LASER looks for more useful safety-evaluation examples

What changed

LASER combines lightweight classifiers with reasoning-model graders to find diverse, ambiguous examples for safety evaluations.

Source: OpenAI Alignment

Our analysis

Builder’s view: improve a test set by looking for difficult edge cases, rather than repeatedly testing familiar examples.

Keep in mind

This is evaluation-data curation, not a general guarantee that a deployed model is safe.

Source: OpenAI Alignment