When external rewards or task supervision are absent, human behavior relies on intrinsic motivation—neural signals that drive curiosity, creativity, play and other essential aspects of intelligence. Intrinsic motivation plays a crucial developmental role in building robust world models that enable generalization and adaptation. Existing work leaves open how a broad spectrum of intrinsic motivations interacts with world‑model representation learning to support downstream control, making it hard to identify which intrinsically gathered experiences yield reusable models. To bridge this gap, we embed 24 intrinsic objectives—including established methods and explicit variants spanning curiosity, entropy and competency—into a common model‑based reinforcement learning framework. We evaluate the resulting world models via zero‑shot, few‑shot and multi‑task control in locomotion, manipulation and open‑ended gaming environments, using three representation‑learning strategies: observation reconstruction, latent prediction and redundancy reduction. Results show that curiosity‑driven exploration combined with observation reconstruction yields the strongest overall profile, although the best individual pairing depends on the experiment. In compositional vision‑based tasks, intrinsic motivation derived from both episodic and lifelong novelty—providing a complementary memory to the world model’s generic recurrent state—achieves the best zero‑shot transfer and few‑shot adaptation. These findings motivate joint design of intrinsic motivation and representation learning. Our work demonstrates world models as a powerful tool for measuring the quality of intrinsically motivated experience and positions intrinsic motivation as a key ingredient for autonomously building world models.
Review: The paper offers a comprehensive integration of diverse intrinsic motivations and validates their impact on world‑model learning through a unified experimental setup, providing clear guidance for future autonomous learning system design.