Evaluating a large language model is harder than evaluating almost any software that came before it, because the thing you want to measure — general competence and good behavior across open-ended tasks — has no single correct answer to check against. A calculator either returns 4 or it does not; an LLM asked to summarize a document, write code, or refuse a harmful request can succeed or fail along a dozen axes at once. The whole discipline of LLM evaluation is a set of imperfect proxies for that unmeasurable ideal, and the most important skill is knowing what each proxy really measures and where it quietly lies.\n\nCapability benchmarks score knowledge and reasoning against fixed answer keys. The familiar leaderboard numbers come from standardized test sets: MMLU for broad multiple-choice knowledge across dozens of subjects, GSM8K and MATH for grade-school and competition mathematics, HumanEval for writing correct code, HellaSwag and ARC for commonsense, and aggregate suites like BIG-bench that bundle hundreds of tasks. Each reduces a messy skill to a gradeable score, which is exactly their appeal and their weakness — they are convenient and comparable, but a single accuracy percentage flattens away how and why a model fails.\n\nThe benchmark numbers are systematically undermined by contamination and saturation. The deepest problem is data contamination: because models train on scrapes of the whole internet, the test questions themselves often leak into the training data, so a high score may reflect memorization rather than skill. Benchmarks also saturate — once frontier models cluster near the ceiling, the test stops discriminating between them and stops being informative. And strong benchmark performance routinely fails to predict real-world usefulness, because neatly formatted multiple-choice questions look nothing like the sprawling, ambiguous requests real users send. This is why the field keeps having to build harder benchmarks and why no serious evaluation rests on one number.\n\nBehavioral evaluation measures how a model acts, and increasingly uses judges and humans rather than answer keys. Beyond raw capability sit the qualities that decide whether a model is actually good to use: does it follow instructions, stay honest instead of hallucinating confident falsehoods, refuse genuinely harmful requests without over-refusing benign ones, and resist adversarial jailbreaks. Because these have no answer key, evaluation turns to two moves — LLM-as-a-judge, where a strong model grades another's outputs at scale (fast and cheap, but biased and gameable), and human preference, most visibly the Chatbot Arena, where people vote on anonymized head-to-head responses and an Elo rating emerges. Human preference is the closest thing to ground truth for open-ended quality, which is why it anchors the field despite being slow and expensive. Hovering over all of this is the debate over emergent abilities — skills that appear abruptly at scale — and whether they are real phase changes or artifacts of how we chose to measure.\n\n| Evaluation type | Examples | Measures | Main pitfall |\n|---|---|---|---|\n| Capability benchmark | MMLU, GSM8K, HumanEval | Knowledge, reasoning, coding | Contamination, saturation |\n| Behavioral / safety | Instruction following, jailbreak, refusal | How the model acts | No answer key, subjective |\n| LLM-as-a-judge | Model grades model outputs | Scalable quality scores | Judge bias, gameable |\n| Human preference | Chatbot Arena (Elo) | Real open-ended quality | Slow, costly, popularity bias |\n\n``svg\n\n``\n\nThe unhelpful way to think about LLM evaluation is to treat the leaderboard as a scoreboard and the top number as the winner. The useful way is to see every metric as a proxy standing in for something you cannot measure directly — genuine competence and trustworthy behavior — and to ask of each one what it captures and what it hides. Capability benchmarks are convenient but contaminated and saturating; behavioral evals matter most but resist automation; LLM judges scale but carry bias; human preference is the nearest thing to truth but is slow and rewards charm. Read LLM evaluation through a what-behavior-do-I-actually-care-about lens rather than a which-model-tops-the-leaderboard lens, and you stop chasing a single score and start doing what real evaluation demands: triangulating many imperfect signals toward the capability and conduct you were trying to measure all along.