Method lets humans oversee AI agents that exceed their own abilities

Original: Calibrating Conservatism for Scalable Oversight

Source: arxiv.org ↗