Explain the difference between pointwise and pairwise evaluation in one breath.

There are two ways to ask a judge (human or LLM) to evaluate model outputs. Understanding the difference matters because the choice affects both the quality and cost of your evaluation. **Pointwise evaluation** shows the judge one output at a time and asks for a score on an absolute scale, like 1 to 5. The judge reads the output, considers the rubric, and assigns a number. This is fast: evaluating n outputs takes n judgments. Pairwise evaluation shows the judge two outputs side by side and asks which is better (or whether they are tied). The judge makes a relative comparison rather than an absolute one. This is more reliable because humans and LLMs both find it easier to compare two things directly than to assign consistent absolute numbers. The tradeoff is cost. To produce a full ranking of n models, pairwise evaluation needs to compare every pair, which is n times (n minus 1) divided by 2 comparisons. For 10 models, that is 45 comparisons versus 10 pointwise scores. For 100 models, it is 4,950 comparisons. The quadratic cost is why systems like Chatbot Arena use statistical techniques (Bradley-Terry models, Elo ratings) to approximate the full ranking from a subset of comparisons.

**Pointwise evaluation** scores each output independently on an absolute scale (1 to 5, or pass/fail). The judge sees one output, applies a rubric, and assigns a number. It is fast, with n judgments for n outputs, but prone to calibration drift: the judge's internal standard can shift across items, making scores inconsistent. Pairwise evaluation shows two outputs side by side and asks which is better. Relative comparison is cognitively easier for both humans and LLM judges, which makes pairwise scores more consistent and reliable. The downside is cost: ranking n models requires comparing every pair, which scales as O(n squared). In practice, pairwise evaluation is the standard for comparing model quality. Chatbot Arena uses pairwise human votes to compute Elo ratings. MT-Bench uses an LLM judge in pointwise mode for efficiency. AlpacaEval 2 uses pairwise comparison against a fixed baseline to avoid the full quadratic cost. The design choice depends on what you are optimizing. Use pointwise when you need absolute scores for many items quickly (regression testing, CI gates). Use pairwise when ranking accuracy matters more than speed (model selection, leaderboard construction).