LLM Evaluation and Benchmarks

For a classifier, accuracy on a held-out test set settles most questions. For a model that can write anything, evaluation is much harder.

The main approaches:

• Static benchmarks. Fixed question sets with checkable answers: multiple-choice knowledge tests (such as MMLU), math problems with exact answers, coding tasks verified by unit tests. Cheap and reproducible.
• Human preference. People compare two anonymous answers and pick the better one; many comparisons produce a leaderboard ranking. Captures overall helpfulness, but also rewards style and length.
• LLM-as-a-judge. A strong model grades answers against a rubric. Scales well, but judges have biases, for example favoring longer answers or their own style.
• Your own evals. A set of real tasks from your application with clear grading. Usually the most informative.

Watch out for:

• Contamination: benchmark questions leak onto the web and into pretraining data.
• Saturation: top models score near 100%, so the benchmark no longer separates them.
• Goodhart’s law: once a number becomes a target, it stops measuring what you care about (compare reward hacking).

Evaluation also covers safety: red teaming probes for harmful behavior that standard benchmarks miss, and hallucination rates need their own tests.