Where should world modeling go next?
Traditionally, a world model is a passive predictor of environment dynamics: given a state and an action, it forecasts what comes next. Useful, but too narrow once the goal is an agent that continually improves through interaction.
Real environments are expensive to access, slow to run, hard to parallelize, and sometimes unsafe. Yet agents need rich, iterative feedback to refine plans, learn from mistakes, and weigh risky actions without paying the full cost of real-world trial and error. So what, exactly, should such an agent learn from?
Costly & slow
Agents need vast trial and error, but real interactions are expensive and slow to run.
Risky & non-rollbackable
Wrong actions in the real world are hard to undo and can cause irreversible harm.
Feedback comes too late
Real feedback is passive and after-the-fact — it only reveals what already happened.
Hard to parallelize
Real environments are limited and can't be replicated at scale like simulators.
Rather than asking what the environment will look like next, we ask what information an agent will receive after an interaction.
From World Model to Agent-Centric World Proxy
We generalize the world model into a World Proxy: an environment-grounded interface an agent can query before acting, training, or updating. It need not stop at the next physical state; it can return an information transition — a rendered view, an execution result, a retrieved memory or skill, or a verdict on a plan — as agent-usable feedback.
Interaction steps, not physical time
An agent interaction does not always advance physical time — it may query a viewpoint, simulate an execution, retrieve an experience, or verify a plan. We index each interaction by a step $\ell$, and treat the proxy's output as agent-usable feedback selected by a proxy function $\mathcal{F}$:
What makes a good proxy?
Not every intermediate module is an effective World Proxy. Within the agent's closed loop, a good proxy should be grounded, controllable, actionable, scalable, and forward-looking.
Groundedness
Relies on real environment data, rules, trajectories, or interaction evidence — not detached generation.
Controllability
Supports agent-initiated queries, actions, or interventions and returns corresponding feedback.
Feedback usefulness
Improves planning, decision making, policy learning, or continual improvement.
Cost & scalability
Reduces real-world interaction cost while scaling safely across tasks and settings.
Forward-looking ability
Anticipates outcomes, counterfactuals, or risks before real execution.
Six Proxy Functions
Each function answers a different question for the agent, but all return agent-usable feedback. Select one on the wheel:
Browse the full curated list on GitHub.
Three Empowerment Levels
Orthogonal to what a proxy returns is how it improves the agent. The same function can graduate from advisor to teacher to partner.
| Level | What it does | Retained across episodes | Representative example |
|---|---|---|---|
| L1 · Guidance | Advises a frozen agent at decision time. | No — within-episode only. | Plan via imagined rollouts; critique a plan. |
| L2 · Training | Generates experience or supervision that updates the agent's parameters. | Yes — in the policy. | Learn a policy in imagination; reward signals for RL. |
| L3 · Co-Evolution | Agent and proxy improve each other in a continual loop. | Yes — both improve. | Online model learning on a robot; a co-evolving proxy. |
One agent, one function, three roles
Consider a web agent and an execution proxy climbing the three levels (illustrative composite):
- L1. The agent imagines the result page of a click before committing — planning against a simulated outcome (WebDreamer-style).
- L2. Those imagined rollouts become synthetic experience that trains the policy with RL (DreamGym-style).
- L3. Agent and proxy improve together: real trajectories sharpen the proxy, and the sharper proxy trains a better agent (WebEvolver-style).
Open Challenges
A proxy is only as useful as it is trustworthy, and trust is the hardest part to guarantee. Four problems set the agenda:
Fidelity & the limits of imagination
Generative models can look convincing while violating the dynamics they claim to model; errors compound over long rollouts. Calibrated uncertainty, not sharper pixels, is the missing ingredient.
Knowing when to trust the proxy
An agent must decide online whether to act on proxy feedback or return to the real environment. Treating the proxy as an oracle invites silent failure.
Reward hacking & safety
When the proxy becomes the reward or verifier, the agent is incentivized to exploit its blind spots — the same sandbox that makes exploration safe opens a new attack surface.
Evaluation that measures information gain
Today's benchmarks grade realism and fidelity in isolation. We need agent-centric benchmarks: did the feedback actually help the agent plan, learn, or improve?
The most valuable world models of the coming years will be judged less by how vividly they dream and more by how much better they make the agents that query them.
Key terms
A compact glossary of the vocabulary this position paper introduces.
An environment-grounded interface an agent can query before acting, training, or updating. It returns agent-usable feedback rather than only the next physical state.
The proxy's output at an interaction step — a rendered view, an execution result, a retrieved memory or skill, or a verdict on a plan — not necessarily the next physical state.
The index of an agent query. Unlike physical time, a step need not advance the world: it may imagine a viewpoint, simulate an execution, or verify a plan.
The type of feedback a proxy returns: Dynamics, Spatial, Execution, Memory, Skill, or Reward / Verification.
How a proxy improves the agent: L1 inference-time guidance, L2 training-time optimization, L3 agent–proxy co-evolution.
The proposed evaluation criterion — how much the feedback helps the agent plan, decide, learn, verify, or improve, rather than how visually realistic it is.
Citation
@article{yang2026worldproxy,
title = {Quo Vadis, World Modeling?},
author = {Yu Yang and Xuemeng Yang and Licheng Wen and Lingdong Kong and Xiaobin Hu and Dongyue Lu and Wei Chow and Xiyan Huang and Yuxiang Feng and Yue Liao and Jianbiao Mei and Daocheng Fu and Rong Wu and Pinlong Cai and Ran Yi and Ying Tai and Jiangning Zhang and Botian Shi and Yong Liu and Shuicheng Yan},
journal = {arXiv preprint arXiv:2608.02713},
year = {2026}
}
Curated reading list: worldbench/awesome-agentic-world-model.
