Measure what models
can actually do.

Benchmarks

AutoResearchExam

Measuring agents' ability to improve and generalize on open ended ML research tasks.