A paper proposes Agent-Centric Interactive World Proxies to move world models past physical-state prediction

Key facts

• A paper titled 'Quo Vadis, World Modeling?' (arXiv 2608.02713), posted August 3, 2026, lists 20 authors led by Yu Yang.
• It argues that classical world models, built to predict future physical states, are too narrow for agents that need actionable feedback.
• It proposes Agent-Centric Interactive World Proxies, sorted into six functional forms by feedback type: dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies.
• It lays out three progressive levels of how proxies can empower agents: Level 1 (Inference-Time Guidance), Level 2 (Training-Time Optimization), and Level 3 (Agent-Proxy Co-Evolution).
• The paper reports no experiments or benchmark results and names no existing agent or product already implementing the framework.

Why it matters

Agents meant to keep improving need a steady stream of feedback on what works, not just the static data they were trained on, but according to the paper, getting that feedback by acting directly in the real world is costly, slow, unsafe, and hard to parallelize. World modeling exists to give agents a cheaper, more controllable substitute: query the model instead of the world before committing to a real action. The paper's argument is that the standard way of building that substitute, training a model to predict the next physical state of the world, is useful but too narrow, since it does not hand an agent feedback it can act on beyond raw state transitions. Its proposed fix widens what a world model is allowed to output: instead of only a predicted physical state, an Agent-Centric Interactive World Proxy can hand back execution outcomes, retrieved experience or skills, and verification signals, information an agent can use directly. The authors frame this as a shift from predicting physics to predicting what an agent needs to know, and describe it as a roadmap for building agents that plan better, learn faster, and evolve continually.

Who it affects

The paper is aimed at researchers and engineers building AI agents meant to keep improving after deployment, and specifically at anyone already building one piece of a world model, such as a next-state predictor, a memory or skill retriever, or a reward or verification model, without a shared framework connecting the pieces. It offers that group a common vocabulary, the six functional forms, and a common roadmap, the three levels, rather than a tool to install. No existing agent, product, or company is named in the paper as already using the framework, so the immediate audience is people deciding how to organize their own agent research, not users of a released system.

How to use it

There is no software release to install here: the contribution is a taxonomy and a roadmap, not a product. The paper's own framing suggests using the six-form breakdown (dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies) to classify the feedback sources a team already has, then using the three-level ladder to decide how far to take them. At Level 1, Inference-Time Guidance, a proxy's output simply enriches the context an agent uses to make a decision, with no training involved. At Level 2, Training-Time Optimization, proxy output becomes a reward, a critique, or a synthetic rollout that a policy learns from. At Level 3, Agent-Proxy Co-Evolution, real-environment evidence keeps updating both the proxy and the agent together, so the two improve as a pair. The paper's HuggingFace listing also links a project page and a GitHub repository of related resources (29 stars at last check), styled as a curated list rather than a working implementation of the framework.

How solid is it

This is explicitly a conceptual and position paper. It reports no experiments, benchmarks, or empirical evaluation, and the six forms and three levels are the authors' proposed way of organizing the field rather than a tested one. The paper lists 20 authors; on HuggingFace's paper board it had drawn 30 upvotes and 2 comments at last check, a sign of some reader attention rather than independent verification. The abstract itself names no author institutions. HuggingFace's listing does carry an organization badge for Shanghai AI Laboratory, but that is platform listing metadata rather than an affiliation stated in the paper's own text, so it should not be read as a confirmed claim about who produced the work.

Risks and caveats

The main risk with any taxonomy-and-roadmap paper is that its categories may not survive contact with real systems, and since the paper presents no experiments, that is untested here. No existing agent, product, or system is named as already implementing the full framework, so whether Agent-Proxy Co-Evolution works as described in practice remains open rather than demonstrated. The paper also has not been published at a stated peer-reviewed venue: as of this posting it exists as an arXiv preprint carried on HuggingFace's daily papers board, with no conference or journal named.

“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.”

— the authors, in the paper's abstract