Quo Vadis, World Modeling?

The Paradigm Shift: From Physics Prediction to Agent Feedback

World models have traditionally been framed as tools for predicting future physical states—answering "what will happen next?" But this formulation is too narrow. The authors argue that world modeling should instead center on what agents actually need: actionable feedback that helps them make better decisions, learn faster, and improve continuously. Rather than predicting raw state transitions, world proxies should deliver execution outcomes, retrieved past experiences, skills, verification signals, or reward estimates—information an agent can directly use to improve itself.

This reframing opens the design space significantly. Instead of one type of world model, the authors identify six functional forms based on feedback modality: dynamics proxies (traditional physical state prediction), spatial proxies (scene understanding), execution proxies (outcome forecasting), memory/experience proxies (retrieved past interactions), skill proxies (learned reusable behaviors), and reward/verification proxies (evaluation signals). Each serves different purposes depending on what an agent needs to improve.

Three Levels of Agent Empowerment

The authors organize how these proxies support agent improvement across three progressive levels:

L.1 Inference-Time Guidance is the simplest: proxy outputs enrich the information available when an agent makes decisions. Think of it as giving the agent better context at decision time—retrieved memories, spatial understanding, or execution predictions that inform the next action.

L.2 Training-Time Optimization goes deeper. Here, proxy outputs generate training signals: rewards, critiques, or synthetic rollouts that the agent uses to update its policy. The proxy doesn't just inform decisions; it teaches the agent.

L.3 Agent-Proxy Co-Evolution is the most sophisticated. Both the agent and the proxy improve together based on real-world evidence. When the agent acts in the true environment, that feedback updates both the proxy and the agent's policy in a continuous loop.

Why This Matters

The core motivation is practical: real-world interaction is expensive, slow, unsafe, and hard to parallelize. A well-designed world proxy lets agents simulate, plan, and learn in a cheaper, faster, controllable environment before committing to real actions. But world proxies are only useful if they provide the right kind of information for the agent's improvement loop.

By mapping the design space explicitly—six proxy types across three empowerment levels—the authors give researchers and practitioners a systematic language for thinking about world modeling. Rather than asking "How do we predict the next frame?", the question becomes "What feedback does this agent need, and which proxy type should generate it?" This shift unlocks possibilities beyond pixel prediction: using world proxies for retrieval, skill discovery, reward modeling, and continuous adaptation.

The paper doesn't present empirical breakthroughs on standard benchmarks. Instead, it offers a conceptual roadmap. It resets the conversation around world modeling from a narrow technical problem (accurate state prediction) to a broader strategic one: how to build agent-centric feedback systems that improve continuously, safely, and efficiently.