Benchmark
A fixed set of tasks, scored the same way, used to compare systems on one measure.
What it means
A benchmark gives different tools the same questions and a shared scoring method, so the results can be compared. For AI research tools, a benchmark usually uses a set of research tasks, a rubric or judge, and an average score. DRACO is one example. It was built by Perplexity and Harvard and spans 10 domains, and the October 2026 run on Hyperresearch's method page used 10 tasks.
Read benchmark results with care. A score tells you how a system did on those tasks with that judge, not how it will do on your question. Check the method, the number of tasks, whether some systems used published reports rather than fresh runs, and who ran the test. Then run a few of your own questions before you choose.
Continue this reading path
Cite this entry
https://bestaiforresearch.com/glossary/benchmark/