World modeling gets a rethink: Agent-centric proxies promise cheaper, faster AI feedback

World modeling has traditionally focused on predicting future physical states, but that's too narrow for agents that need feedback beyond raw transitions. In the arXiv preprint 'Quo Vadis, World Modeling?' (2608.02713), a team of 20 researchers proposes a new paradigm: Agent-Centric Interactive World Proxies. Instead of simulating physics, these proxies simulate agent-usable information like execution outcomes, retrieved experiences, skills, and verification signals. The authors classify world proxies into six functional forms based on feedback modality: dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies. Together, they cover the main ways a world model can serve agent improvement.

The framework then maps how these proxies support agents at three progressive levels: L1, inference-time guidance, where proxy outputs enrich in-context information for better decisions; L2, training-time optimization, where proxy outputs generate rewards, critiques, or synthetic rollouts for policy learning; and L3, agent-proxy co-evolution, where real-environment evidence continuously updates both the proxy and the agent. This layered approach broadens world modeling from a rigid physics simulator to a flexible, feedback-rich environment. The work is a conceptual roadmap for building world proxies that let agents plan better, learn faster, and evolve continually—potentially reducing reliance on costly, slow, or unsafe real-world trials. It's a timely reframing for embodied AI, robotics, and reinforcement learning communities.