LLM evaluation is the practice of systematically measuring how well a language model's outputs meet quality, accuracy, and safety requirements. Unlike traditional software testing with deterministic pass/fail outcomes, LLM evaluation deals with probabilistic outputs that can vary across runs.
Evaluation approaches fall into three categories. Automated metrics use algorithms to score outputs — BLEU and ROUGE for text similarity, exact match for factual accuracy, format validation for structured outputs, and custom rules for domain-specific criteria. These are fast and cheap but often miss nuances that humans catch.
Human evaluation involves subject matter experts reviewing model outputs against rubrics. This captures quality dimensions that automated metrics miss — helpfulness, tone, factual accuracy in context, and overall coherence — but is slow, expensive, and doesn't scale to continuous testing.
Model-based evaluation (also called "LLM-as-judge") uses a capable model to evaluate the outputs of another model. This approach offers a middle ground: more nuanced than rule-based metrics, faster and cheaper than human review. The judging model is given the input, the output, evaluation criteria, and optionally a reference answer, then produces a score and explanation.
Effective evaluation strategies combine all three approaches. Automated metrics provide continuous regression detection in CI/CD pipelines. Model-based evaluation covers qualitative dimensions at scale. Human evaluation calibrates the automated systems and handles high-stakes decisions.
Evaluation should be integrated into the prompt development workflow. Every new prompt version should be evaluated against a test suite before publishing to production, similar to how code changes go through CI before deployment.