Citing and Describing TGraphX Accurately in Papers and Reports

If you use a research tool in published work, citing it correctly is both courtesy and reproducibility. For TGraphX this is easy, because the project ships its own citation metadata — but describing it accurately, without drifting into claims the software does not make, takes a little care. This short reference gives you the BibTeX, the version and authorship facts, and a set of conservative descriptions you can paste into a paper or report with confidence.

It pairs with the integrity-minded articles Benchmark Disclaimers and Research Integrity in Benchmarks.

The canonical citation

The README and CITATION.cff agree on the preferred citation — the arXiv preprint:

@misc{sajjadi2025tgraphxtensorawaregraphneural,
            title={TGraphX: Tensor-Aware Graph Neural Network for Multi-Dimensional Feature Learning},
            author={Arash Sajjadi and Mark Eramian},
            year={2025},
            eprint={2504.03953},
            archivePrefix={arXiv},
            primaryClass={cs.CV},
            url={https://arxiv.org/abs/2504.03953},
        }
        

If your venue prefers software citation, CITATION.cff (machine-readable, used by GitHub's "Cite this repository" button) records the same authorship with an ORCID for the lead author and an MIT licence. Cite the preprint for the method; cite the software entry and the version when reproducibility of a specific run matters.

The facts you should state

Always pin the version: tgx.__version__ returns it at runtime, and including it makes a result reproducible against the right code. Note too that the README states TGraphX is an independent research software project and does not represent an official product or endorsement of the University of Saskatchewan — useful to reflect if you describe its provenance.

A conservative description to reuse

When you need a sentence or two of prose, use wording the software supports:

TGraphX (Sajjadi & Eramian, 2025; arXiv:2504.03953) is a tensor-aware graph learning framework for PyTorch in which nodes, edges, or graph-level attributes may carry structured tensor representations such as [C, H, W] feature maps. We used version X.Y.Z (Beta) under the MIT licence.

Claims to make — and to avoid

The project's own CHANGELOG repeatedly states that it makes no superiority claims, and its benchmark guidance is explicit that performance numbers require your own profiling. Mirroring that restraint in your write-up is not just polite — it is what keeps your paper's claims defensible. See Benchmark Disclaimers for the reasoning.

Honest framing

Two caveats. First, the cff file records a software version field that may lag the package version — cite the preprint for the method and the actual installed version (from tgx.__version__) for reproducibility, rather than relying on a single number. Second, "Beta" is a real status; describing the framework as production-hardened would overstate it. Accurate citation is mostly a matter of using the project's own conservative language.

A ready-to-adapt methods paragraph

For convenience, here is a methods-section paragraph you can adapt directly, filling in your specifics:

We used TGraphX (Sajjadi & Eramian, 2025; arXiv:2504.03953), a tensor-aware graph learning framework for PyTorch, version X.Y.Z, under the MIT licence. Graphs were represented with tensor-valued node features of shape [C, H, W]; we trained <model> for <N> epochs with seed <s> under deterministic mode, and report mean and standard deviation over <k> seeds.

Three things make this defensible: it cites the method by its preprint, pins the exact version (so the code is identifiable), and reports a spread over seeds rather than a single number. Adapt the shapes, model, and counts to your run, and drop the deterministic-mode clause if you trained on GPU without it. Avoid inserting performance superlatives; if you compare against another framework, describe complementarity and state what you measured rather than asserting an ordering. A methods paragraph that names the version and the seeds is one a reviewer can trust and a future reader can reproduce — which is the entire point of citing software carefully.

And if you build on a specific TGraphX capability — say the knowledge-graph models or the evolutionary optimiser — cite the relevant docs/ page or module alongside the preprint, so a reader can find the exact feature you used rather than inferring it from the framework name alone. Specificity in a citation is a courtesy to whoever tries to reproduce you.

Related guides

• TGraphX Benchmark Disclaimers
• Research Integrity in Graph-Learning Benchmarks
• Where Tensor-Native Graph Learning Fits

Conclusion

Citing TGraphX is straightforward: use the arXiv BibTeX for the method, the CITATION.cff/version for the software, and pin tgx.__version__ for reproducibility. Describing it well is about restraint — call it a tensor-aware, Beta research framework, state what it is designed for, and avoid performance or superiority claims you have not measured. Borrow the project's own conservative wording and your citations will be both accurate and defensible.