for Code Evaluation (2026.findings-eacl)

Copied to clipboard

Challenge: Large language models (LLMs) are increasingly used as evaluators for code evaluation tasks . however, whether they can handle superficial variations remains unclear .
Approach: They define six types of potential biases in code evaluation and reveal their impact on LLM judges.
Outcome: The proposed method can be used to evaluate semantically equivalent code with superficial variations without reference implementations.
Benchmarking Debiasing Methods for LLM-based Parameter Estimates (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) are expensive yet powerful ways to annotate text, and can be inconsistent when compared with experts.
Approach: They propose to combine LLM annotations with a limited number of expensive expert annotations to produce valid estimates.
Outcome: The proposed methods produce consistent estimates under theoretical assumptions, but they are not comparable across finite datasets.
Humans or LLMs as the Judge? A Study on Judgement Bias (2024.emnlp-main)

Copied to clipboard

Challenge: Proprietary models such as GPT-4, Claude, Gemini-Pro and others are being democratized to improve evaluations of LLMs.
Approach: They propose a framework that is free from referencing groundtruth annotations for investigating **Misinformation Oversight Bias**, **Gender Bia**,**Authority Bia* and **Beauty Bia's** on LLM and human judges.
Outcome: The proposed framework investigates **Misinformation Oversight Bias**, **Gender Bia**,**Authority Bia* and **Beauty Bia' on LLM and human judges.
Curse of Knowledge: Your Guidance and Provided Knowledge are biasing LLM Judges in Complex Evaluation (2025.findings-emnlp)

Copied to clipboard

Challenge: a recent study has focused on simple settings, but their reliability in complex tasks remains understudied.
Approach: They propose to use large language models as judges to evaluate reliability in complex tasks . they use a challenge benchmark to expose and quantify Auxiliary Information Induced Biases .
Outcome: The proposed benchmark exposes and quantifies Auxiliary Information Induced Biases across 12 basic and 3 advanced scenarios.
Beyond the Surface: Measuring Self-Preference in LLM Judgments (2025.emnlp-main)

Copied to clipboard

Challenge: Existing methods measure self-preference bias by comparing the scores a judge model assigns to its own responses with those assigned to other models.
Approach: They propose to use gold judgments as proxies for the actual quality of responses . they propose to measure self-preference bias as the difference between the judge model's own and other models' scores .
Outcome: The proposed method can assess self-preference bias across large language models . it uses gold judgments as proxies for the ground truth scores of the judge model .
LLMs Are Biased Towards Output Formats! Systematically Evaluating and Mitigating Output Format Bias of LLMs (2025.naacl-long)

Copied to clipboard

Challenge: Using format-following capabilities, state-of-the-art large language models (LLMs) can be leveraged to tailor outputs to specific task formats.
Approach: They propose to define a format bias evaluation metric and establish effective strategies to reduce it.
Outcome: The proposed evaluation reduces the variance in ChatGPT’s performance among wrapping formats from 235.33 to 0.71 (%2)
Assistant-Guided Mitigation of Teacher Preference Bias in LLM-as-a-Judge (2025.findings-emnlp)

Copied to clipboard

Challenge: LLM-as-a-Judge uses large language models to evaluate the quality of LLM generated responses, but training proxy judge models using evaluation data generated by powerful teacher models introduces a critical yet previously overlooked issue: teacher preference bias.
Approach: They propose a new setting that incorporates an additional assistant model, which is not biased toward the teacher model’s responses, to complement the training data.
Outcome: The proposed model reduces teacher preference bias while maintaining strong performance across six evaluation benchmarks.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations