decision-oriented-training

This paper introduces Agent-Authored World Modeling (AAWM), a training procedure that constructs world-model supervision based on the policy's own decision needs rather than next-observation prediction, aligning the learning objective with the dynamics required for effective decision-making.