reward-disentanglement

This paper introduces GRID, a social learning method that extracts universal behaviors from heterogeneous agents by decomposing per-agent reward functions into general and specific rewards, enabling a generalist agent that avoids mode-averaging bias.