Polymind for researchers: how to cite our data

If you are using Polymind in research, cite the method and the data snapshot you actually used. The leaderboard changes as new public runs arrive, so a citation should make the retrieval date and surface clear.

The short version:

1. Cite the methodology page for how ranks are computed.
2. Cite the specific leaderboard page you used, such as all models, code, or research.
3. Use the machine-readable JSON or CSV endpoint when you need exact numbers.
4. Include the retrieval date.

What Polymind data represents

Polymind is a multi-model debate system. A user prompt goes to several panelist models. The panelists answer, optionally critique one another, and a judge model writes the final synthesis. The judge also names the panelists it leaned on. Those names become picks.

The public leaderboard aggregates picks over appearances. It answers a narrow question:

Which panelist models do Polymind judges lean on most often across public runs?

It does not answer "which model is objectively best" in every setting. It measures judge preference inside this product shape. That is still useful data, but the scope matters.

Preferred citation targets

For methodology, cite:

• Polymind leaderboard methodology

For the current all-domain leaderboard, cite:

• Polymind all-domain leaderboard
• JSON twin
• CSV snapshot

For domain-specific slices, cite the matching page and JSON endpoint:

The page-level "Cite this leaderboard" dialog also provides BibTeX, APA, HTML, and Markdown formats.

What to report

When quoting Polymind numbers, report at least:

• the leaderboard URL;
• retrieval date;
• domain;
• appearances;
• picks;
• raw win rate;
• Wilson lower bound;
• methodology version if shown on the page.

Do not cite only a rank. Rank without sample size is easy to misunderstand. A model ranked first on a small slice and a model ranked first after many appearances are different claims.

License and attribution

Polymind leaderboard data is intended to be citeable and reusable with attribution. The dataset schema points to a CC-BY-4.0 license, and the machine-readable endpoints are public so researchers can inspect the same numbers the HTML renders.

Attribution should name Polymind, link to the relevant leaderboard or methodology page, and include the retrieval date. If you transform the data, say what you changed.

Limits to state in papers

If you use Polymind data in a paper, state these limitations.

The judge is an LLM, so the picks reflect judge preference, not ground truth. The prompt distribution is user-driven, not a controlled benchmark suite. Domain labels are produced by a classifier and can be noisy near boundaries. Public run volume affects confidence, especially for narrow slices.

Those limits do not make the data unusable. They make it specific. The right citation is not "Polymind proves model X is best." It is closer to: "In Polymind's public judge-pick leaderboard, model X had the highest Wilson lower bound on the retrieved date."

That sentence is less flashy. It is also the kind a reader can verify.