Overfitting to Benchmarks
When a model is trained on or sees benchmark questions during pretraining, inflating its scores.
Think of it like
Like studying by memorizing the exact exam questions, then only the official exam.
Example
A model trained on Common Crawl data saw the MMLU questions during pretraining. Its MMLU score is high, but real-world knowledge is the same as a model without that exposure.
How it actually works
This is rampant because benchmark datasets are public and often mixed into training corpora. Hard to fix: you can remove known benchmarks from training, but new benchmarks get contaminated quickly. The solution is retiring benchmarks (replacing them with new ones) and being honest about contamination. Some labs now train on data cut off before benchmark release dates.
For product teams
Ask vendors whether their model was trained on benchmark data. High scores on old benchmarks matter less.
For engineers
Version your training data with cutoff dates; document known contamination; retire old benchmarks regularly.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome