The MMLU, Chatbot Arena, HELM and MT-Bench papers and OpenAI's evaluation guide cited below were read on 11 October 2026. Leaderboards and tools change quickly, so we quote no current rankings; check each project's own page for today's numbers.

LLM Evaluation· Chatbot quality and testing

LLM evaluation is the practice of measuring how well a language model, or an application built on one, performs a defined task. It combines a set of test inputs, a way of scoring the outputs, and a decision about what score is good enough.

By Chatbotscape Editorial· Methodology· Published 12 October 2026· Updated 12 October 2026

LLM Evaluation — Measuring Whether a Model Is Good Enough for the Job

Quick answer: A language model is not good or bad in general; it is good or bad at a task. LLM evaluation is how you measure that. Public benchmarks such as MMLU and Chatbot Arena compare models on shared questions and are useful for a shortlist. They cannot say how a model handles your customers' questions, your policies and your tone. For that you run an evaluation on your own questions, and then keep checking after launch.

What an evaluation is made of

Every evaluation, public or private, has the same three parts. OpenAI's evaluation guide describes them as a dataset of inputs, the outputs a model produces for each, and graders that assess those outputs, with human annotations serving as the ground truth for desired behavior. The practical questions are the same whether you are a research lab or a five-person support team:

Which inputs? The questions the model will be asked. The closer they are to real traffic, the more the score means.
How is an output scored? An exact match, a rule, a human reading it, or another model reading it.
What counts as a pass? A threshold you set before you see results, so the result cannot move the goalposts.

Four kinds of evaluation

Kind	What it tells you	Main weakness
Public benchmark	How a model scored on a shared, fixed set of questions	Not your questions; can be gamed or leaked
Task-specific test (your own)	How a model handles your real cases	Costs time to build; only as good as its sample
Human review	Whether an answer is actually right, helpful and on-brand	Slow, costly, and reviewers disagree
LLM-as-a-judge	A fast, cheap approximation of human grading at scale	Known biases; must be checked against humans

What three well-known public evaluations measure

MMLU. Introduced in 2020 by Dan Hendrycks and colleagues, it is a test of 57 tasks across subjects such as elementary mathematics, US history, computer science and law, meant to measure academic and professional knowledge. It says nothing about customer-service questions, and the authors designed it to reveal knowledge gaps in the models of that time.
Chatbot Arena. A crowdsourced platform, described in a 2024 paper by Wei-Lin Chiang and colleagues, where people compare two models' answers head to head and vote. The authors report over 240,000 votes at the time of writing and say the votes agree well with expert raters. It measures general human preference, which is not the same as correctness on your help-center content.
HELM. The Holistic Evaluation of Language Models, led by Percy Liang and Rishi Bommasani, evaluated 30 models on 42 scenarios using seven metrics: accuracy, calibration, robustness, fairness, bias, toxicity and efficiency. Its contribution is the idea that accuracy alone is too narrow a view.

These are research results at the time of publication. Models and leaderboards have moved on since, so read the current versions on the projects' own pages.

Offline and online evaluation

Offline evaluation runs before a change reaches customers: a fixed set of questions run against a candidate model, prompt or knowledge base, scored, and compared with the current version. It is repeatable, which is what makes it useful for comparing choices.

Online evaluation happens on live traffic: sampling real conversations, collecting customer ratings, tracking resolution rate and escalations, and watching for drift. It sees what no test set anticipated, but it sees it after customers did. The two are complements: offline catches what you can foresee, online catches what you cannot. Online evaluation shares tooling with LLM observability.

What a score cannot tell you

Whether it fits your content. A model that tops a knowledge benchmark can still invent a return policy. Measuring that is the job of a test on your own documents; see the RAG evaluation guide for the retrieval case and AI hallucination for the failure itself.
Whether the difference is real. On a small test set, a gap of a few questions is within chance. The companion guide shows the arithmetic.
Whether the benchmark was seen in training. Public questions can leak into training data, which would inflate a score without improving ability. Treat any single public number with caution.
Whether it stays true. Providers update models. A score from last quarter describes last quarter's model; see model drift.
Cost and speed. A better answer that is slower or several times more expensive may not be the better choice; see LLM tokens for how cost is metered.

FAQ

What is LLM evaluation?

Measuring how well a language model or a model-based application performs a defined task, using test inputs, a scoring method and a pass threshold.

What is the difference between a benchmark and an evaluation on my own data?

A benchmark uses shared public questions so many models can be compared. An evaluation on your own data uses your real questions, so it predicts your results. Benchmarks help you shortlist; your own test helps you decide.

What are the main ways to evaluate an LLM?

Public benchmarks, task-specific tests, human review and LLM-as-a-judge. Most teams combine them: benchmarks to shortlist, a test set plus a judge to compare, and human review to check the judge.

What is the difference between offline and online evaluation?

Offline runs a fixed test set before release and is repeatable. Online measures live conversations after release and catches what the test set missed.

Are LLM leaderboards reliable for choosing a chatbot model?

They are a reasonable starting point and a poor final answer. They measure general ability or general preference, not your content, your policies or your cost limits.

How big should a test set be?

Large enough that the difference you care about stands out from chance. A few dozen questions can show a big gap; they cannot separate two models a few points apart.

LLM-as-a-judge — using a model to grade answers, and its known biases.
Golden dataset — the fixed, human-checked set of questions behind a repeatable test.
Large language model — what is being evaluated.
LLM observability — the production side of evaluation.
Model drift — why a past score goes stale.
Model distillation — why you may need to compare a large and a small model.

Sources

Hendrycks et al., Measuring Massive Multitask Language Understanding — arxiv.org/abs/2009.03300, read 11 October 2026. Source of the 57-task scope and the subjects named.
Chiang et al., Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference — arxiv.org/abs/2403.04132, read 11 October 2026. Source of the pairwise-vote design, the 240,000-vote figure and the agreement-with-experts claim (authors' claims).
Liang et al., Holistic Evaluation of Language Models — arxiv.org/abs/2211.09110, read 11 October 2026. Source of the 42 scenarios, seven metrics and 30 models.
Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena — arxiv.org/abs/2306.05685, read 11 October 2026. Source of the judge-agreement claim and the named judge biases.
OpenAI, Getting started with evals — developers.openai.com/api/docs/guides/evaluation-getting-started, read 11 October 2026. Source of the dataset, outputs, annotations and graders description; the page also says OpenAI is deprecating its Evals platform, with a shutdown scheduled for 30 November 2026.
Chatbotscape evaluation methodology. /methodology (continuously updated).