🏢 University of Toronto Organization

The article argues that current AI benchmarks are incomplete because they measure what systems can do, not whether they do what users actually intend. To address this gap, the authors propose a new me