Should world models predict states or return agent feedback?

The paper redefines what a world model is for. Classical world models are described as pursuing "future physical-state prediction," a formulation "useful yet narrow for agents that require actionable feedback beyond raw state transitions." In its place the authors propose "Agent-Centric Interactive World Proxies," which move "from physical state transitions to agent-usable information transitions, such as execution outcomes, retrieved experiences or skills, and verification signals." The conclusion states the shift as one "from predicting the world to serving the agent": the proxy is "an environment-grounded mechanism that returns the information transition an agent needs," whether "a future state, rendered view, execution result, retrieved memory or skill, or verdict on a plan."

The reasoning starts from a ceiling on static supervision. Agents are mostly taught from "expert trajectories, human annotations, or supervised fine-tuning corpora," data that "cannot answer back when the agent strikes out on its own," so competence is "bounded by the distribution it was trained on." Continual improvement needs "dynamic interaction feedback," yet direct real-environment interaction is "costly, slow, unsafe, and hard to parallelize." A proxy is the intermediate step: agents "query lower-cost, more controllable feedback before committing to real actions." To map the design space, the paper sorts proxies "into six functional forms based on their feedback modalities: dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies." The axis is what kind of feedback comes back to the agent, not how the proxy is built.

This gives an agent-side reading to several neighboring notes. It shares the diagnosis in Can agents learn beyond what their training data shows?, where competence is bounded by curated coverage, but it names a remedy that avoids real-environment cost. It contrasts with What blocks scaling from language models to autonomous agents?, which scales real environments, while this paper treats real interaction as the expensive thing to be ahead of. Can language models learn to simulate agent environments? is next-state prediction at scale. My reading, which the excerpt does not state, is that it sits nearest the "dynamics" form, one of six. Can agent deployment itself generate training signals automatically? and Can agents learn better from their failures than successes? plausibly map onto the execution and memory/experience forms, though that mapping is also mine.

The excerpt is abstract, introduction and conclusion only, and it reads as a conceptual argument: it reports no experiments, benchmarks or comparisons between proxy forms. It does not show that information-transition proxies outperform state predictors, that the six forms are exhaustive, or how a proxy stays faithful to the real environment. The conclusion's list of ways proxies "empower agents" is cut off after its first item, inference-time guidance. What follows at this strength is a vocabulary. The Qwen-AgentWorld line of work can be asked which feedback modality it supplies, and other agent-improvement mechanisms can be classed by the same question. Whether the reframing pays off is not something this excerpt can settle.