Datasets:

matybohacek
/

benchmarking-cultures-25

benchmark_id stringlengths 2 22	benchmark_name stringlengths 2 26	benchmark_full_name stringlengths 12 71 ⌀	paper_id stringlengths 4 22 ⌀	paper_href stringlengths 27 84 ⌀	paper_published_at stringdate 2015-04-19 00:00:00 2025-12-03 00:00:00 ⌀
agi-eval	AGIEval	null	2304.06364	https://arxiv.org/abs/2304.06364	2023-04-13
ai2-arc	ARC	AI2 Reasoning Challenge	1803.05457	https://arxiv.org/abs/1803.05457	2018-03-14
ai2d	AI2D	AI2 Diagrams	1603.07396	https://arxiv.org/abs/1603.07396	2016-03-24
aider-polyglot	Aider Polyglot	null	null	null	null
aime-24	AIME 2024	American Invitational Mathematics Examination 2024	null	null	null
aime-25	AIME 2025	American Invitational Mathematics Examination 2025	null	null	null
aitz	Android In The Zoo	null	2403.02713	https://arxiv.org/abs/2403.02713	2024-03-05
alpaca-eval	AlpacaEval	null	2404.04475	https://arxiv.org/abs/2404.04475	2024-04-06
amc	AMC	American Mathematics Competitions	null	null	null
android-control	Android Control	null	2406.03679	https://arxiv.org/abs/2406.03679	2024-06-06
android-world	Android World	null	2405.14573	https://arxiv.org/abs/2405.14573	2024-05-23
arc-agi-2	ARC-AGI-2	Abstraction and Reasoning Corpus 2	2505.11831	https://arxiv.org/abs/2505.11831	2025-05-17
arc-challenge	ARC-AGI-1	Abstraction and Reasoning Corpus	1911.01547	https://arxiv.org/abs/1911.01547	2019-11-05
arena-hard-auto	Arena-Hard-Auto	null	2406.11939	https://arxiv.org/abs/2406.11939	2024-06-17
asdiv	ASDIV	Academia Sinica Diverse MWP Dataset	2106.15772	https://arxiv.org/abs/2106.15772	2021-06-30
basic-skills	Basic Skills	null	null	null	null
bbh	BigBenchHard	null	2210.09261	https://arxiv.org/abs/2210.09261	2022-10-17
bcb	BigCodeBench	null	2406.15877	https://arxiv.org/abs/2406.15877	2024-06-22
bfcl	BFCL	Berkeley Function-Calling Leaderboard	bfcl	null	null
bfcl-3	BFCL 3	Berkeley Function-Calling Leaderboard V3	bfcl-3	https://gorilla.cs.berkeley.edu/blogs/13_bfcl_v3_multi_turn.html	2024-09-19
bird-sql	Bird-SQL (Dev)	null	2305.03111	https://arxiv.org/abs/2305.03111	2023-05-04
blink	BLINK	null	2404.12390	https://arxiv.org/abs/2404.12390	2024-04-18
browse-comp	BrowseComp	null	2504.12516	https://arxiv.org/abs/2504.12516	2025-04-16
browse-comp-plus	BrowseComp-Plus	null	2508.06600	https://arxiv.org/abs/2508.06600	2025-08-08
browse-comp-zh	BrowseComp-ZH	null	2504.19314	https://arxiv.org/abs/2504.19314	2025-05-01
c-eval	C-Eval	null	2305.08322	https://arxiv.org/abs/2305.08322	2023-05-15
c-math	CMath	null	2306.16636	https://arxiv.org/abs/2306.16636	2023-06-29
c-simple-qa	ChineseSimpleQA	null	2411.07140	https://arxiv.org/abs/2411.07140	2024-11-11
c3	C3	null	2507.22968	https://arxiv.org/abs/2507.22968	2025-07-30
cc-bench	CCBench	null	cc-bench	https://www.nature.com/articles/s41698-025-00916-7	2025-05-03
cc-ocr	CC-OCR	null	2412.02210	https://arxiv.org/abs/2412.02210	2024-12-03
ccpm	CCPM	null	2106.01979	https://arxiv.org/abs/2106.01979	2021-06-03
cfb	ComplexFuncBench	null	2501.10132	https://arxiv.org/abs/2501.10132	2025-01-17
charades-sta	Charades-STA	null	1705.02101	https://arxiv.org/abs/1705.02101	2017-05-05
chart-qa	ChartQA	null	2203.10244	https://arxiv.org/abs/2203.10244	2022-03-19
charvix	CharXiv-Reasoning	null	2406.18521	https://arxiv.org/abs/2406.18521	2024-06-26
chinese-word	Chinese Word	null	2508.02324	https://arxiv.org/abs/2508.02324	2025-08-04
clasi	CLASI	null	2407.21646	https://arxiv.org/abs/2407.21646	2024-07-31
clue-wsc	ClueWSC	null	null	null	null
cm17k	CM17K	null	2107.01431	https://arxiv.org/abs/2107.01431	2021-07-03
cmmlu	CMMLU	null	2306.09212	https://arxiv.org/abs/2306.09212	2023-06-15
cmrc	CMRC	Chinese Machine Reading Comprehension	cmrc	https://aclanthology.org/D19-1600/	2019-11-03
cmteb	Chinese MTEB	Chinese Massive Text Embedding Benchmark	2309.07597	https://arxiv.org/abs/2309.07597	2023-09-14
cnmo	CNMO	China National Mathematical Olympiad	null	null	null
codeforces	Codeforces	null	2501.01257	https://arxiv.org/abs/2501.01257	2025-01-02
collie	COLLIE	null	2307.08689	https://arxiv.org/abs/2307.08689	2023-07-17
coqa	ConversationalQA	null	1808.07042	https://arxiv.org/abs/1808.07042	2018-08-21
count-bench	CountBench	null	2302.12066	https://arxiv.org/abs/2302.12066	2023-02-23
covost2	CoVoST2 (21 lang)	null	2002.01320	https://arxiv.org/abs/2002.01320	2020-02-04
creative-writing	Creative Writing	null	creative-writing	https://github.com/EQ-bench/creative-writing-bench	2025-03-28
crux-eval	CRUXEval	null	2401.03065	https://arxiv.org/abs/2401.03065	2024-01-05
cs-qa	CommonsenseQA	null	1811.00937	https://arxiv.org/abs/1811.00937	2018-11-02
cv-15	Common Voice 15	null	1912.06670	https://arxiv.org/abs/1912.06670	2019-12-13
cv-bench	CVBench	Cross Video Bench	2508.19542	https://arxiv.org/abs/2508.19542	2025-08-27
deepmind-math	Deepmind Math	null	1904.01557	https://arxiv.org/abs/1904.01557	2019-04-02
deepseek-leetcode	DeepSeek Leetcode	null	deepseek-leetcode	https://github.com/deepseek-ai/DeepSeek-Coder/commits/main/Evaluation/LeetCode	2024-01-26
doc-vqa	DocVQA	Document Visual Question Answering	2007.00398	https://arxiv.org/abs/2007.00398	2020-07-01
dpg	DPG Bench	Dense Prompt Graph Benchmark	2403.05135	https://arxiv.org/abs/2403.05135	2024-03-08
drop	DROP	Discrete Reasoning Over Paragraphs	1903.00161	https://arxiv.org/abs/1903.00161	2019-03-01
ds-1000	DS-1000	null	2211.11501	https://arxiv.org/abs/2211.11501	2022-11-18
ego-schema	EgoSchema	null	2308.09126	https://arxiv.org/abs/2308.09126	2023-08-17
eq-bench	EQ-Bench	null	2312.06281	null	null
erqa	ERQA	null	2503.20020	https://arxiv.org/abs/2503.20020	2025-03-25
eval-plus	EvalPlus	null	2305.01210	https://arxiv.org/abs/2305.01210	2023-05-02
fact-score	FActScore	null	2305.14251	https://arxiv.org/abs/2305.14251	2023-05-23
facts	FACTS Grounding	null	2501.03200	https://arxiv.org/abs/2501.03200	2025-01-06
finance-agent	Finance Agent	null	2508.00828	https://arxiv.org/abs/2508.00828	2025-05-20
fleurs	FLEURS	null	2205.12446	https://arxiv.org/abs/2205.12446	2022-05-25
follow-ir	FollowIR	null	2403.15246	https://arxiv.org/abs/2403.15246	2024-03-22
frames	FRAMES	Factuality, Retrieval, And reasoning MEasurement Set	2409.12941	https://arxiv.org/abs/2409.12941	2024-09-19
frontier-math	FrontierMath	null	2411.04872	https://arxiv.org/abs/2411.04872	2024-11-07
full-stack-bench	FullStackBench	null	2412.00535	https://arxiv.org/abs/2412.00535	2024-11-30
gdp-val	GDPval	null	2510.04374	https://arxiv.org/abs/2510.04374	2025-10-05
gedit-bench	GEdit Bench	null	2504.17761	https://arxiv.org/abs/2504.17761	2025-04-24
gen-eval	GenEval	null	2310.11513	https://arxiv.org/abs/2310.11513	2023-10-17
global-mmlu	Global MMLU (Lite)	null	2412.03304	https://arxiv.org/abs/2412.03304	2024-12-04
global-piqa	Global PIQA	null	2510.24081	https://arxiv.org/abs/2510.24081	2025-10-28
gpqa-diamond	GPQA Diamond	Graduate-Level Google-Proof Q&A Benchmark Diamond	2311.12022	https://arxiv.org/abs/2311.12022	2023-11-20
gpqa-main	GPQA Main	Graduate-Level Google-Proof Q&A Benchmark Main	2311.12022	https://arxiv.org/abs/2311.12022	2023-11-20
gpqa-science	GPQA Science	Graduate-Level Google-Proof Q&A Benchmark Science	2311.12022	https://arxiv.org/abs/2311.12022	2023-11-20
graph-walks	GraphWalks	null	graph-walks	https://openai.com/index/gpt-4-1/	2025-04-14
gsm-symbolic	GSM Symbolic	null	2410.05229	https://arxiv.org/abs/2410.05229	2024-10-07
gsm8k	GSM8k	null	2110.14168	https://arxiv.org/abs/2110.14168	2021-10-27
gso	GSO	null	2505.23671	https://arxiv.org/abs/2505.23671	2025-05-29
gtzan	GTZAN	null	null	null	null
hallusion-bench	HallusionBench	null	2310.14566	https://arxiv.org/abs/2310.14566	2023-10-23
health-bench-hard	HealthBench Hard	null	2505.08775	https://arxiv.org/abs/2505.08775	2025-05-13
heatlh-bench	HealthBench	null	2505.08775	https://arxiv.org/abs/2505.08775	2025-05-13
hellaswag	HellaSwag	null	1905.07830	https://arxiv.org/abs/1905.07830	2019-05-19
hidden-math	HiddenMath	null	null	null	null
hle	Humanity's Last Exam	null	2501.14249	https://arxiv.org/abs/2501.14249	2025-01-24
hmmt	HMMT 2025	null	null	null	null
hr-bench	HR-Bench	null	2408.15556	https://arxiv.org/abs/2408.15556	2024-08-28
human-eval	HumanEval	null	2107.03374	https://arxiv.org/abs/2107.03374	2021-07-07
human-eval-plus	HumanEval+	null	2305.01210	https://arxiv.org/abs/2305.01210	2023-05-02
if-bench	IF Bench	null	2507.02833	https://arxiv.org/abs/2507.02833	2025-07-03
if-eval	IF Eval	null	2311.07911	https://arxiv.org/abs/2311.07911	2023-11-14
img-edit	ImgEdit	null	2505.20275	https://arxiv.org/abs/2505.20275	2025-05-26
imo-answer-bench	IMOAnswerBench	null	2511.01846	https://arxiv.org/abs/2511.01846	2025-11-03
include	INCLUDE	null	2411.19799	https://arxiv.org/abs/2411.19799	2024-11-29

End of preview. Expand in Data Studio

Benchmarking-Cultures-25 Dataset

This dataset accompanies the Unsteady Metrics and Benchmarking Cultures of AI Model Builders paper submitted to FAccT 2026 by Stefan Baack, Christo Buschek and Maty Bohacek.

The dataset contains the following parts:

core: The curated Benchmarking-Cultures-25 dataset.
derived: Datasets that were derived from the core dataset and informed the FAccT submission.
figures: Figures generated from derived data and used in the paper.
docs: Data dictionaries describing the core data files.
code: Scripts used for the automated parts of the data curation.

Additionally, an interactive UI is available at https://bench-cultures.net, where the data can introspected.

Data curation

The data was collected with the following process. All steps were done manually unless otherwise noted.

Based on the scope of our study, we included 11 model publishers (see Section 3 of the paper for selection methodology). Each entity was manually profiled through desk research to determine geographic headquarters, primary domain, and institutional affiliation.
We monitored official corporate blogs, technical reports, and primary dissemination channels (Hacker News) to identify new model releases. A benchmark was recorded as "explicitly highlighted" only if it met a strict evidence threshold: it must be cited by name with associated performance metrics presented in a primary table or figure.
New benchmarks were integrated into a master table. For existing benchmarks, we recorded incremental highlights.
For every identified benchmark, we isolated and recorded the authoritative primary source (typically, a peer-reviewed paper or arXiv pre-print). We explicitly noted cases where benchmarks (e.g., specific math competitions) lacked traditional academic release papers to maintain data transparency.
We fetched the list of authors from the ArXiv API for every paper that was found on ArXiv using an automated script, which is included in this repository. We extracted the list of all authors for every benchmark from these API responses. For the papers that were not published on ArXiv we manually extracted the authors from their respective papers and added them to the list as well.
To focus our analysis on those driving the research agenda and benchmark production, we isolated core contributors, manually excluding broad-scale technical staff or distal advisors where such distinctions were explicitly noted in the source acknowledgments.
For every author we extracted the affiliation as stated in the paper and associated the institution with a country, a domain and one of five sectors based on desk research. This step was skipped if no clear affiliation could be determined on the paper alone.
In a final qualitative pass, we conducted a close reading of the authoritative papers to develop an inductive taxonomy of competencies. Benchmarks were mapped to one or more categories based strictly on the claims of the original authors regarding what the metric is intended to evaluate.

Quality Assurance

To ensure the computational integrity of all rankings, statistics, and visualizations, we implemented a blind cross-verification protocol. Independent team members were tasked with reimplementing the data transformation/analysis scripts using only the data in the core folder.

This replication process ensures that the artifacts in the derived folder are fully reproducible and free from errors. A final parity audit is underway to ensure a complete alignment between the internal and public release version.

`core` data files

models.csv

The list of all model releases captured in 2025.

model_id: Unique identifier for the model (slug).
model_name: The display name of the model.
model_family: The name of the model family, e.g. Gemini or DeepSeek
model_version: The version of the model, e.g. 2.5 or V3.1.
model_variant: The variant of the model, e.g. Flash or Terminus.
model_subvariant: A subvariant of the model, e.g Lite.
model_is_base: A flag indicating if the model is a base model.
model_total_parameters: The number of total parameters of the model.
model_active_parameters: The number of active parameters of the model.
model_href: URL to the model's press release or blog post.
model_published_at: The date the model was released.
model_access: The access level of the model. Options: Closed, Open-Weight or Open-Source.
model_has_highlight: A flag indicating if the model has any benchmark highlights in it's release announcement.
organization_name: The name of the organization releasing this model.
organization_sector: The sector of the organization. Options: Industry, Academia or Non-Profit.
organization_country: The country of origin of this organization.
organization_domain: The domain of influence this organization belongs to. Options: China or West.

benchmarks.csv

The list of all benchmarks that were highlighted in the model's press release or blog post (model_href).

benchmark_id: Unique identifier for the benchmark (slug).
benchmark_name: The display name of the benchmark.
paper_id: Unique identifier for the paper announcing the benchmark (ArXiv ID or custom slug).
paper_href: URL to the paper announcing the benchmark.
paper_published_at: The date the paper was published. This was taken as the benchmark release date (version 1 if more than one was provided).

highlights.csv

A join table connecting models and highlighted benchmarks.

benchmark_id: Unique identifier for the benchmark (slug).
model_id: Unique identifier for the model (slug).
prescribed_competency: The competency that model builders prescribe to this benchmark for that model release. This field remains empty if the model builder didn't assign a competency but highlighted the benchmark anyway.
prescribed_category: A generalized categorization of the prescribe_competency.

affiliations.csv

Connect benchmark releases (papers) with their authors and associate them with their institutional affiliations at the time of the paper release. In cases where a distinction was made between author/core-contributor and another role, only the former were included.

paper_id: Unique identifier for the release paper (ArXiv ID or custom slug).
author_name: The name of an author for this paper.
organization_name: The name of the organization affiliated with the author.
organization_sector: The sector of the organization. Options: Industry, Academia or Non-Profit.
organization_country: The country of origin of this organization.
organization_domain: The domain of influence this organization belongs to. Options: China or West.

categories.csv

The taxonomy of benchmark competencies developed by the authors, i.e. a list of categories (competencies) that the authors assigned to benchmarks.

benchmark_category: Granular functional classification. Options: Audio-visual pattern recognition, Audio-visual understanding, Base model capabilities, Coding, Commonsense, Embodied spatial understanding, Factuality, Generic, Health, Instruction following, Instruction retention, Long-context, Math, Multilingual performance, Multimodal generation, Reasoning and knowledge, Rule adherence, Semantic search, Strategic planning, Tool orchestration, Translation or Writing style.
benchmark_meta_category: High-level classification of the benchmark. Options: Agentic task execution, Foundational capabilities, General knowledge application, Information retrieval, Multilingual capabilities, Multimodal processing, Preference-Alignment or Specialized knowledge application.
benchmark_category_definition: A description of the meaning for the category.

categorizations.csv

The assigned category for every benchmark. Some benchmarks have more than one category assigned.

benchmark_id: Unique identifier for the benchmark (slug).
benchmark_category: Granular functional classification. Options: Audio-visual pattern recognition, Audio-visual understanding, Base model capabilities, Coding, Commonsense, Embodied spatial understanding, Factuality, Generic, Health, Instruction following, Instruction retention, Long-context, Math, Multilingual performance, Multimodal generation, Reasoning and knowledge, Rule adherence, Semantic search, Strategic planning, Tool orchestration, Translation or Writing style.

knowledge-subjects.csv

A breakdown of the subjects covered in the top 15 benchmarks assigned to "Reasoning and knowledge" (GPQA Diamond, Humanity's Last Exam, MMLU, MMLU-Pro and MMMU).

benchmark_id: Unique identifier for the benchmark (slug).
subject: The subject as it was named in the benchmark data set.
field: A mapping of the subject to a field. Options: Art & Design, Business, Health & Medicine, Humanities & Social Sciences, Law, Science, Tech & Engineering or nil.
science_discipline: A mapping of the science field to a concrete discipline.
n: The number of questions related to this subject in the benchmark.
p: The percentage of questions related to this subject in the benchmark.

`derived` data files

affiliations-benchmark-creators-multiple.csv

Table 2. Distribution of Multiple Affiliations of Benchmark Creators. Shows the distribution of combinations of the different affiliations for benchmark creators, that have more than one affiliation.

combination: The collapsed combination of benchmark creator affiliation.
n: The number of benchmark creators for this combination.
p: The percentage of benchmark creators for this combination.

affiliations-benchmark-creators.csv

Table 1. Affiliation of Benchmark Creators. Creators’ affiliations are categorized into organization categories. A breakdown is provided for benchmarks published in 2025 and all benchmarks present in the dataset (labeled as "Overall").

organization_sector: The sector of the organization. Options: Industry, Academia, Non-Profit, Government, Independent or nil.
n_overall: The number of benchmark authors overall affiliated with this sector.
p_overall: The percentage of benchmark authors overall affiliated with this sector.
n_overall_2025: The number of authors of benchmarks released in 2025 overall affiliated with this sector.
p_overall_2025: The percentage of authors of benchmarks released in 2025 overall affiliated with this sector.
n_west: The number of benchmark authors from the West affiliated with this sector.
p_west: The percentage of benchmark authors from the West affiliated with this sector.
n_west_2025: The number of authors from the West of benchmarks released in 2025 affiliated with this sector.
p_west_2025: The percentage of authors from the West of benchmarks released in 2025 affiliated with this sector.
n_china: The number of benchmark authors from China affiliated with this sector.
p_china: The percentage of benchmark authors from China affiliated with this sector.
n_china_2025: The number of authors from China of benchmarks released in 2025 affiliated with this sector.
p_china_2025: The percentage of authors from China of benchmarks released in 2025 affiliated with this sector.

affiliations-benchmark-wide.csv

Breakdown of the author affiliations for every benchmark. Same data as in affiliations-benchmark.csv but as a wide table with each sector as it's own column.

benchmark_id: Unique identifier for the benchmark (slug).
benchmark_name: The display name of the benchmark.
benchmark_published_at: The date the benchmark was published. This is based on the release date of the associated paper.
p_industry: The percentage of all authors affiliated with industry for benchmark_id. p_non-profit
p_academia: The percentage of all authors affiliated with academia for benchmark_id.
p_independent: The percentage of all independent authors for benchmark_id.
p_government: The percentage of all authors affiliated with a government for benchmark_id.
p_na: The percentage of all authors for benchmark_id where no affiliation could be determined.

affiliations-benchmark.csv

Breakdown of the author affiliations for every benchmark.

benchmark_id: Unique identifier for the benchmark (slug).
benchmark_name: The display name of the benchmark.
benchmark_published_at: The date the benchmark was published. This is based on the release date of the associated paper.
organization_sector: The sector of the organization. Options: Industry, Academia or Non-Profit.
n: The number of benchmark authors affiliated with benchmark_id and organization_sector.
p: The percentage of benchmark authors affiliated with benchmark_id and organization_sector.

benchmark-adoption.csv

Table 6: Distribution of Benchmark Adoption. The percentage of model builders and models that highlight a benchmark X number of times.

highlights: The number of times a benchmark was highlighted.
n_publishers: The number of model builders highlighting a benchmark X number of times.
p_publishers: The percentage of model builders highlighting a benchmark X number of times.
n_publishers_china: The number of model builders in China highlighting a benchmark X number of times.
p_publishers_china: The percentage of model builders in China highlighting a benchmark X number of times.
n_publishers_west: The number of model builders in the West highlighting a benchmark X number of times.
p_publishers_west: The percentage of model builders in the West highlighting a benchmark X number of times.
n_models: The number of models highlighting a benchmark X number of times.
p_models: The percentage of models highlighting a benchmark X number of times.
n_models_closed: The number of closed models highlighting a benchmark X number of times.
p_models_closed: The percentage of closed models highlighting a benchmark X number of times.
n_models_open-source: The number of open source models highlighting a benchmark X number of times.
p_models_open-source: The percentage of open source models highlighting a benchmark X number of times.
n_models_open-weight: The number of open weight models highlighting a benchmark X number of times.
p_models_open-weight: The percentage of open weight models highlighting a benchmark X number of times.

benchmark-creators.csv

RQ1: Counts of the total number of benchmark authors compared to the number of authors with multiple affiliations.

n_benchmark_creators: The number of all benchmark creators.
n_benchmark_creators_multiple: The number of all benchmark creators with multiple affiliations.
p_benchmark_creators_multiple: The percentage of all benchmark creators with multiple affiliations.

benchmark-released-by-year.csv

RQ4: Publication year of benchmarks highlighted in 2025.

year: The year of the release.
n: The number of benchmarks released in that year.
p: The percentage of benchmarks released in that year.

evaluated-competencies-by-year.csv

*Table 5 and Table 19: Yearly Benchmark Releases by Selected Tested Competencies.

benchmark_meta_category: High-level classification of the benchmark. Options: Agentic task execution, Formalized comprehension & reasoning, Information retrieval, Multilingual capabilities, Multimodal processing, Preference-Alignment, Self-contained foundational capabilities, Unstructured comprehension & reasoning.
benchmark_category: Granular functional classification. Options: Audio-visual pattern recognition, Audio-visual understanding, Coding, Commonsense, Embodied spatial understanding, Factuality, Foundational skills, Generic, Health, Instruction following, Instruction retention, Long-context, Math, Multilingual performance, Multimodal generation, Reasoning and knowledge, Rule adherence, Semantic search, Strategic planning, Tool orchestration, Translation or Writing style.
n_2014: The number of benchmarks with the assigned competency for 2014.
n_2015: The number of benchmarks with the assigned competency for 2015.
n_2016: The number of benchmarks with the assigned competency for 2016.
n_2017: The number of benchmarks with the assigned competency for 2017.
n_2018: The number of benchmarks with the assigned competency for 2018.
n_2019: The number of benchmarks with the assigned competency for 2019.
n_2020: The number of benchmarks with the assigned competency for 2020.
n_2021: The number of benchmarks with the assigned competency for 2021.
n_2022: The number of benchmarks with the assigned competency for 2022.
n_2023: The number of benchmarks with the assigned competency for 2023.
n_2024: The number of benchmarks with the assigned competency for 2024.
n_2025: The number of benchmarks with the assigned competency for 2025.
n_NA: The number of benchmarks with the assigned competency where the publication year is unknown.
p_2014: The percentage of benchmarks with the assigned competency for 2014.
p_2015: The percentage of benchmarks with the assigned competency for 2015.
p_2016: The percentage of benchmarks with the assigned competency for 2016.
p_2017: The percentage of benchmarks with the assigned competency for 2017.
p_2018: The percentage of benchmarks with the assigned competency for 2018.
p_2019: The percentage of benchmarks with the assigned competency for 2019.
p_2020: The percentage of benchmarks with the assigned competency for 2020.
p_2021: The percentage of benchmarks with the assigned competency for 2021.
p_2022: The percentage of benchmarks with the assigned competency for 2022.
p_2023: The percentage of benchmarks with the assigned competency for 2023.
p_2024: The percentage of benchmarks with the assigned competency for 2024.
p_2025: The percentage of benchmarks with the assigned competency for 2025.
p_NA: The percentage of benchmarks with the assigned competency where the publication year is unknown.

evaluated-competencies.csv

Table 3. Distribution of Evaluated Competencies in the Top 15 Most Popular Benchmarks.

benchmark_category: Granular functional classification. Options: Audio-visual pattern recognition, Audio-visual understanding, Coding, Commonsense, Embodied spatial understanding, Factuality, Foundational skills, Generic, Health, Instruction following, Instruction retention, Long-context, Math, Multilingual performance, Multimodal generation, Reasoning and knowledge, Rule adherence, Semantic search, Strategic planning, Tool orchestration, Translation or Writing style.
benchmarks: The list of benchmarks for this category, separated by a comma. All benchmarks in the "Reasoning and knowledge" category are also used to evaluate "Math" competency. Hence, they are listed twice.
n: The number of benchmark competencies for this category.
p: The percentage of benchmark competencies for this category.

evaluated-meta-categories.csv

Distribution of Evaluated Meta Categories in the Top 15 Most Popular Benchmarks.

benchmark_meta_category: High-level classification of the benchmark. Options: Agentic task execution, Formalized comprehension & reasoning, Information retrieval, Multilingual capabilities, Multimodal processing, Preference-Alignment, Self-contained foundational capabilities, Unstructured comprehension & reasoning.
benchmarks: The list of benchmarks for this category, separated by a comma. All benchmarks in the "Reasoning and knowledge" category are also used to evaluate "Math" competency. Hence, they are listed twice.
n: The number of benchmark competencies for this category.
p: The percentage of benchmark competencies for this category.

highlights-2025.csv

Figure 2: Highlights of benchmarks released in 2025.

benchmark_id: Unique identifier for the benchmark (slug).
model_id: Unique identifier for the model (slug).
paper_published_at: The date the paper was published. This was taken as the benchmark release date (version 1 if more than one was provided).
model_published_at: The date the model was released.

highlighted-competencies-by-month.csv

Figure 3: Highlights of selected competencies by model builders.

month: The model publication date rounded to the nearest rounded month for this highlight.
benchmark_category: Granular functional classification. Options: Audio-visual pattern recognition, Audio-visual understanding, Coding, Commonsense, Embodied spatial understanding, Factuality, Foundational skills, Generic, Health, Instruction following, Instruction retention, Long-context, Math, Multilingual performance, Multimodal generation, Reasoning and knowledge, Rule adherence, Semantic search, Strategic planning, Tool orchestration, Translation or Writing style.
benchmark_meta_category: High-level classification of the benchmark. Options: Agentic task execution, Formalized comprehension & reasoning, Information retrieval, Multilingual capabilities, Multimodal processing, Preference-Alignment, Self-contained foundational capabilities, Unstructured comprehension & reasoning.
n: The number of highlights in this month for this benchmark_category.

prescribed-competencies-coding.csv

Figure 1: Prescribed competencies by model publishers within "Coding" benchmarks.

benchmark_id: Unique identifier for the benchmark (slug).
model_id: Unique identifier for the model (slug).
prescribed_category: A generalized categorization of the competency prescribed by model builders.

prescribed-competencies-lcb.csv

RQ2: Prescribed competencies by model publishes for the LiveCodeBench benchmark.

prescribed_category: A generalized categorization of the competency prescribed by model builders.
n: The number of model releases highlighting this prescribed competency.
p: The percentage of model releases highlighting this prescribed competency.

prescribed-competencies-reasoning-knowledge.csv

Figure 4: Prescribed competencies by model publishers within top 15 "Reasoning and knowledge" benchmarks.

benchmark_id: Unique identifier for the benchmark (slug).
model_id: Unique identifier for the model (slug).
prescribed_category: A generalized categorization of the competency prescribed by model builders.

rankings.csv and rankings-top-15.csv

Table 4: The ranking and geometric mean scores for every benchmark. The rank, score, n_publisher and n_models fields repeat for every domain or access type of models using this benchmark.

benchmark_id: Unique identifier for the benchmark (slug).
rank: A rank describing how influential the benchmark is in the field. A lower number means higher influence. Overall rank of this benchmark based on the score.
n_publishers: The number of publishers using this benchmark in at least one model release.
p_publishers: The percentage of publishers using this benchmark in at least one model release.
n_models: The number of models highlighting a benchmark.
p_models: The percentage of models highlighting a benchmark.
n_publishers_china: The number of publishers in China using this benchmark in at least one model release.
p_publishers_china: The percentage of publishers in China using this benchmark in at least one model release.
n_publishers_west: The number of publishers using this benchmark in at least one model release.
p_publishers_west: The percentage of publishers in China using this benchmark in at least one model release.
n_models_closed: The number of closed models highlighting a benchmark.
p_models_closed: The percentage of closed models highlighting a benchmark.
n_models_open_weight: The number of open weight models highlighting a benchmark.
p_models_open_weight: The percentage of open weight models highlighting a benchmark.
n_models_open_source: The number of open source models highlighting a benchmark.
p_models_open_source: The percentage of open source models highlighting a benchmark.
rank_china: A rank describing how influential the benchmark is in the field for publishers in China. A lower number means higher influence.
rank_west: A rank describing how influential the benchmark is in the field for publishers in China. A lower number means higher influence.
rank_closed: A rank describing how influential the benchmark is in the field for closed models. A lower number means higher influence.
rank_open_weight: A rank describing how influential the benchmark is in the field for open weight models. A lower number means higher influence.
rank_open_source: A rank describing how influential the benchmark is in the field for open weight models. A lower number means higher influence.
score: The score calculated as the geometric mean of the number of publishers and number of models. A higher score represents a higher influence.
score_closed: The score calculated as the geometric mean of the number of publishers and number of closed models. A higher score represents a higher influence.
score_open_weight: The score calculated as the geometric mean of the number of publishers and number of open weight models. A higher score represents a higher influence.
score_open_source: The score calculated as the geometric mean of the number of publishers and number of open source models. A higher score represents a higher influence.
score_west: The score calculated as the geometric mean of the number of publishers in the West and number of models. A higher score represents a higher influence.
score_china: The score calculated as the geometric mean of the number of publishers in China and number of models. A higher score represents a higher influence.

reasoning-knowledge-unique-models.csv

Unique models highlighting at least one of the "Reasoning and knowledge" benchmarks in the Top 15 most popular benchmarks.

n: The number of unique models highlighting one the "Reasoning and knowledge" benchmarks.
p: The percentage of unique models highlighting one the "Reasoning and knowledge" benchmarks.

science-subjects-covered.csv

Table 9: Breakdown of science questions by discipline in top 15 "Reasoning and knowledge" benchmarks: Table 9

science_discipline: The discipline a science question belongs to. Options: Biology, Chemistry, Geography, Math and Physics.
n: The number of questions in this discipline.
p: The percentage of questions in this discipline.

subjects-covered.csv

Table 8: Distribution of subjects covered in top 15 "Reasoning and knowledge" benchmarks.

field: The fields a question is categorized in. Options: Art & Design, Business, Health & Medicine, Humanities & Social Sciences, Law, Science and Tech & Engineering.
n: The number of questions in this field.
p: The percentage of questions in this field.

`figures` graphs

prescribed-competencies-coding.png Figure 1: Count of Prescribed Competencies by Model Builders for The Top 5 "Coding" Benchmarks.
highlights-2025.png Figure 2: Cumulative adoption of benchmarks released in 2025.
highlighted-competencies-by-month.png Figure 3: Selected competencies highlighted by model releases in 2025.
prescribed-competencies-reasoning-knowledge.png Figure 4: Count of Prescribed Competencies by Model Builders for The Top 5 "Reasoning and knowledge" Benchmarks.
highlighted-competencies-by-month-facets.png Figure 5: All competencies highlighted by model releases in 2025.

`code` scripts

The three scripts were used to automatically fetch and extract paper metadata and author information from ArXiv. They are meant to be run sequentially and are designed to fit into our larger data collection/annotation process.

`get-arxiv-paper-metadata.sh`

Given a list of ArXiv paper IDs (inputs/benchmarks/raw/arxiv-ids.txt), fetch the metadata for every id and store the XML API response in a file. The output path is currently hardcoded to inputs/benchmarks/raw/arxiv-metadata.

bash get-arxiv-paper-metadata.sh

`arxiv-metadata.R`

Read the list of XML files that were fetched from the ArXiv API and extract the metadata into a CSV file. The CSV contains the id, the link to the paper on ArXiv, the timestamp when the it was published, the timestamp when it was last updated and the number of versions for this paper.

Rscript code/arxiv-metadata.R -i inputs/benchmarks/raw/arxiv-metadata -o inputs/benchmarks/raw/papers.csv

`arxiv-metadata-by-authors.R`

Extract the names of authors for every paper. This script was designed to run incrementally, so it merges the extracted information with data that was already manually annotated (e.g. institutional affiliation).

Rscript code/arxiv-metadata-by-authors.R -i inputs/benchmarks/raw/arxiv-metadata -o inputs/benchmarks/raw/authors.csv

Downloads last month: 331

Paper for matybohacek/benchmarking-cultures-25

Unsteady Metrics and Benchmarking Cultures of AI Model Builders

Paper • 2605.14164 • Published 8 days ago