A benchmark is a standardised test suite that allows models to be compared with each other on the same task and metric. It arose in response to the inadequacy of perplexity alone, which measures the ability to predict text but not usefulness. Typologically, benchmarks split into knowledge and reasoning sets with closed answers that are scored automatically, generative tasks requiring a judge – either human or a stronger model acting as evaluator – and comparisons based on anonymous human preferences in so-called arenas. The fundamental methodological weakness is contamination: benchmarks are public, so they very likely end up in the training corpus, turning a test of knowledge into a test of memory. There is also Goodhart’s law – once a metric becomes the target of optimisation, it stops being a good measure – plus sensitivity of results to prompt wording, and failure to cover exactly those capabilities that matter in production. The practical recommendation is therefore to build your own private evaluation set derived from your organisation’s real data and cases, and to treat public leaderboards only as a rough guide.
A benchmark is like a school entrance exam. Useful, because otherwise you would be comparing applicants on gut feeling. Treacherous, because once the questions leak online, everybody starts studying for that exact test – and you end up measuring who had the better prep course, not who is more talented. And once parents learn that the number of books read is what counts, children start merely flipping through books. That is why good employers, alongside generic tests, always invent their own task that nobody has seen in advance – one that closely resembles the actual work.