LLM evaluation can measure factual accuracy, instruction following, reasoning on a defined task, formatting, refusal behavior and consistency. Methods range from exact-match checks and unit tests to expert review and model-based grading. Each method has blind spots, so a single score rarely captures overall quality.
For a support summarizer, an evaluation set might hold 100 past tickets with a human-written reference summary and a checklist: does it name the problem, the action taken and the next step? Run the same set on each model or prompt version and compare failures side by side. When one model grades another model's output, spot-check the grades with a person, because automated graders can favor longer or more confident answers. Results guide prompt engineering changes.
Build a small, representative evaluation set before changing models or prompts. Separate the examples used to tune a system from those used to judge it, and inspect failures rather than relying on averages alone. Related: AI Agent Evaluation, AI Benchmark.