Crux

New open-world AI evaluations use real-world tasks for capability assessment

Researchers have introduced a new evaluation method called open-world evaluations, which complements traditional benchmark-based assessments for frontier AI capabilities. These evaluations focus on long-horizon, complex…