What an evaluation score tells you

An evaluation score answers a question about a model on a particular set of inputs under a particular procedure. Before interpreting the number, identify those three things.

Prediction loss and perplexity

On held-out text, next-token cross-entropy averages $-\ln p$ for the observed next tokens. Perplexity is the exponential of that average. A mean loss of $\ln4$ gives perplexity 4.

Lower perplexity means better prediction of that evaluation text. It does not by itself establish factual accuracy or usefulness. Compare scores using the same tokenization, data, and evaluation procedure; otherwise the numbers may measure different units or tasks.

Task accuracy

MMLU contains multiple-choice questions across 57 subjects. One evaluation method compares the model’s scores for the answer options. Other setups generate an answer and parse it. The method, prompt, and allowed tools can affect the result.

These illustrative scores show that a high-probability choice can be wrong. The contamination switch demonstrates a possible effect of prior exposure; it does not measure or diagnose contamination in a real model.

If a model answers 8 of 10 questions correctly, its observed accuracy is 80%. That small test does not establish 80% accuracy on every population of future questions. Sampling variation and how the questions were selected matter.

Prior exposure changes the interpretation

If evaluation examples appeared in training, success may partly reflect memorization. Check for overlap and use suitable held-out tasks. Making questions harder is not, by itself, proof that they are uncontaminated.

Also distinguish validation, used to choose prompts or models, from a final test, used after those choices are frozen. Repeatedly choosing the best system on the same public benchmark can make its score less representative of new tasks.

Pick measurements that match the intended use

For a coding task, test whether the program behaves correctly, including on edge cases. For source-based answers, check whether the cited source supports each claim. For a service, consider latency, cost, and failures across relevant groups as well as average accuracy.

These measurements answer different questions. An aggregate score should not conceal a serious failure on the cases that matter most. The HELM framework is one example of evaluation across several dimensions.

Next: common failure patterns.

Sources · 4

1. Jurafsky, D., & Martin, J. H. (2026). Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition with Language Models (3rd ed., draft of January 6, 2026). Stanford University. Ch. 7 (Large Language Models); Ch. 10 (Masked Language Models).
2. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2021). Measuring Massive Multitask Language Understanding. International Conference on Learning Representations (ICLR 2021). arXiv:2009.03300.
3. Zhang, A. K., Klyman, K., Mai, Y., Levine, Y., Zhang, Y., Bommasani, R., & Liang, P. (2024). Language model developers should report train-test overlap. arXiv:2410.08385. (Position paper, ICML 2025.)
4. Liang, P., Bommasani, R., Lee, T., Tsipras, D., Soylu, D., Yasunaga, M., et al. (2023). Holistic Evaluation of Language Models (HELM). Transactions on Machine Learning Research (TMLR). arXiv:2211.09110.

Full bibliography →