RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation
Apple researchers have introduced RISED, a framework designed to improve how AI agents are trained across multiple, varied environments. The method uses self-distillation to refine performance, addressing the limitations of existing strategies that often rely solely on local rewards or rigid data selection. By enabling more efficient learning across diverse tasks, this approach aims to advance the development of generalist agents capable of functioning effectively in different contexts.