Cite AOBench¶
If AOBench contributed to work you are publishing, please cite the software. Citing research software is how the people who maintain it get credit that counts, and it lets readers find the exact version your numbers came from.
Cite the software¶
GitHub renders a "Cite this repository" button from CITATION.cff, which is the authoritative metadata. In BibTeX:
@software{aobench,
title = {{AOBench}: A Trace-Driven, Role-Aware Benchmark for Agent
Operations in Realistic Environments},
author = {Seyedkazemi Ardebili, Mohsen and Bartolini, Andrea},
year = {2026},
version = {0.4.1},
license = {Apache-2.0},
doi = {10.5281/zenodo.21854863},
url = {https://doi.org/10.5281/zenodo.21854863},
publisher = {Zenodo}
}
Cite the version you actually ran, not main. Every release is tagged, and the version field above should match the tag you used. Results from different versions are not necessarily comparable — scoring profiles and the task corpus evolve between minor versions.
Machine-readable metadata¶
| File | Standard | Consumed by |
|---|---|---|
CITATION.cff | Citation File Format 1.2.0 | GitHub, Zenodo, citation managers |
codemeta.json | CodeMeta 3.0 | Research-software registries, OpenAIRE |
.zenodo.json | Zenodo deposition metadata | Zenodo, when a release is archived |
All three carry the same authors, ORCIDs, affiliation, licence and keywords. If you change one, change all three — divergent metadata is worse than none, because downstream registries will disagree about who wrote the software.
Archival DOI¶
Every release is archived on Zenodo and carries a DOI.
| DOI | Use it when |
|---|---|
10.5281/zenodo.21854862 | You mean AOBench in general — the concept DOI always resolves to the newest release |
10.5281/zenodo.21854863 | You mean v0.4.1 exactly — use this for reproducibility, so a reader gets the code you ran |
For a paper reporting benchmark numbers, cite the version DOI. A concept DOI resolves to whatever is newest at the time of reading, which is not what your results were produced with.
Also cite the data, where it applies¶
Six of the 29 environment snapshots and eight of the 88 tasks are grounded in the public Marconi100 ExaData dataset. If your work depends on those, cite the dataset paper too — the environments are derived from it, not independent of it:
@article{m100exadata,
title = {{M100 ExaData}: a data collection campaign on the {CINECA}'s
{Marconi100} {Tier-0} supercomputer},
author = {Borghesi, Andrea and Di Santi, Carmine and Molan, Martin and
Seyedkazemi Ardebili, Mohsen and Mauri, Alessio and
Guarrasi, Massimiliano and Galetti, Daniela and Cestari, Mirko and
Barchi, Francesco and Benini, Luca and Beneventi, Francesco and
Bartolini, Andrea},
journal = {Scientific Data},
volume = {10},
number = {1},
pages = {288},
year = {2023},
doi = {10.1038/s41597-023-02174-3}
}
What to report alongside a number¶
If you publish an AOBench score, these four fields let a reader reproduce it. Without them a score is not checkable, and an uncheckable benchmark number is not evidence:
| Report | Why |
|---|---|
Version tag (e.g. v0.4.0) | Task corpus and scoring profiles change between versions |
Scoring profile (e.g. default_hpc_v01) | Dimension weights differ between profiles |
Split (dev or test) | The test split is held out; dev and test numbers are not interchangeable |
| Adapter and model identifier | openai:gpt-4o is not the same system as direct_qa |
See Reproducing results for the full contract.