Published on July 5, 2026, the item describes a method that trains agents to create rubrics and use them for planning, searching, reviewing, and answering.
Published on July 5, 2026, the item presents RubricEM, which trains research agents to create rubrics before starting work and use them to plan, search, review, and answer. A judge scores stages of the trajectory, and evaluated attempts are turned into lessons for future tasks.
The approach combines stage-level rewards and reusable lessons rather than relying only on a single final score. To check the scope and findings, consult the original item record and compare its description with the paper or technical material it cites. If using AI to study or apply the method, do not include sensitive organizational data without authorization, and follow internal data-protection policies.