Why MMLU and GPQA scores don't tell you if your app works
A benchmark is a fixed public list of questions with known answers, which labs use to compare one model against another. None of those questions ask about your refund policy, your codebase or your customers. Use benchmarks to shortlist two or three models, then let product evals built from your own production logs decide which one ships.
What a benchmark is
A benchmark is an exam for models. MMLU has multiple-choice questions across 57 subjects, from law to physics. GPQA has 448 hard science questions written by PhD-level experts. HumanEval and SWE-bench test code, and Humanity's Last Exam collects questions designed to be hard for current models.
Labs report these scores so you can compare models on the same questions. That is useful, and it answers a different question from the one you have.
What a product eval is
A product eval is built from your own production logs, around the mistakes your system actually makes. It asks things like: did the support bot quote the right refund window for this customer's region, did the agent check the order status before promising a delivery date, did the summary keep the penalty clause.
It is the only test that tells you whether your app got better or worse after a change, because it's the only one built from your app.
| Benchmark | Product eval | |
|---|---|---|
| Questions | Public, fixed | From your own traffic |
| Measures | General capability | Your failure modes |
| Answers | Which model is generally stronger? | Did this change help my users? |
| Good for | Shortlisting models | Shipping decisions |
Why benchmark scores mislead
Benchmarks run out of room
When every top model scores over 90%, a benchmark can't separate them anymore. The Humanity's Last Exam paper (2025) notes that LLMs now score over 90% on MMLU, and replaces it with 2,500 harder questions.
Benchmarks have mistakes in the answer key
The MMLU-Redux team re-checked 5,700 MMLU questions and estimated that 6.49% of the benchmark contains errors. In the virology questions they checked, 57% were wrong. A model can lose points for giving the right answer.
Benchmarks measure what they measure
GPQA is good at its job: experts scored 65%, and skilled non-experts with 30 minutes or more of web search scored 34%. That measures deep scientific reasoning, which tells you little about whether your support bot follows your policy.
Leaderboards can be gamed
In April 2025, an experimental version of Llama 4 Maverick built for chat ranked near the top of LMArena, while the released model landed around 32nd. LMArena apologised and changed its rules.
Public answers leak into training data
In February 2026, OpenAI stopped reporting SWE-bench Verified, because every frontier model it tested could reproduce some of the original human-written fixes. When the answers are on the internet, a high score can mean the model has seen them.
How to use benchmarks
I use benchmarks for one job: picking two or three models to try. Then my own product evals decide which one ships.
- Use benchmarks and leaderboards related to your task (coding, maths, long context, your language) to shortlist candidates.
- Run the shortlisted models on your own eval set.
- Compare pass rate, cost and latency on your data.
- Choose, and keep the eval set to test the next model that comes out.
Every model decision in your team should cite your own eval results, with the benchmark only as the reason a model made the shortlist.
Where the real work is
Hamel Husain and Shreya Shankar say they spend 60 to 80% of development time on error analysis and evals. That sounds like a lot until you see what the alternative looks like: shipping changes without knowing what they did.
From the floor
In the session on how LLMs are trained, I showed Humanity's Last Exam as one of the hardest exams for LLMs, with low accuracy for every model, and benchmarks like HumanEval for code, as the way labs compare models after training. Those scores tell you which model is stronger in general, and they're where your own testing starts, not where it ends.
Related
Sources
- Measuring Massive Multitask Language Understanding (MMLU), Hendrycks and others (2020)
- GPQA: A Graduate-Level Google-Proof Q&A Benchmark, Rein and others (2023)
- Are We Done with MMLU? (MMLU-Redux), Gema and others (2024)
- Humanity's Last Exam, Phan and others (2025)
- Meta's vanilla Maverick AI model ranks below rivals on a popular chat benchmark, TechCrunch (11 April 2025)
- Why we no longer evaluate SWE-bench Verified, OpenAI (23 February 2026)
- Husain and Shankar, LLM evals FAQ