Glossary

Plain-language definitions of terms that show up across these briefs. It grows as the bundle does.

• Preprint — a paper its authors have posted publicly before independent experts have reviewed it. Most arXiv papers are preprints. Useful and early — not yet vetted.
• Peer review — the check where independent experts in the field read a paper and judge whether the method and claims hold up before a journal or conference accepts it.
• Benchmark — a fixed test used to measure and compare AI systems on a task, so two models can be scored on the same yardstick.
• Baseline — the existing method a new result is measured against, to show whether the new approach is actually better.
• Ablation — an experiment that removes one piece of a system to see how much that piece was contributing. It's how authors show which part did the work.
• Fine-tuning — further training of an existing model on a narrower set of data to make it better at a specific task.
• Parameters — the adjustable values a model learns during training. Their count — often in the billions — is a rough proxy for a model's size.
• State of the art (SOTA) — the best published result on a given task so far. A paper claiming SOTA is claiming it beat every prior result on that benchmark.