What is LLM evaluation? Methods, metrics, and graders
Also called LLM evals, evals, model evaluation
Definition
LLM evaluation is the practice of measuring how well a language model or LLM-based application performs on defined tasks, using test datasets, automated graders, model-based judges, and human review.
LLM evaluation, usually shortened to evals, is how teams decide whether a prompt change, model upgrade, or new agent design made an application better or worse. LLM outputs vary from run to run and rarely have a single correct string, so evaluation relies on test sets of representative inputs, explicit success criteria, and graders that score each output. OpenAI's evaluation best practices recommend evaluating early and often and growing the test set from production logs.
How LLM evaluation works#
An eval runs a set of tasks through the system, scores each result, and aggregates the scores. Anthropic's engineering guide to evals for AI agents groups graders into three types:
| Grader | How it scores | Strengths | Weaknesses |
|---|---|---|---|
| Code-based | Exact match, regex, unit tests, schema checks | Fast, cheap, reproducible | Limited on subjective quality |
| Model-based | An LLM judge applies a rubric or compares outputs | Flexible; scales to open-ended output | Non-deterministic; needs calibration |
| Human | Expert review, spot checks, A/B tests | Highest quality | Slow and expensive |
Most suites combine all three: code checks for hard requirements, model judges for qualities such as faithfulness or tone, and periodic human review to calibrate the judges. For AI agents, the same guide separates the transcript, the full record of steps and tool calls, from the outcome, the final state of the environment, such as whether a reservation actually exists in the database.
Anthropic also distinguishes capability evals from regression evals. Capability evals start with a low pass rate and measure progress. Regression evals should stay near a 100% pass rate and catch backsliding when prompts or models change.
LLM evaluation vs benchmarks#
Public LLM benchmarks compare models on fixed, shared tasks. Application evals test your system, including its prompts, retrieval, and tools, on your own data and success criteria. A model that leads a leaderboard can still fail your eval, so benchmarks help shortlist candidates and application evals decide what ships.
Why LLM evaluation matters#
Without evals, every change is judged by spot-checking a few outputs, which misses regressions in cases nobody looked at. Offline evals run before deployment; online evaluation scores sampled production traffic, which is where evaluation and LLM observability meet. Evals are also how teams measure hallucination rates and guardrail false positives instead of guessing at them.
The LLM evaluation guide goes deeper on designing eval suites.
Related terms
Guides that cover llm evaluation
- LLM evaluation: a practical guide to evals, metrics, and judges
- AI agent evaluation: metrics, graders, and how to test agents
- LLM benchmarks explained: what they measure and how to read them
- LLM observability: what to capture, which metrics, and tools
- How to build an AI agent: a step-by-step guide in Python
Cite this definition
The AI Engineering Editors. "What is LLM evaluation? Methods, metrics, and graders." The AI Engineering, updated Nov 06, 2026. https://theaiengineering.com/glossary/llm-evaluation
- Title
- What is LLM evaluation? Methods, metrics, and graders
- Author
- The AI Engineering Editors
- Publisher
- The AI Engineering
- Updated