Deep research agents are designed to plan, search, evaluate evidence, and synthesize long-form reports. Training these agents with reinforcement learning is difficult because they operate over long horizons and their outputs lack exact, verifiable rewards—making it hard to provide clear feedback at each step.

The paper introduces RubricEM, a method that tackles this problem through meta-RL with rubric-guided policy decomposition. The approach aims to move beyond verifiable rewards, using rubrics to structure the learning process and decompose policies into more manageable components.

By guiding policy decomposition with rubrics, RubricEM offers a potential path for training agents on complex, long-horizon tasks where traditional reward signals are insufficient. The abstract does not provide further implementation details, but the proposed framework directly addresses a core bottleneck in deep research agent training.