Search

Benchmark

MMLU-PRO

Reported in 25 papers on Zotpaper. The latest, each as the paper states it:

• Training AI from partially misaligned states makes behavior fixes last longer58.7% clean accuracy

All benchmarks →