In a new research paper, Hugging Face researchers argue that continually improving AI agents need dynamic interaction feedback beyond static supervision. However, direct real-environment interaction is costly, slow, unsafe, and difficult to parallelize. World modeling offers a natural intermediate proxy, allowing agents to query lower-cost, more controllable feedback before committing to real actions.
Classical world models primarily predict future physical states, a formulation that is useful but narrow for agents requiring actionable feedback beyond raw state transitions. The paper conceptualizes Agent-Centric Interactive World Proxies, shifting the paradigm from physical state transitions to agent-usable information transitions—such as execution outcomes, retrieved experiences or skills, and verification signals. This broadens world modeling to provide versatile feedback for continually improving agents.
To systematically map this design space, the authors organize world proxies into six functional forms based on feedback modalities: dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies. These characterize the primary ways world modeling serves agent improvement.
The paper further analyzes how these proxies empower agents across three progressive levels: L.1 Inference-Time Guidance, where proxy outputs enrich in-context information for superior decisions; L.2 Training-Time Optimization, where proxy outputs yield rewards, critiques, or synthetic rollouts for policy learning; and L.3 Agent-Proxy Co-Evolution, where real-environment evidence continuously updates both the proxy and the agent for co-evolution.
Ultimately, this work recasts world modeling into an agent-centric paradigm, establishing a roadmap for building world proxies that empower agents to plan better, learn faster, and evolve continually.