AI Agent Papers2026
Tests decision models as actors, trajectory judges, and reinforcement-learning teachers for long-horizon agents. It finds that a model can be a weak actor yet still improve policy learning by evaluating actions after seeing their consequences.