Researchers propose a new framework for designing benchmarks for AI systems that perform knowledge work. They introduce four explicit fields to describe what part of the work is represented, under what conditions it's tested, what product the system produces, and what aspect of that product is evaluated. The team applies this representation to three existing benchmark datasets, showing how different aspects of work can be captured within a single framework. --- Why it matters: This matters because AI systems are increasingly being used for complex tasks like completing workflows and producing work products. A standardized way to design benchmarks will help developers compare the performance of these systems across different domains and applications. Source: https://arxiv.org/abs/2605.23262