Saved in:
Bibliographic Details
Main Authors: Wang, Qian, Lou, Zhanzhi, Tang, Zhenheng, Chen, Nuo, Zhao, Xuandong, Zhang, Wenxuan, Song, Dawn, He, Bingsheng
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2504.09946
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916695550984192
author Wang, Qian
Lou, Zhanzhi
Tang, Zhenheng
Chen, Nuo
Zhao, Xuandong
Zhang, Wenxuan
Song, Dawn
He, Bingsheng
author_facet Wang, Qian
Lou, Zhanzhi
Tang, Zhenheng
Chen, Nuo
Zhao, Xuandong
Zhang, Wenxuan
Song, Dawn
He, Bingsheng
contents Large Reasoning Models (LRMs) like DeepSeek-R1 and OpenAI-o1 have demonstrated remarkable reasoning capabilities, raising important questions about their biases in LLM-as-a-judge settings. We present a comprehensive benchmark comparing judging biases between LLMs and LRMs across both subjective preference-alignment datasets and objective fact-based datasets. Through investigation of bandwagon, authority, position, and distraction biases, we uncover four key findings: (1) despite their advanced reasoning capabilities, LRMs remain susceptible to the above biases; (2) LRMs demonstrate better robustness than LLMs specifically on fact-related datasets; (3) LRMs exhibit notable position bias, preferring options in later positions; and (4) we identify a novel "superficial reflection bias" where phrases mimicking reasoning (e.g., "wait, let me think...") significantly influence model judgments. To address these biases, we design and evaluate three mitigation strategies: specialized system prompts that reduce judging biases by up to 19\% in preference alignment datasets and 14\% in fact-related datasets, in-context learning that provides up to 27\% improvement on preference tasks but shows inconsistent results on factual tasks, and a self-reflection mechanism that reduces biases by up to 10\% in preference datasets and 16\% in fact-related datasets, with self-reflection proving particularly effective for LRMs. Our work provides crucial insights for developing more reliable LLM-as-a-Judge frameworks, especially as LRMs become increasingly deployed as automated judges.
format Preprint
id arxiv_https___arxiv_org_abs_2504_09946
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Assessing Judging Bias in Large Reasoning Models: An Empirical Study
Wang, Qian
Lou, Zhanzhi
Tang, Zhenheng
Chen, Nuo
Zhao, Xuandong
Zhang, Wenxuan
Song, Dawn
He, Bingsheng
Computers and Society
Computation and Language
Large Reasoning Models (LRMs) like DeepSeek-R1 and OpenAI-o1 have demonstrated remarkable reasoning capabilities, raising important questions about their biases in LLM-as-a-judge settings. We present a comprehensive benchmark comparing judging biases between LLMs and LRMs across both subjective preference-alignment datasets and objective fact-based datasets. Through investigation of bandwagon, authority, position, and distraction biases, we uncover four key findings: (1) despite their advanced reasoning capabilities, LRMs remain susceptible to the above biases; (2) LRMs demonstrate better robustness than LLMs specifically on fact-related datasets; (3) LRMs exhibit notable position bias, preferring options in later positions; and (4) we identify a novel "superficial reflection bias" where phrases mimicking reasoning (e.g., "wait, let me think...") significantly influence model judgments. To address these biases, we design and evaluate three mitigation strategies: specialized system prompts that reduce judging biases by up to 19\% in preference alignment datasets and 14\% in fact-related datasets, in-context learning that provides up to 27\% improvement on preference tasks but shows inconsistent results on factual tasks, and a self-reflection mechanism that reduces biases by up to 10\% in preference datasets and 16\% in fact-related datasets, with self-reflection proving particularly effective for LRMs. Our work provides crucial insights for developing more reliable LLM-as-a-Judge frameworks, especially as LRMs become increasingly deployed as automated judges.
title Assessing Judging Bias in Large Reasoning Models: An Empirical Study
topic Computers and Society
Computation and Language
url https://arxiv.org/abs/2504.09946