LLM Evaluation

LLM evaluation is the practice of measuring how well a large language model, or a product built on one, performs on defined tasks such as answering correctly, following instructions, writing working code, staying safe and citing real sources. It ranges from public benchmarks that compare frontier models to small private test sets that check whether one company's chatbot gives the right refund policy.

Most evaluation combines four methods:

• Static benchmarks with fixed questions and scored answers, such as GPQA Diamond for graduate-level science, Humanity's Last Exam for expert knowledge, and SWE-bench variants for real software fixes
• Human preference ratings, where people compare two anonymous answers and vote, as on Arena (renamed from LMArena in January 2026), which reported more than 5 million monthly users
• Model-graded evaluation, where a strong LLM scores outputs against a rubric, which scales well but inherits the grader's biases
• Task-specific and adversarial tests built from real usage, including red-teaming for safety and prompt injection

Public leaderboards move quickly. In October 2026, Scale's Humanity's Last Exam (Diamond) board had GPT-6 Astra on top at 60.6%, followed by Claude Opus 5.5, while Arena's text leaderboard had Gemini 4 Argon first. On October 15, 2026, Scale and Reflection released SWE-Bench Pro V2, which removed 89 invalid tasks and locked down network access during runs because earlier versions could be gamed. That pattern repeats: once a benchmark is widely reported, models are tuned toward it, test questions leak into training data, and scores saturate.

The numbers are also more fragile than they look. A September 2026 study, 'There Is No Neutral Harness', ran 12 open-weight models on the same 3,679 benchmark items under 26 reasonable test configurations. One model's score ranged from 31% to 89% depending only on how answers were prompted and scored, and four different models reached rank one under some configuration. OpenAI's 2025 research on hallucinations made a related point: accuracy-only scoring rewards models for guessing instead of saying they do not know, which pushes up hallucination rates. Good evaluation reports errors and abstentions separately from correct answers.

For teams building products, the practical rule is to evaluate on your own tasks. Collect a few hundred real inputs, define what a good output looks like, score candidates the same way each time, and rerun the set whenever the model, prompt or retrieval setup changes. Systems that use retrieval-augmented generation need two layers: retrieval evaluation (did the system fetch the right documents?) and generation evaluation (did the answer use them faithfully and cite them correctly?). For a deeper look at how benchmark suites are built, see AI benchmarks.

For marketing and GEO teams, evaluation has a second meaning: checking what models say about you. The same discipline applies. Fix a prompt set that reflects real buyer questions, run it on the assistants your customers use, and score each answer for whether your brand appears, whether the facts are right and whether the cited sources are accurate. Repeat it on a schedule, because answers change with every model update. Our framework to fact-check what AI says about your brand walks through the scoring step by step.