@inproceedings{ducel-etal-2026-benchmarks,
title = "Are Benchmarks a Source of Bias in {LLM}s? {MMLU}, {C}ommon{S}ense{QA} and {MGSM} under the Microscope",
author = {Ducel, Fanny and
Shariff, Shayan Ahmed and
Al Kotob, Ibrahim and
Arakkal Remesh, Binesh and
Digoin-Caparros, Lucie and
N{\'e}v{\'e}ol, Aur{\'e}lie and
Fort, Kar{\"e}n},
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.768/",
pages = "17826--17841",
ISBN = "979-8-89176-459-0",
abstract = "Evaluation is an essential process in the development and use of NLP systems. Benchmarks are widely used to calibrate systems, determine the efficiency of techniques, but also to advertise recently launched language models. Not only do benchmarks allow for evaluation and system comparisons, but they are also commonly utilized as training and validation corpora. In this paper, we aim at characterizing the ``world-knowledge'' that is put forth by benchmarks by closely inspecting their content. Through qualitative manual annotations (+15k instances) as well as quantitative automatic analyses, we uncover the presence of both explicit and implicit biases in MMLU, CommonSenseQA and MGSM, ranging from representational imbalance to bluntly sexist and racist instances. Our findings demonstrate that current NLP evaluation practices pose numerous quality issues, reinforce stereotypes and even contain explicitly harmful content. The current operationalization of benchmarks is questioned, and we provide recommendations and directions for further ethical evaluation. Warning: this paper contains content that may be offensive or upsetting."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ducel-etal-2026-benchmarks">
<titleInfo>
<title>Are Benchmarks a Source of Bias in LLMs? MMLU, CommonSenseQA and MGSM under the Microscope</title>
</titleInfo>
<name type="personal">
<namePart type="given">Fanny</namePart>
<namePart type="family">Ducel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shayan</namePart>
<namePart type="given">Ahmed</namePart>
<namePart type="family">Shariff</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ibrahim</namePart>
<namePart type="family">Al Kotob</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Binesh</namePart>
<namePart type="family">Arakkal Remesh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lucie</namePart>
<namePart type="family">Digoin-Caparros</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aurélie</namePart>
<namePart type="family">Névéol</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Karën</namePart>
<namePart type="family">Fort</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Evaluation is an essential process in the development and use of NLP systems. Benchmarks are widely used to calibrate systems, determine the efficiency of techniques, but also to advertise recently launched language models. Not only do benchmarks allow for evaluation and system comparisons, but they are also commonly utilized as training and validation corpora. In this paper, we aim at characterizing the “world-knowledge” that is put forth by benchmarks by closely inspecting their content. Through qualitative manual annotations (+15k instances) as well as quantitative automatic analyses, we uncover the presence of both explicit and implicit biases in MMLU, CommonSenseQA and MGSM, ranging from representational imbalance to bluntly sexist and racist instances. Our findings demonstrate that current NLP evaluation practices pose numerous quality issues, reinforce stereotypes and even contain explicitly harmful content. The current operationalization of benchmarks is questioned, and we provide recommendations and directions for further ethical evaluation. Warning: this paper contains content that may be offensive or upsetting.</abstract>
<identifier type="citekey">ducel-etal-2026-benchmarks</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.768/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17826</start>
<end>17841</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Are Benchmarks a Source of Bias in LLMs? MMLU, CommonSenseQA and MGSM under the Microscope
%A Ducel, Fanny
%A Shariff, Shayan Ahmed
%A Al Kotob, Ibrahim
%A Arakkal Remesh, Binesh
%A Digoin-Caparros, Lucie
%A Névéol, Aurélie
%A Fort, Karën
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F ducel-etal-2026-benchmarks
%X Evaluation is an essential process in the development and use of NLP systems. Benchmarks are widely used to calibrate systems, determine the efficiency of techniques, but also to advertise recently launched language models. Not only do benchmarks allow for evaluation and system comparisons, but they are also commonly utilized as training and validation corpora. In this paper, we aim at characterizing the “world-knowledge” that is put forth by benchmarks by closely inspecting their content. Through qualitative manual annotations (+15k instances) as well as quantitative automatic analyses, we uncover the presence of both explicit and implicit biases in MMLU, CommonSenseQA and MGSM, ranging from representational imbalance to bluntly sexist and racist instances. Our findings demonstrate that current NLP evaluation practices pose numerous quality issues, reinforce stereotypes and even contain explicitly harmful content. The current operationalization of benchmarks is questioned, and we provide recommendations and directions for further ethical evaluation. Warning: this paper contains content that may be offensive or upsetting.
%U https://aclanthology.org/2026.emnlp-main.768/
%P 17826-17841
Markdown (Informal)
[Are Benchmarks a Source of Bias in LLMs? MMLU, CommonSenseQA and MGSM under the Microscope](https://aclanthology.org/2026.emnlp-main.768/) (Ducel et al., EMNLP 2026)
ACL
- Fanny Ducel, Shayan Ahmed Shariff, Ibrahim Al Kotob, Binesh Arakkal Remesh, Lucie Digoin-Caparros, Aurélie Névéol, and Karën Fort. 2026. Are Benchmarks a Source of Bias in LLMs? MMLU, CommonSenseQA and MGSM under the Microscope. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 17826–17841, Budapest, Hungary. Association for Computational Linguistics.