Regression Testing
Running old tests after a change to make sure you didn't break anything that was working before.
Think of it like
Like checking that your new recipe still makes good bread, not just that it doesn't burn.
Example
After updating a model, you re-evaluate it on your previous gold set to make sure accuracy didn't drop.
How it actually works
Regression testing catches silent failures—changes that improve one metric but degrade another. It's essential before shipping. The catch: you need good regression tests (comprehensive, non-brittle) and a clean baseline to compare against. Many teams skip this and later wonder why quality dropped.
For product teams
Regression testing is cheap insurance. Always run it before pushing to production.
For engineers
Maintain a stable test suite; add new tests for each bug; bisect changes if regressions appear.
Related
- Eval Set — Part of this process.
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