Paper deep dive
Quo Vadis, World Modeling?
Yu Yang, Xuemeng Yang, Licheng Wen, Lingdong Kong, Xiaobin Hu, Dongyue Lu, Wei Chow, Xiyan Huang, Yuxiang Feng, Yue Liao, Jianbiao Mei, Daocheng Fu, Rong Wu, Pinlong Cai, Ran Yi, Ying Tai, Jiangning Zhang, Botian Shi, Yong Liu, Shuicheng Yan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/5/2026, 4:27:53 AM
Summary
The paper 'Quo Vadis, World Modeling?' proposes shifting from classical physical-state world models to 'Agent-Centric Interactive World Proxies' that provide diverse, agent-usable information transitions (e.g., execution outcomes, skills, verification). It categorizes these proxies into six functional forms (dynamics, spatial, execution, memory/experience, skill, reward/verification) and analyzes their empowerment of agents across three levels: L1 Inference-Time Guidance, L2 Training-Time Optimization, and L3 Agent-Proxy Co-Evolution.
Entities (15)
Relation Signals (13)
Agent-Centric Interactive World Proxy → enables → L3 Agent-Proxy Co-Evolution
confidence 95% · L.3 Agent-Proxy Co-Evolution, where real-environment evidence continuously updates both the proxy and the agent for co-evolution
Agent-Centric Interactive World Proxy → enables → L1 Inference-Time Guidance
confidence 95% · L.1 Inference-Time Guidance, where proxy outputs enrich in-context information for superior decisions
Agent-Centric Interactive World Proxy → enables → L2 Training-Time Optimization
confidence 95% · L.2 Training-Time Optimization, where proxy outputs yield rewards, critiques, or synthetic rollouts for policy learning
Agent-Centric Interactive World Proxy → includes → Execution Proxy
confidence 95% · organize world proxies into six functional forms... execution... proxies
Agent-Centric Interactive World Proxy → includes → Memory/Experience Proxy
confidence 95% · organize world proxies into six functional forms... memory/experience... proxies
Agent-Centric Interactive World Proxy → includes → Skill Proxy
confidence 95% · organize world proxies into six functional forms... skill... proxies
Agent-Centric Interactive World Proxy → includes → Reward/Verification Proxy
confidence 95% · organize world proxies into six functional forms... reward/verification proxies
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Continually improving agents require dynamic interaction feedback beyond static supervision, yet direct real-environment interaction is costly, slow, unsafe, and hard to parallelize. World modeling offers a natural intermediate proxy that allows agents to query lower-cost, more controllable feedback before committing to real actions. Classical world models instantiate this proxy primarily through future physical-state prediction, a formulation useful yet narrow for agents that require actionable feedback beyond raw state transitions. In this work, we conceptualize Agent-Centric Interactive World Proxies, shifting the fundamental paradigm from physical state transitions to agent-usable information transitions, such as execution outcomes, retrieved experiences or skills, and verification signals, broadening the scope of world modeling to provide versatile feedback for continually improving agents. To systematically map this design space, we organize world proxies into six functional forms based on their feedback modalities: dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies, which together characterize the primary ways world modeling serves agent improvement. We further analyze 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.
Tags
Links
- Source: https://arxiv.org/abs/2608.02713v1
- Canonical: https://arxiv.org/abs/2608.02713v1
Trouble viewing inline? Open PDF directly →
Full Text
106,399 characters extracted from source content.
Expand or collapse full text
[ Path=fonts/, Scale=0.88, Ligatures=NoCommon, BoldFont=Optimistic.ttf, BoldFeatures=FakeBold=1.25 ] -Regular.otf Quo Vadis, World Modeling? Towards Interactive World Proxies for Continually Improving Agents KnowledgeX Lab @ Shanghai AI Laboratory APRIL Lab @ Zhejiang University LV-Lab @ National University of Singapore Continually improving agents require dynamic interaction feedback beyond static supervision, yet direct real-environment interaction is costly, slow, unsafe, and hard to parallelize. World modeling offers a natural intermediate proxy that allows agents to query lower-cost, more controllable feedback before committing to real actions. Classical world models instantiate this proxy primarily through future physical-state prediction, a formulation useful yet narrow for agents that require actionable feedback beyond raw state transitions. In this work, we conceptualize Agent-Centric Interactive World Proxies, shifting the fundamental paradigm from physical state transitions to agent-usable information transitions, such as execution outcomes, retrieved experiences or skills, and verification signals, broadening the scope of world modeling to provide versatile feedback for continually improving agents. To systematically map this design space, we organize world proxies into six functional forms based on their feedback modalities: dynamics, spatial, execution, memory/experience, skill, and reward/verification proxies, which together characterize the primary ways world modeling serves agent improvement. We further analyze 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. Project Page: https://worldbench.github.io/awesome-agentic-world-model GitHub Repo: https://github.com/worldbench/awesome-agentic-world-model Date: Figure 1: Conceptual shift and design space of Agent-Centric World Proxies. We shift from world models for physical state prediction to interactive world proxies for information transition prediction, enabling continual agent improvement through six proxy functions and three empowerment levels: L.1 inference-time guidance, L.2 training-time optimization, and L.3 Agent-Proxy co-evolution. Contents 1 Motivation: Why Improving Agents Need World Modeling 1.1 From Static Supervision to Continual Improvement 1.2 Bottlenecks of Direct Real-Environment Interaction 1.3 World Modeling as an Intermediate Proxy 2 Definition: From World Models to Agent-Centric World Proxies 2.1 Rethinking World Models 2.2 From World Model to World Proxy 2.3 Agent-Centric World Proxy Definition 2.4 What Makes a Good World Proxy? 3 Empowerment: How World Proxies Improve Agents 3.1 L1: Inference-Time Guidance 3.2 L2: Training-Time Optimization 3.3 L3: Agent-Proxy Co-Evolution 4 Instantiations: Functional Forms of Agent-Centric World Proxies 4.1 Overview 4.2 Dynamics Proxy (World Model) 4.3 Spatial Proxy 4.4 Execution Proxy 4.5 Memory / Experience Proxy 4.6 Skill Proxy 4.7 Reward / Verification Proxy 4.8 Putting It Together: Functions Meet Levels 5 Conclusion: Quo Vadis, World Modeling? 6 List of Contributors References 1 Motivation: Why Improving Agents Need World Modeling 1.1 From Static Supervision to Continual Improvement A truly capable agent does more than finish the task in front of it. It explores the unfamiliar, draws feedback from the world, and turns every success and failure into momentum for the next attempt [203, 157]. Capability of this kind is not granted once at training time; it is earned, again and again, through interaction. Yet most agents today are still taught the way students cram for an exam: from static, offline data such as expert trajectories, human annotations, or supervised fine-tuning corpora [122, 10]. Such data captures patterns that already exist, but it cannot answer back when the agent strikes out on its own. The agent’s competence is therefore bounded by the distribution it was trained on, and the new information that only active trial and error can reveal stays out of reach [155]. This is the ceiling that genuine self-improvement must break through. How can an agent move beyond static supervision, gather useful feedback through active interaction, and use it to improve without end? The most direct answer is to set the agent loose in the real environment and let it learn from whatever comes back: Agent→Action→Real Environment→Feedback→Agent Improvement Agent~→~ Action~→~ Real Environment~→~ Feedback~→~ Agent Improvement Figure 2: Basic Agent-Environment Interaction Loop. The agent executes real actions and receives feedback for planning, learning, and improvement. The environment answers with observations, execution results, rewards, or errors, and the agent folds those signals back into its planning, decision making, policy learning, and continual improvement [146, 136, 59]. It is a clean loop, and for a long time it was the whole story. 1.2 Bottlenecks of Direct Real-Environment Interaction The real environment offers the most faithful feedback, and yet it makes a poor sole training ground once an agent needs to err thousands of times, branch across many possibilities at once, and reason several steps ahead before committing. The cracks show up along four recurring axes, collected in Table 1: it is costly, risky, stubbornly backward-looking, and hard to parallelize. In short, the real world is indispensable for grounding but ill-suited to scale. To improve efficiently, an agent needs something in between: a place to try, to predict, and to reason about what-ifs without paying the full price of real execution. Table 1: Bottlenecks of direct real-environment interaction. Bottleneck Explanation Examples Low Training Efficiency and High Cost The agent needs a large amount of trial and error, while real interactions are usually expensive and slow. • Embodied Agents: robotic trial and error incurs device wear and safety costs [173]. • Digital Agents: web, GUI, or game agents consume computation and interaction resources during large-scale sampling [213, 177, 103, 77]. Non-Rollbackable and Risky Interaction Incorrect operations in the real environment are often difficult to undo and may even cause irreversible consequences [198, 94]. • Embodied Agents: robots may collide with or damage objects, and autonomous vehicles may hit pedestrians. • Digital Agents: web agents may incorrectly submit, delete, or send information. Limited Forward-Looking Feedback Real-environment feedback is usually passive and after-the-fact: it reveals only the outcome of an executed action, making future prediction or multi-step reasoning difficult [125]. • Embodied Agents: robots cannot easily know in advance whether an action will cause a collision. • Digital Agents: web or code agents often need to execute before knowing the result. Difficult to Parallelize Real-environment instances are limited and difficult to replicate at scale like simulated environments [111, 150, 134, 149, 90, 121]. • Embodied Agents: physical robot platforms are limited in number. • Digital Agents: online services and user environments are difficult to replicate in batches. 1.3 World Modeling as an Intermediate Proxy That something is a proxy that sits between the agent and the world [57, 146]: Agent↔World Modeling↔Real Environment. Agent~ ~ World Modeling~ ~ Real Environment. Figure 3: Agent-Centric World Modeling as an Intermediate Proxy. World Modeling provides lower-cost feedback between the Agent and the Real Environment. Crucially, World Modeling here is not a stand-in for reality, nor a contest to render the most photorealistic future. Its job is humbler and more useful: to act as an intermediate proxy, grounded in the real environment, that hands the agent interaction feedback in a cheaper, more controllable, and more predictable form. Concretely, such a proxy lets the agent probe its options before committing to any of them. What might happen if I take this action? What would I see from another viewpoint? What would this command or tool call return? Have I faced something like this before, and is the plan safe to run? Each is a question the proxy can answer cheaply, and each spares the agent a costly or irreversible move in the real world. Seen this way, World Modeling for continually improving agents is no longer confined to next-state prediction; it becomes a broader, agent-facing proxy, grounded in real-environment evidence, that can simulate, retrieve, guide, and verify. What makes a useful proxy? What separates such a proxy from a mere simulator or database is what it is optimized for. To genuinely serve a continually improving agent, Agent-Centric World Modeling should meet at least three requirements: Key Requirements 1. Agent-Facing Closed Loop: Support agent-initiated queries, actions, or interventions and return conditioned feedback for decision making and improvement. 2. Real-Environment Grounding: Learn from real-environment data, rules, trajectories, or interaction evidence, approximating the outcomes and feedback relevant to agent decisions. 3. Actionable Information Gain: Optimize for better context, safer decisions, more effective exploration, or higher-quality training signals, not only visual realism or prediction accuracy. 2 Definition: From World Models to Agent-Centric World Proxies 2.1 Rethinking World Models To ask where world modeling should go, it helps to recall where it began. A classical World Model is usually defined as a state transition model [57, 59, 88]: s^t+1=ℳ(st,at). s_t+1=WM(s_t,a_t). That is, at physical time step t, the system is in state sts_t, executes action ata_t, and the World Model predicts the next physical state s^t+1 s_t+1. This form is suitable for robotics, model-based RL, video prediction, and related settings [119, 173, 18]. Its core focus is: Current State+Action→Future State. Current State+ Action→ Future State. In other words, a traditional World Model traces a single thread of state transitions along physical time. Even modern self-supervised variants that predict in a learned latent space [194] rather than in raw pixels [5, 6] inherit this backbone: a state goes in, an action is applied, and a future comes out. For a continually improving agent, however, the feedback worth having reaches well past the next state. Over the course of solving a task, such an agent may need to: • simulate the result of an action, command, API call, or tool call [52, 148, 106]; • retrieve relevant information from memory, experience, or failure cases [124, 140, 123]; • query reusable skills or sub-policies for the current task [157, 2, 98]; • verify whether a plan, trajectory, or action is safe and feasible [32, 99, 112]; • obtain reward, critique, preference, or error-diagnosis signals for training [30, 122, 128]. What unites these is the kind of answer they return. None is simply a next state; each is a different species of information, a consequence, a memory, a skill, a judgment, that the agent can immediately act on. Thus the classical World Model is an indispensable starting point, but only a starting point: it captures one mechanism of Agent-World interaction, not the full repertoire a continually improving agent relies on. 2.2 From World Model to World Proxy To embrace this wider set of interactions, we generalize the World Model into a World Proxy. The move is less a replacement than a broadening of the same core idea: an intermediate mechanism, grounded in the real environment, that supplies feedback the agent would otherwise have to win through direct execution. What changes is the reach. A World Proxy need not stop at predicting the next state; it may also simulate execution results, retrieve experience, offer skill guidance, or verify and evaluate behavior [52, 124, 157, 99]. Figure 4: From World Model to World Proxy. World Proxy is an agent-facing proxy layer that may instantiate as dynamics prediction, spatial synthesis, execution simulation, memory retrieval, skill guidance, or reward / verification feedback. Concretely, this broader view stretches the classical definition along three axes, summarized in Table 2: Table 2: From environment-centric prediction to agent-centric interaction modeling. Rethinking From To Key Idea Physical Time Step → Interaction Step t→t+1t→ t+1 ℓ→ℓ+1 → +1 Agent interaction does not always advance physical time. It may query a viewpoint, simulate execution, retrieve experience, or verify a plan. We use interaction step ℓ to describe each Agent-World interaction. State Transition → Information Transition st→s^t+1s_t→ s_t+1 sℓ→s^ℓ+1s_ → s_ +1 The output of a World Proxy is not limited to next-state prediction. It can be viewed as an information transition, such as a new observation, execution result, retrieved memory or skill, reward, verification feedback, or other agent-usable information. External Condition → Agent-Initiated Interaction external uℓℱu_ ^F The interaction condition should be agent-initiated: a query, action, or intervention actively proposed by the Agent according to its goal and context, rather than only an externally given condition. Note. ℱF denotes different proxy functions, such as dynamics prediction, spatial rendering, execution simulation, memory retrieval, skill guidance, and reward / verification. Therefore, the key question for an Agent-Centric World Proxy is not only whether the prediction is accurate, but: What new information will the agent obtain after this interaction? In this sense, the output of a World Proxy should be organized as agent-usable feedback, rather than only as an environmental state. The shift is subtle but freeing: a memory lookup, a code execution, and a reward estimate can all be treated as one move of the same game, each delivering information the agent did not possess an instant earlier [140, 148, 30]. 2.3 Agent-Centric World Proxy Definition Pulling these threads together, we can now state the idea precisely. We define an Agent-Centric World Proxy as: An Agent-Centric World Proxy is an environment-grounded proxy that models or predicts information transitions conditioned on agent-initiated interactions, aiming to provide information gain for agent improvement. Formally, it is expressed as: s^ℓ+1=(sℓ,uℓℱ),sℓ∈. s_ +1=WP (s_ ,u_ ^F ), s_ . where: • ℓ : interaction step, representing the ℓ -th interaction between the Agent and the World Proxy, not limited to physical time; • S: information state space, a generalized space that may include physical state, observation, memory, knowledge, execution result, verification, guidance, and more; • ℱF: proxy function, such as dynamics prediction, spatial rendering, execution simulation, memory retrieval, skill guidance, or reward / verification; • uℓℱu_ ^F: the query, action, or intervention actively proposed by the Agent under a specific proxy function; • s^ℓ+1 s_ +1: the feedback returned by the World Proxy, such as a future state, novel observation, execution result, retrieved memory / skill, reward, or verification result. A single prediction, though, is not yet improvement. To serve a continually improving agent, the proxy cannot remain a one-shot oracle; it must close the loop with the agent [157, 140, 41]: Agent→uℓℱ→→s^ℓ+1→Agent. Agent~→~u_ ^F → s_ +1~→~ Agent. Figure 5: Agent-in-the-Loop World Proxy. The agent queries the World Proxy, which predicts, simulates, retrieves, or verifies the resulting information transition and returns feedback for agent improvement. At the ℓ -th interaction step, this loop can be divided into four steps: • Step 1: Agent proposes an interaction The Agent actively proposes an interaction request uℓℱu_ ^F according to its current goal, context state, and other conditions. • Step 2: World Proxy predicts information transition The World Proxy predicts or generates feedback conditioned on the current information state sℓs_ and the agent interaction uℓℱu_ ^F. • Step 3: Proxy feedback is returned to the Agent The predicted s^ℓ+1 s_ +1 is returned as new feedback or information gain, such as a future state, predicted observation, execution result, retrieved guidance, or verification result. • Step 4: Agent uses feedback for improvement The Agent uses the feedback for planning, decision making, policy learning, or continual improvement. Run once, this is a single prediction; iterated over many steps, it becomes a trajectory of improvement, since each answer quietly reshapes the next question the Agent thinks to ask. 2.4 What Makes a Good World Proxy? Not every intermediate module is an effective World Proxy. For continually improving agents, a good World Proxy should provide feedback that is grounded, controllable, actionable, scalable, and forward-looking within the agent’s closed loop. Table 3: Criteria for effective World Proxies. Criterion Meaning 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 and Scalability Reduces real-world interaction cost while scaling safely across tasks and settings. Forward-Looking Ability Anticipates outcomes, counterfactuals, or risks before real execution. In practice, useful World Proxies must balance real-environment grounding, scalability, and actionable feedback; proxies that drift too far from grounding soon stop being useful, and confidently wrong ones can be worse than none [153]. Takeaway 1. From World Model to World Proxy: Extending beyond future-state prediction to spatial rendering, execution simulation, memory retrieval, skill guidance, reward modeling, and verification. 2. From Realism to Information Gain: Moving beyond prediction accuracy to prioritize agent-usable feedback for planning, decision making, learning, and continual improvement. 3. From one-shot prediction to closed loop: Use agent-initiated interaction and proxy feedback to drive iterative improvement. 3 Empowerment: How World Proxies Improve Agents With the World Proxy defined, the obvious question is what it actually buys us: How can World Proxies help an Agent plan, learn, and evolve? One way to answer is to grade the world model by its own intrinsic capability, as Chu et al. [31] do: • L.1 Predictor: one-step / local transition prediction; • L.2 Simulator: long-horizon, action-conditioned rollout; • L.3 Evolver: world model self-reflection. Figure 6: L1 to L3 World Proxies for Agent Improvement. L1 augments inference-time context, L2 provides training signals, and L3 enables Agent-Proxy co-evolution. These levels grade the world model’s own competence. Our question is orthogonal: not how capable the model is in isolation, but how much better it makes the agent. We deliberately reuse the L1-L3 shorthand for this agent-centric axis, so the two scales rhyme without being identical; Table 4 sets them side by side to keep the two readings distinct. World-Proxy-Driven Agent Improvement We take an agent-centric perspective: instead of asking how powerful the World Model itself is, we ask how a World Proxy drives Agent improvement. This leads to three levels: L.1 Inference-Time Guidance • Proxy Role: augments inference-time context. • Agent Effect: better decisions. L.2 Training-Time Optimization • Proxy Role: provides reward, verification, or simulation signals. • Agent Effect: optimizes the policy. L.3 Agent-Proxy Co-Evolution • Proxy Role: closes the Agent-Proxy-Environment loop. • Agent Effect: continual co-evolution. Table 4: Two orthogonal readings of the L1-L3 shorthand: a world model’s intrinsic capability [31] versus the agent-centric empowerment used in this article. The rungs align in spirit, not in definition. Level Intrinsic WM Capability (Chu et al.) Agent-Centric Empowerment (Ours) L.1 Predictor: one-step / local transition prediction Inference-Time Guidance: enrich context for better decisions L.2 Simulator: long-horizon, action-conditioned rollout Training-Time Optimization: reshape the agent policy L.3 Evolver: world-model self-reflection Agent-Proxy Co-Evolution: continual mutual improvement Read together, the three levels form a ladder of increasing commitment. L1 leaves the agent untouched and merely informs its next move; L2 reaches into the agent’s parameters and rewrites its policy; L3 lets agent and proxy reshape each other over time. Capability grows at every rung, and so does the burden of proof on the proxy, because the deeper its feedback reaches, the more a mistake costs. 3.1 L1: Inference-Time Guidance Figure 7: L1 Inference-Time Guidance via World Proxy. The Agent queries the World Proxy, which retrieves, simulates, or verifies relevant information. The returned guidance is added to the Agent’s context for better inference-time decisions. We begin with the lightest touch. At L1, the World Proxy never alters the Agent’s parameters; it simply hands over extra context, experience, skill, or verification feedback at inference time, so the decision the Agent is about to make is a better-informed one [169, 85, 189, 188, 62, 52]. Because nothing is retrained, L1 is cheap and fully reversible; its ceiling, though, is the agent’s existing competence, since it can only recombine what the agent already knows. Formally: s^ℓ+1guide=(sℓagent,uℓℱ),sℓagent+=sℓagent⊕s^ℓ+1guide. s_ +1^guide=WP (s_ ^agent,u_ ^F ), s_ ^agent+=s_ ^agent s_ +1^guide. Here, s^ℓ+1guide s_ +1^guide denotes guidance returned by the World Proxy, and ⊕ denotes adding it to the Agent’s current context for information augmentation. Process The flow runs in one direction, from the agent’s question to a richer context for its very next move; Fig. 7 gives the corresponding visual layout: Agent Query ⟶\; \; Proxy Retrieval / Simulation / Verification ⟶\; \; Guidance ⟶\; \; Augmented Context ⟶\; \; Agent Decision Implementations • Memory / Skill Retrieval: retrieves historical experience, strategies, skills, or tool-use rules according to the current task [124, 144, 140, 157, 135, 193]; • Execution Simulation: simulates the result of an action or action sequence before real execution [52, 148, 132]; • Verification Feedback: determines whether the current plan / action is safe, feasible, and consistent with constraints [32, 99, 112]. In practice, this is the agent looking before it leaps: a web agent about to click Purchase can first ask the proxy to imagine the resulting page, then revise its plan if that page shows an error or an unintended charge, all without touching the live site [52]. Key Point L1 lets the Agent gain richer context within the current episode, so it can “see more, reason more accurately, and act more robustly.” 3.2 L2: Training-Time Optimization L2 raises the stakes. The World Proxy now does more than whisper context at inference time; it serves as a reward model, verifier, critic, or simulator, producing the training signals that reshape the Agent’s policy itself [30, 122, 32, 99, 28]. This lifts the agent’s ceiling rather than merely its current context, but the gain is only as trustworthy as the signal behind it: a biased reward quietly teaches biased behavior. Formally: s^ℓ+1opt=(sℓagent,uℓℱ),agent+=Train(agent,s^ℓ+1opt). s_ +1^opt=WP (s_ ^agent,u_ ^F ), ^+=Train (agent, s_ +1^opt ). Here, s^ℓ+1opt s_ +1^opt denotes training signals generated by the World Proxy, such as reward, verification, critique, or simulated rollouts. These signals can be converted into objectives such as SFT, DPO [128], PPO [137], or GRPO [138]. Compared with L1, the key change in L2 is: Proxy Output as Context ⟶\; \; Proxy Output as Training Signal Figure 8: L2 Training-Time Optimization Driven by World Proxy. The Agent generates rollouts, which are evaluated, verified, or simulated by the World Proxy. The resulting reward, critique, preference, or verification signal is converted into a training objective to optimize the Agent policy. Process Now the loop bends back into learning: the proxy’s verdict on a rollout becomes a gradient on the policy rather than a hint for the moment; Fig. 8 gives the corresponding visual layout. Agent Rollout ⟶\; \; Proxy Verification / Simulation ⟶ Reward / Critique / Preference ⟶\; \; Training Objective ⟶\; \; Agent Policy Optimization Implementations • Proxy-as-Reward: The World Proxy or verifier scores trajectories to form rewards [30, 122, 10]. • Proxy-as-Critic: Identifies failure causes and outputs critiques, error diagnoses, or constraint violations [112, 109, 200]. • Proxy-as-Simulator: Generates synthetic trajectories or constructs preference pairs for DPO / RLHF / GRPO [28, 168, 161]. • Proxy-Guided Curriculum: Generates harder or more targeted tasks based on the Agent’s current failure modes [167, 206]. In practice, the agent’s own rollouts become training fuel: the proxy scores, verifies, or replays them into synthetic trajectories and preference pairs, so a policy can be optimized at a scale that collecting real interactions could never reach [28, 168]. Key Point L2 upgrades the World Proxy from an “inference-time advisor” to a “training-time judge, teacher, or simulated environment,” directly optimizing the Agent policy. 3.3 L3: Agent-Proxy Co-Evolution Figure 9: L3 Agent-Proxy Co-Evolution. Real-environment evidence updates the World Proxy, while useful proxy knowledge is distilled back into the Agent policy for continual improvement. L3 completes the arc by closing a continual loop among the Agent, the World Proxy, and the real environment. Real trajectories, failures, and fresh discoveries update the proxy; the sharpened proxy then guides, verifies, and trains the Agent, so both improve together rather than one after the other [157, 140, 41, 130]. This is the most powerful rung, but it also requires machinery that keeps proxy and policy aligned as both change. Formally: s^ℓ+1proxy=(sℓagent,uℓℱ),(agent+,+)=CoEvolve(agent,,s^ℓ+1proxy,sℓ+1env) s_ +1^proxy=WP (s_ ^agent,u_ ^F ), (agent^+,WP^+)=CoEvolve (agent,WP, s_ +1^proxy,s_ +1^env ) Here, s^ℓ+1proxy s_ +1^proxy denotes proxy feedback, sℓ+1envs_ +1^env denotes real-environment evidence, and CoEvolve(⋅)CoEvolve(·) updates both the Agent and the World Proxy. Process The loop now uses real-environment gaps as the update signal; Fig. 9 gives the corresponding visual layout: Agent Rollout ⟶\; \; Proxy Prediction / Verification ⟶\; \; Real-Environment Evidence ⟶\; \; Disagreement / Error Diagnosis ⟶\; \; Update Proxy + Distill to Agent ⟶\; \; Co-Evolution Implementations • Proxy → Agent Internalization: memory, skills, constraints, verification rules, or reward signals are distilled into Agent parameters or policies, turning external feedback into internal capability [157, 98, 2]. • Agent → Proxy Update: real interaction trajectories, effective experiences, failure cases, and environmental feedback update the World Proxy’s memory, skill library, verifier, simulator, or reward model [140, 41, 20]. In practice, picture a web agent whose every deployment both draws on the proxy’s predictions and feeds fresh trajectories back into it: proxy and policy are retrained in tandem, each cycle shrinking the gap between what the agent imagines and what the world actually returns [41]. Key Point L3 makes the World Proxy more than a static tool: it becomes a co-evolving partner that is continually updated through real-environment evidence. Worked Example: One Web Agent up the Ladder To see the three levels as one continuum rather than three separate tricks, follow a single web agent completing a multi-step online purchase, and watch the same execution proxy deepen its role at each rung: • L.1 Guidance. Before clicking Purchase, the agent asks the proxy to imagine the resulting page; if that page shows an error or an unintended charge, it revises its plan, without ever touching the live site [52]. • L.2 Optimization. Those imagined rollouts are not discarded: the proxy scores and replays them into synthetic trajectories and preference pairs, and the agent’s policy is optimized on this fabricated experience at a scale live interaction could never reach [28]. • L.3 Co-Evolution. Once deployed, the agent’s real trajectories flow back to retrain the proxy, and the sharpened proxy in turn yields better guidance and training signal, each cycle shrinking the gap between what the agent imagines and what the web actually returns [41]. One agent, one proxy function, three escalating roles: advisor, teacher, and partner. Takeaway World Proxies empower agents through three progressive levels: • L.1 Inference-Time Guidance. The World Proxy provides memory, skill, simulation, or verification feedback to enrich inference-time context and support better decisions. • L.2 Training-Time Optimization. The World Proxy acts as a reward model, critic, verifier, or simulator, generating training signals that directly optimize the Agent policy. • L.3 Agent-Proxy Co-Evolution. Real-environment evidence updates the World Proxy, whose knowledge is distilled back into the Agent for continual improvement. 4 Instantiations: Functional Forms of Agent-Centric World Proxies If Section 3 asked how a World Proxy helps, this section asks in what form it appears. The single symbol ℱF in our definition quietly stands in for a whole family of proxy functions; here we unpack it into six concrete forms: dynamics prediction, spatial observation, execution simulation, memory retrieval, skill guidance, and reward/verification. Each has grown into a research area in its own right, and several recent surveys map them in depth [215, 95, 44, 151]. Figure 10: Functional forms of Agent-Centric World Proxies. A World Proxy may simulate dynamics, render spatial observations, predict execution outcomes, retrieve memory, suggest skills, or provide reward / verification feedback. 4.1 Overview The six forms differ less in their machinery than in the question each one answers for the agent. Table 5 lines them up at a glance, pairing every function with the input it consumes, the slice of the world it stands in for, and what it hands back; we then take each in turn. 4.2 Dynamics Proxy (World Model) Formula s^ℓ+1,r^ℓ+1=dyn(sℓ,uℓdyn). s_ +1, r_ +1=WP^dyn (s_ ,u_ ^dyn ). Meaning The Dynamics Proxy is the classical form of a World Model. Given the current state, history, and an agent action or future query, it predicts the future state and may also predict reward. It is the proxy in its most literal sense, a learned stand-in for the dynamics of the environment, and the closest of the six forms to the textbook world model. Table 5: Functional forms of Agent-Centric World Proxies. Proxy Function Agent Input What it Proxies Proxy Output Typical Examples Dynamics (World Model) state / history + action / future query real-world dynamics and temporal transitions future state, rollout, predicted reward video prediction, action-conditioned generation, robotics dynamics model, model-based RL Spatial scene context + viewpoint / pose / location query spatial observation under alternative viewpoints novel observation, rendered view, spatial representation NeRF, 3D Gaussian Splatting, novel view synthesis, spatial imagination Execution code / command / click / API call / tool call consequences of executable interactions in digital environments execution result, state change, stdout / stderr browser simulator, GUI simulator, code execution predictor, API response simulator Memory / Experience task context + retrieval query reusable past interaction evidence retrieved experience, failure cases, constraints experience memory, reflection memory, failure memory Skill goal / context + skill query reusable action knowledge or behavior priors skill suggestion, action prior skill library, reusable behavior module Reward / Verification plan / trajectory / answer / action evaluation feedback, preferences, criteria reward, critique, preference, verification reward model, verifier, critic, trajectory evaluator, LLM-as-Judge Here: • sℓs_ : current state or historical observations; • uℓdynu_ ^dyn: an action, action sequence, or future query proposed by the Agent; • s^ℓ+1 s_ +1: predicted future state; • r^ℓ+1 r_ +1: predicted reward (optional). When the interaction step ℓ aligns with the physical time step t, this form reduces to classical next-state prediction: s^t+1=ℳ(st,at). s_t+1=WM(s_t,a_t). Typical Examples • video prediction / future frame prediction [63, 156, 66, 182, 8, 102, 172]; • action-conditioned video generation [19, 184, 18, 1, 50, 35, 9, 113, 187, 207, 49, 172, 154, 192]; • interactive game-world simulation [154, 35, 54, 25, 192, 50]; • robotics dynamics prediction [173, 59, 60, 61, 6, 57, 76]; • autonomous driving trajectory prediction [47, 67, 166, 211, 15, 181, 118, 204, 97, 179, 133, 185, 114]; • model-based reinforcement learning [57, 136, 59, 76, 116, 4, 205, 58, 79]. Key Point A Dynamics Proxy focuses on: how the environment would change if the Agent executed a given action. 4.3 Spatial Proxy Formula o^ℓ+1view=spatial(sℓ,uℓspatial). o_ +1^view=WP^spatial (s_ ,u_ ^spatial ). Meaning The Spatial Proxy generates observations under spatial or viewpoint conditions. The Agent queries a new location, camera pose, or viewpoint, and the World Proxy returns the corresponding observation or spatial representation. In effect it lets the agent look before it moves, turning an expensive physical relocation into a cheap query about what a yet-unseen vantage point would reveal. Here: • sℓs_ : known visual, geometric, or spatial information; • uℓspatialu_ ^spatial: the queried viewpoint, pose, or spatial position; • o^ℓ+1view o_ +1^view: the predicted or rendered observation from that viewpoint. Typical Examples • NeRF-based neural rendering [117, 13, 190, 120, 26, 14, 147]; • 3D Gaussian Splatting [81, 171, 107, 195, 27, 42]; • 3D/4D scene reconstruction [86, 165, 89, 96, 202, 158, 163, 183, 159]; • 3D/4D scene generation [186, 145, 45, 64, 201, 176, 29, 191, 92, 170]; • visual imagination in navigation (Navigation WM [12, 83, 38, 69, 37]); • spatial reasoning and manipulation [199, 164, 209, 105, 24, 74, 208, 71, 22, 160, 197, 141, 142, 73, 17, 16]. Key Point A Spatial Proxy focuses on: what the Agent would observe from another viewpoint or position. 4.4 Execution Proxy Formula s^ℓ+1exec,y^ℓ+1exec=exec(sℓ,uℓexec). s_ +1^exec, y_ +1^exec=WP^exec (s_ ,u_ ^exec ). Meaning The Execution Proxy simulates executable interactions. The Agent issues code, a command, a web click, an API call, or a tool call, and the World Proxy predicts the resulting state and feedback. Where a dynamics proxy models continuous physics, an execution proxy models the discrete and often brittle logic of digital systems, in which a single misplaced character or click can flip the outcome entirely. Here: • sℓs_ : current web, code, file, program, or tool state; • uℓexecu_ ^exec: the executable interaction issued by the Agent; • s^ℓ+1exec s_ +1^exec: predicted post-execution state; • y^ℓ+1exec y_ +1^exec: predicted feedback, such as stdout, stderr, an error message, a test result, or a page change. Typical Examples • Browser / Web interaction simulator: simulates page states and feedback after clicks, inputs, or navigation [52, 23, 175, 43, 48, 139]; • GUI environment simulator: predicts interface changes after button clicks, window switching, or form filling [108, 21, 53, 212, 93, 84, 3, 68]; • Code execution predictor: predicts stdout, error messages, or test results after running code [148, 34, 33, 110, 129, 20]; • Shell / file-system transition model: predicts file states, stdout, or stderr after command execution [132, 53, 148, 34, 33, 110, 129, 177]; • API / tool-call response simulator: predicts the returned result after an API or tool call [131, 46, 55, 106, 70, 135, 62]. Key Point An Execution Proxy focuses on: what the digital environment would return after the Agent performs an operation. 4.5 Memory / Experience Proxy Formula m^ℓ+1=mem(sℓ,uℓmem). m_ +1=WP^mem (s_ ,u_ ^mem ). Meaning The Memory / Experience Proxy retrieves task-relevant information from past interactions, trajectories, failures, or constraints. Its output is not a physical state, but experiential feedback for planning and decision making. Here: • sℓs_ : current task context, Agent memory, or environment information; • uℓmemu_ ^mem: the retrieval query issued by the Agent; • m^ℓ+1 m_ +1: retrieved experience, failure case, constraint, or risk hint. An ordinary static memory store is not necessarily a World Proxy. It becomes a Memory / Experience Proxy only when a store-retrieve dynamic system returns environment-, task-, or decision-relevant information in response to the Agent’s query and feeds it back into planning, decision making, or improvement. A generative agent recalling the most relevant slices of its past before it acts, or a reflective agent remembering exactly why its last attempt failed, both fall under this view. Typical Examples • Experience-based world memory: retrieves reusable experience from successful or failed trajectories [124, 28, 126, 140, 144, 123, 174]; • Constraint / failure memory: retrieves constraints, risks, and avoidance strategies from past failures [130, 140, 20, 28, 168, 41]; • Reflection memory model: returns prior mistakes, corrective feedback, or improvement hints [140, 144, 124, 123, 130, 126]. Key Point A Memory / Experience Proxy focuses on: the prior experience, failure cases, or constraints the Agent needs for the current decision. 4.6 Skill Proxy Formula g^ℓ+1skill=skill(sℓ,uℓskill). g_ +1^skill=WP^skill (s_ ,u_ ^skill ). Meaning The Skill Proxy retrieves or recommends reusable behavioral knowledge. While the Memory / Experience Proxy emphasizes what has happened before, the Skill Proxy emphasizes what the Agent can do now. It recommends reusable skills, tool-use routines, or action priors based on the current goal, task context, and environment state. An open-world agent, for instance, can bank a routine it once worked out, crafting a particular tool or completing a multi-step form, and later summon it whole rather than rediscovering it move by move. Here: • sℓs_ : current task, goal, environment state, or Agent context; • uℓskillu_ ^skill: the skill or policy query issued by the Agent; • g^ℓ+1skill g_ +1^skill: a retrieved skill, tool-use routine, reusable behavior module, or action prior. Typical Examples • Skill library: retrieves executable skills for the current task [157, 75, 2, 98, 72, 20]; • Tool-use routine: retrieves tool-calling workflows or operation templates [131, 135, 46, 55, 62, 103]; • Reusable behavior module: provides a reusable behavior strategy [157, 75, 2, 98, 72, 149]. Key Point A Skill Proxy focuses on: the reusable skills or behavior strategies available for the current task. 4.7 Reward / Verification Proxy Formula v^ℓ+1eval=eval(sℓ,uℓeval). v_ +1^eval=WP^eval (s_ ,u_ ^eval ). Meaning The Reward / Verification Proxy evaluates an Agent’s behavior, trajectory, answer, or plan. It acts as a reward model, verifier, critic, or evaluator. Rather than telling the agent what the world will look like next, it tells the agent how good its behavior is, collapsing an entire rollout into a single, actionable verdict. Here: • sℓs_ : current task, context, environment information, or historical trajectory; • uℓevalu_ ^eval: the plan, trajectory, answer, or action submitted by the Agent; • v^ℓ+1eval v_ +1^eval: predicted reward, verification result, critique, preference, or failure reason. This is a feedback-oriented World Proxy: it does not need to simulate the full environment, but predicts how the Agent’s behavior would be judged by the environment, rules, or evaluation system. Typical Examples • reward model [30, 216, 143, 122, 10, 11]; • preference model [128, 7, 196, 40, 65, 115, 11]; • verifier [32, 152, 99, 162, 138, 112]; • critic model [112, 109, 140, 200, 11, 104]; • trajectory evaluator [91, 127, 87, 213, 177, 103, 153]; • LLM-as-Judge [210, 104, 82, 39, 51, 100]. Key Point A Reward / Verification Proxy focuses on: whether the Agent’s behavior is correct, safe, and feasible, and how it should improve. 4.8 Putting It Together: Functions Meet Levels The two axes of this article are orthogonal: any proxy function (Section 4) can empower an agent at any level (Section 3). Reading the six functions against L1-L3 turns the design space into a simple map, with representative systems in each cell (Table 6). The mapping is illustrative rather than exclusive, since many systems span more than one level, and the sparser cells mark territory that remains largely open. Table 6: Functions × Levels. Representative ways each proxy function empowers agents across L1, L2, and L3. Sparse cells flag underexplored directions. Function L.1 Inference-Time Guidance L.2 Training-Time Optimization L.3 Agent-Proxy Co-Evolution Dynamics plan via imagined rollouts [136, 62] learn a policy in imagination [60, 76] online model learning on a real robot [173] Spatial imagine views to navigate [12, 83] train on synthesized observations [214] underexplored Execution simulate actions before acting [52, 23] synthesize experience for RL [28] self-improve via a coevolving model [41] Memory retrieve experience at decision time [124, 144] learn from past successes and failures [130, 174] memory updated, then distilled back [20] Skill reuse skills from a library [157] create and train new skills [75] skill library grows with the agent [157, 75] Reward / Verif. critique a plan at inference [200, 100] reward / verifier signals for RL [168, 161] emerging Two patterns stand out. Reading down a column shows that a single level admits many functional realizations; reading across a row shows that the same function can graduate from advisor to teacher to partner as it climbs L1 to L3. The blank corners, spatial and reward proxies at the co-evolution level, are not accidents but invitations. Takeaway Different proxy functions, one purpose: agent-usable feedback. • Beyond future-state prediction: a World Proxy may predict dynamics, render spatial observations, simulate execution, retrieve experience, suggest skills, or provide reward / verification feedback. • Function defines feedback type: dynamics, spatial, execution, memory, skill, and verification proxies differ in what information transition they approximate for the Agent. • Actionability is the common criterion: regardless of function, the output should help the Agent plan, decide, learn, verify, or continually improve. 5 Conclusion: Quo Vadis, World Modeling? We opened with a question, so let us close with an answer. For most of its history, world modeling has been pursued as the art of predicting the world: given a state and an action, render the next frame as faithfully as possible [36]. This article has argued for a quieter but consequential shift, from predicting the world to serving the agent. Once the goal becomes continual improvement, the right object is no longer a state predictor but an Agent-Centric World Proxy: 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. That reframing organized the rest of the story. We saw why real environments alone cannot carry continual improvement, what changes when physical state transitions become interactive information transitions, how proxies empower agents across L.1 inference-time guidance, L.2 training-time optimization, and L.3 Agent-Proxy co-evolution, and in what forms they appear: dynamics, spatial, execution, memory, skill, and reward / verification. The unifying thread is not visual realism for its own sake, but actionable information gain. The same reframing also sharpens the open problems. A proxy is only as useful as it is trustworthy, and trust remains the hardest part to guarantee. Open Challenges 1. Fidelity and the limits of imagination. Generative models can look convincing while violating the dynamics they claim to model [153, 80, 56, 78]. Errors compound over long rollouts; calibrated uncertainty, not just sharper pixels, is the missing ingredient. 2. Knowing when to trust the proxy. An agent must decide, online, whether to act on proxy feedback or return to the real environment. Today agents rarely make that judgment well [125]; treating the proxy as an oracle invites silent failure. 3. Reward hacking and safety. When the proxy becomes the reward or verifier (L2), the agent is incentivized to exploit blind spots. The same sandbox that makes risky exploration safe also opens a new attack surface [198, 94, 178, 101, 180]. 4. Evaluation that measures information gain. Current benchmarks score realism, fidelity, controllability, or human-aligned quality [91, 127, 87], but still grade the proxy in isolation. We need agent-centric benchmarks: did the feedback help the agent plan, learn, or improve? None of these are reasons for pessimism; they are the agenda. Each becomes tractable once we stop asking a world model to be a perfect mirror and start asking it to be a useful interlocutor, one whose answers are grounded, calibrated, and continually corrected by contact with reality. The L3 loop, where real-environment evidence keeps the proxy honest and the proxy keeps the agent improving, is as much a safety mechanism as a learning one. So, quo vadis? We expect the most valuable world models of the coming years to be judged less by how vividly they dream and more by how much better they make the agents that query them. If this article nudges the conversation from building world simulators toward building world proxies that agents can learn from, it will have served its purpose. 6 List of Contributors • Concept & Design: Yu Yang, Xuemeng Yang, Licheng Wen • Writing & Editing: Yu Yang, Xuemeng Yang, Licheng Wen, Lingdong Kong, Xiaobin Hu, Dongyue Lu, Wei Chow • Figures & Visual Design: Xiyan Huang, Yuxiang Feng • Discussion & Insights: Yue Liao, Jianbiao Mei, Daocheng Fu, Rong Wu, Pinlong Cai, Ran Yi, Ying Tai, Jiangning Zhang • Advising: Botian Shi, Yong Liu, Shuicheng Yan References Agarwal et al. [2025] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, et al. Cosmos world foundation model platform for physical AI. arXiv preprint arXiv:2501.03575, 2025. URL https://arxiv.org/abs/2501.03575. Ahn et al. [2022] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, et al. Do as i can, not as i say: Grounding language in robotic affordances. In Conf. Robot Learn., 2022. URL https://arxiv.org/abs/2204.01691. Ai et al. [2026] Jiaxin Ai, Tao Hu, Xuemeng Yang, Shu Zou, Hairong Zhang, Daocheng Fu, Yu Yang, Hongbin Zhou, Nianchen Deng, Pinlong Cai, et al. ComAct: Reframing professional software manipulation via COM-as-Action paradigm. arXiv preprint arXiv:2606.13239, 2026. URL https://arxiv.org/abs/2606.13239. Alonso et al. [2024] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in Atari. In Adv. Neural Inf. Process. Syst., volume 37, 2024. URL https://arxiv.org/abs/2405.12399. Assran et al. [2023] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2023. URL https://arxiv.org/abs/2301.08243. Assran et al. [2025] Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint arXiv:2506.09985, 2025. URL https://arxiv.org/abs/2506.09985. Azar et al. [2024] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. A general theoretical paradigm to understand learning from human preferences. In Int. Conf. Artif. Intell. Stat., 2024. URL https://arxiv.org/abs/2310.12036. Babaeizadeh et al. [2021] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. FitVid: Overfitting in pixel-level video prediction. arXiv preprint arXiv:2106.13195, 2021. URL https://arxiv.org/abs/2106.13195. Bai et al. [2025] Jianhong Bai, Menghan Xia, Xiao Fu, Xintao Wang, Lianrui Mu, Jinwen Cao, Zuozhu Liu, Haoji Hu, Xiang Bai, Pengfei Wan, and Di Zhang. ReCamMaster: Camera-controlled generative rendering from a single video. arXiv preprint arXiv:2503.11647, 2025. URL https://arxiv.org/abs/2503.11647. Bai et al. [2022a] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022a. URL https://arxiv.org/abs/2204.05862. Bai et al. [2022b] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, et al. Constitutional AI: Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073, 2022b. URL https://arxiv.org/abs/2212.08073. Bar et al. [2025] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 15791–15801, 2025. URL https://arxiv.org/abs/2412.03572. Barron et al. [2021] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF Int. Conf. Comput. Vis., pages 5835–5844, 2021. URL https://arxiv.org/abs/2103.13415. Barron et al. [2022] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 5470–5479, 2022. URL https://arxiv.org/abs/2111.12077. Bian et al. [2025] Hengwei Bian, Lingdong Kong, Haozhe Xie, Liang Pan, Yu Qiao, and Ziwei Liu. DynamicCity: Large-scale 4D occupancy generation from dynamic scenes. In Int. Conf. Learn. Represent., 2025. URL https://arxiv.org/abs/2410.18084. Brohan et al. [2023a] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Sergey Levine, Yao Lu, Igor Mordatch, Pierre Sermanet, Ted Xiao, Peng Xu, Tianhe Yu, Brianna Zitkovich, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. In Conf. Robot Learn., 2023a. URL https://arxiv.org/abs/2307.15818. Brohan et al. [2023b] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Sergey Levine, Yao Lu, Igor Mordatch, Ofir Nachum, Carolina Parada, Pierre Sermanet, Ted Xiao, Peng Xu, Tianhe Yu, Brianna Zitkovich, et al. RT-1: Robotics transformer for real-world control at scale. In Robot. Sci. Syst., 2023b. URL https://arxiv.org/abs/2212.06817. Brooks et al. [2024] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, et al. Video generation models as world simulators. Technical report, OpenAI, 2024. URL https://openai.com/research/video-generation-models-as-world-simulators. Bruce et al. [2024] Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, et al. Genie: Generative interactive environments. In Int. Conf. Mach. Learn., pages 4603–4623, 2024. URL https://arxiv.org/abs/2402.15391. Butt et al. [2024] Natasha Butt, Blazej Manczak, Auke Wiggers, Corrado Rainone, David W. Zhang, Michaël Defferrard, and Taco Cohen. CodeIt: Self-improving language models with prioritized hindsight replay. In Int. Conf. Mach. Learn., pages 5013–5034, 2024. URL https://arxiv.org/abs/2402.04858. Cao et al. [2026] Yilin Cao, Yufeng Zhong, Zhixiong Zeng, Liming Zheng, Jing Huang, Haibo Qiu, Peng Shi, Wenji Mao, and Guanglu Wan. MobileDreamer: Generative sketch world model for GUI agent. arXiv preprint arXiv:2601.04035, 2026. URL https://arxiv.org/abs/2601.04035. Cao et al. [2025] Ziang Cao, Fangzhou Hong, Zhaoxi Chen, Liang Pan, and Ziwei Liu. PhysX-Anything: Simulation-ready physical 3D assets from single image. arXiv preprint arXiv:2511.13648, 2025. URL https://arxiv.org/abs/2511.13648. Chae et al. [2025] Hyungjoo Chae, Namyoung Kim, Kai Tzu-iunn Ong, Minju Gwak, Gwanwoo Song, Jihoon Kim, Sunghwan Kim, Dongha Lee, and Jinyoung Yeo. Web agents with world models: Learning and leveraging environment dynamics in web navigation. In Int. Conf. Learn. Represent., 2025. URL https://arxiv.org/abs/2410.13232. Chai et al. [2025] Ying Chai, Litao Deng, Ruizhi Shao, Jiajun Zhang, Kangchen Lv, Liangjun Xing, Xiang Li, Hongwen Zhang, and Yebin Liu. GAF: Gaussian action field as a 4D representation for dynamic world modeling in robotic manipulation. arXiv preprint arXiv:2506.14135, 2025. URL https://arxiv.org/abs/2506.14135. Che et al. [2025] Haoxuan Che, Xuanhua He, Quande Liu, Cheng Jin, and Hao Chen. GameGen-X: Interactive open-world game video generation. In Int. Conf. Learn. Represent., 2025. URL https://arxiv.org/abs/2411.00769. Chen et al. [2022] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In Eur. Conf. Comput. Vis., pages 333–350, 2022. URL https://arxiv.org/abs/2203.09517. Chen and Wang [2024] Guikun Chen and Wenguan Wang. A survey on 3D gaussian splatting. ACM Comput. Surv., 2024. URL https://arxiv.org/abs/2401.03890. Chen et al. [2025] Zhaorun Chen, Zhuokai Zhao, Kai Zhang, Bo Liu, Qi Qi, Yifan Wu, Tarun Kalluri, Sara Cao, et al. Scaling agent learning via experience synthesis. arXiv preprint arXiv:2511.03773, 2025. URL https://arxiv.org/abs/2511.03773. Chen et al. [2023] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. SceneDreamer: Unbounded 3D scene generation from 2D image collections. IEEE Trans. Pattern Anal. Mach. Intell., 2023. URL https://arxiv.org/abs/2302.01330. Christiano et al. [2017] Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Adv. Neural Inf. Process. Syst., 2017. URL https://arxiv.org/abs/1706.03741. Chu et al. [2026] Meng Chu, Xuan Billy Zhang, Kevin Qinghong Lin, Lingdong Kong, Jize Zhang, Teng Tu, Weijian Ma, Ziqi Huang, et al. Agentic world modeling: Foundations, capabilities, laws, and beyond. arXiv preprint arXiv:2604.22748, 2026. URL https://arxiv.org/abs/2604.22748. Cobbe et al. [2021] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. URL https://arxiv.org/abs/2110.14168. Copet et al. [2025] Jade Copet, Quentin Carbonneaux, Gal Cohen, Jonas Gehring, Jacob Kahn, Jannik Kossen, Felix Kreuk, Emily McMilin, Michel Meyer, Yuxiang Wei, David Zhang, Kunhao Zheng, Jordi Armengol-Estapé, Pedram Bashiri, Maximilian Beck, et al. CWM: An open-weights LLM for research on code generation with world models. arXiv preprint arXiv:2510.02387, 2025. URL https://arxiv.org/abs/2510.02387. Dainese et al. [2024] Nicola Dainese, Matteo Merler, Minttu Alakuijala, and Pekka Marttinen. Generating code world models with large language models guided by Monte Carlo tree search. In Adv. Neural Inf. Process. Syst., volume 37, 2024. URL https://arxiv.org/abs/2405.15383. Decart et al. [2024] Decart, Julian Quevedo, Quinn McIntyre, Spruce Campbell, Xinlei Chen, and Robert Wachen. Oasis: A universe in a transformer. Blog post, 2024. URL https://oasis-model.github.io. Ding et al. [2025] Jingtao Ding, Yunke Zhang, Yu Shang, Jie Feng, Yuheng Zhang, Zefang Zong, Yuan Yuan, Hongyuan Su, et al. Understanding world or predicting future? A comprehensive survey of world models. ACM Comput. Surv., 2025. URL https://arxiv.org/abs/2411.14499. Dong et al. [2025] Yifei Dong, Fengyi Wu, Guangyu Chen, Lingdong Kong, Xu Zhu, Qiyu Hu, Yuxuan Zhou, Jingdong Sun, Jun-Yan He, Qi Dai, Alexander G. Hauptmann, and Zhi-Qi Cheng. Towards unified world models for visual navigation via memory-augmented planning and foresight. arXiv preprint arXiv:2510.08713, 2025. URL https://arxiv.org/abs/2510.08713. Dong et al. [2026] Yifei Dong, Fengyi Wu, Yilong Dai, Lingdong Kong, Guangyu Chen, Xu Zhu, Qiyu Hu, Tianyu Wang, et al. Language-conditioned world modeling for visual navigation. arXiv preprint arXiv:2603.26741, 2026. URL https://arxiv.org/abs/2603.26741. Dubois et al. [2024] Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. Length-controlled AlpacaEval: A simple way to debias automatic evaluators. In Conf. Lang. Model., 2024. URL https://arxiv.org/abs/2404.04475. Ethayarajh et al. [2024] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. KTO: Model alignment as prospect theoretic optimization. In Int. Conf. Mach. Learn., 2024. URL https://arxiv.org/abs/2402.01306. Fang et al. [2025] Tianqing Fang, Hongming Zhang, Zhisong Zhang, Kaixin Ma, Wenhao Yu, Haitao Mi, and Dong Yu. WebEvolver: Enhancing web agent self-improvement with coevolving world model. In Proc. Conf. Empir. Methods Nat. Lang. Process., pages 8959–8975, 2025. URL https://arxiv.org/abs/2504.21024. Fei et al. [2024] Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3D gaussian as a new era: A survey. IEEE Trans. Vis. Comput. Graph., 2024. URL https://arxiv.org/abs/2402.07181. Feng et al. [2025a] Jichen Feng, Yifan Zhang, Chenggong Zhang, Yifu Lu, Shilong Liu, and Mengdi Wang. Web world models. arXiv preprint arXiv:2512.23676, 2025a. URL https://arxiv.org/abs/2512.23676. Feng et al. [2025b] Tuo Feng, Wenguan Wang, and Yi Yang. A survey of world models for autonomous driving. arXiv preprint arXiv:2501.11260, 2025b. URL https://arxiv.org/abs/2501.11260. Fridman et al. [2023] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. SceneScape: Text-driven consistent scene generation. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2302.01133. Ganapavarapu and Patel [2026] Giridhar Ganapavarapu and Dhaval Patel. MCP-Cosmos: World model-augmented agents for complex task execution in MCP environments. arXiv preprint arXiv:2605.09131, 2026. URL https://arxiv.org/abs/2605.09131. Gao et al. [2024] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. In Adv. Neural Inf. Process. Syst., volume 37, 2024. URL https://arxiv.org/abs/2405.17398. Gao et al. [2025] Yifei Gao, Junhong Ye, Jiaqi Wang, and Jitao Sang. WebSynthesis: World-model-guided MCTS for efficient WebUI-trajectory synthesis. arXiv preprint arXiv:2507.04370, 2025. URL https://arxiv.org/abs/2507.04370. Gao et al. [2026] Zelin Gao, Qiuyu Wang, Yanhong Zeng, Jiapeng Zhu, Ka Leong Cheng, Yixuan Li, Hanlin Wang, Yinghao Xu, et al. Advancing open-source world models. arXiv preprint arXiv:2601.20540, 2026. URL https://arxiv.org/abs/2601.20540. Google DeepMind [2025] Google DeepMind. Genie 3: A new frontier for world models. DeepMind Technical Blog, 2025. URL https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/. Gu et al. [2024] Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, et al. A survey on LLM-as-a-judge. arXiv preprint arXiv:2411.15594, 2024. URL https://arxiv.org/abs/2411.15594. Gu et al. [2025] Yu Gu, Kai Zhang, Yuting Ning, Boyuan Zheng, Boyu Gou, Tianci Xue, Cheng Chang, Sanjari Srivastava, Yanan Xie, Peng Qi, Huan Sun, and Yu Su. Is your LLM secretly a world model of the internet? Model-based planning for web agents. Trans. Mach. Learn. Res., 2025. URL https://arxiv.org/abs/2411.06559. Guan et al. [2026] Yiming Guan, Rui Yu, John Zhang, Lu Wang, Chaoyun Zhang, Liqun Li, Bo Qiao, Si Qin, et al. Computer-using world model. arXiv preprint arXiv:2602.17365, 2026. URL https://arxiv.org/abs/2602.17365. Guo et al. [2025a] Junliang Guo, Yang Ye, Tianyu He, Haoyu Wu, Yushu Jiang, Tim Pearce, and Jiang Bian. MineWorld: a real-time and open-source interactive world model on minecraft. arXiv preprint arXiv:2504.08388, 2025a. URL https://arxiv.org/abs/2504.08388. Guo et al. [2025b] Shangmin Guo, Omar Darwiche Domingues, Raphaël Avalos, Aaron Courville, and Florian Strub. World modelling improves language model agents. arXiv preprint arXiv:2506.02918, 2025b. URL https://arxiv.org/abs/2506.02918. Guo et al. [2026] Zhixiang Guo, Siyuan Liang, András Balogh, Noah Lunberry, Rong-Cheng Tu, Márk Jelasity, and Dacheng Tao. When world models dream wrong: Physical-conditioned adversarial attacks against world models. arXiv preprint arXiv:2602.18739, 2026. URL https://arxiv.org/abs/2602.18739. Ha and Schmidhuber [2018] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In Adv. Neural Inf. Process. Syst., volume 31, 2018. URL https://arxiv.org/abs/1803.10122. Hafner et al. [2019] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In Int. Conf. Mach. Learn., pages 2555–2565, 2019. URL https://arxiv.org/abs/1811.04551. Hafner et al. [2020] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. In Int. Conf. Learn. Represent., 2020. URL https://arxiv.org/abs/1912.01603. Hafner et al. [2025] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse control tasks through world models. Nature, 640(8059):647–653, 2025. URL https://arxiv.org/abs/2301.04104. Hansen et al. [2024] Nicklas Hansen, Hao Su, and Xiaolong Wang. TD-MPC2: Scalable, robust world models for continuous control. In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2310.16828. Hao et al. [2023] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. In Proc. Conf. Empir. Methods Nat. Lang. Process., pages 8154–8173, 2023. URL https://arxiv.org/abs/2305.14992. Ho et al. [2022] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Adv. Neural Inf. Process. Syst., volume 35, 2022. URL https://arxiv.org/abs/2204.03458. Höllein et al. [2023] Lukas Höllein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3D meshes from 2D text-to-image models. In IEEE/CVF Int. Conf. Comput. Vis., 2023. URL https://arxiv.org/abs/2303.11989. Hong et al. [2024] Jiwoo Hong, Noah Lee, and James Thorne. ORPO: Monolithic preference optimization without reference model. In Proc. Conf. Empir. Methods Nat. Lang. Process., 2024. URL https://arxiv.org/abs/2403.07691. Höppe et al. [2022] Tobias Höppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. Trans. Mach. Learn. Res., 2022. URL https://arxiv.org/abs/2206.07696. Hu et al. [2023] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. GAIA-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. URL https://arxiv.org/abs/2309.17080. Hu et al. [2026a] Tao Hu, Jiaxin Ai, Licheng Wen, Xueheng Li, Shu Zou, Siqi Li, Nianchen Deng, Xinyu Cai, Hongbin Zhou, Pinlong Cai, et al. IterCAD: An iterative multimodal agent for visually-grounded CAD generation and editing. arXiv preprint arXiv:2606.13368, 2026a. URL https://arxiv.org/abs/2606.13368. Hu et al. [2026b] Tianshuai Hu, Zeying Gong, Lingdong Kong, Xiaodong Mei, Yiyi Ding, Qi Zeng, Ao Liang, Rong Li, Yangyi Zhong, and Junwei Liang. NavThinker: Action-conditioned world models for coupled prediction and planning in social navigation. arXiv preprint arXiv:2603.15359, 2026b. URL https://arxiv.org/abs/2603.15359. Hu et al. [2026c] Xiaomeng Hu, Yinger Zhang, Fei Huang, Jianhong Tu, Yang Su, Lianghao Deng, Yuxuan Liu, Yantao Liu, Dayiheng Liu, and Tsung-Yi Ho. OccuBench: Evaluating AI agents on real-world professional tasks via language world models. arXiv preprint arXiv:2604.10866, 2026c. URL https://arxiv.org/abs/2604.10866. Huang et al. [2025a] Suning Huang, Qianzhong Chen, Xiaohan Zhang, Jiankai Sun, and Mac Schwager. ParticleFormer: A 3D point cloud world model for multi-object, multi-material robotic manipulation. arXiv preprint arXiv:2506.23126, 2025a. URL https://arxiv.org/abs/2506.23126. Huang et al. [2022] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, et al. Inner monologue: Embodied reasoning through planning with language models. In Conf. Robot Learn., 2022. URL https://arxiv.org/abs/2207.05608. Huang et al. [2023] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. VoxPoser: Composable 3D value maps for robotic manipulation with language models. In Conf. Robot Learn., 2023. URL https://arxiv.org/abs/2307.05973. Huang et al. [2026] Wenlong Huang, Yu-Wei Chao, Arsalan Mousavian, Ming-Yu Liu, Dieter Fox, Kaichun Mo, and Li Fei-Fei. PointWorld: Scaling 3D world models for in-the-wild robotic manipulation. arXiv preprint arXiv:2601.03782, 2026. URL https://arxiv.org/abs/2601.03782. Huang et al. [2025b] Xu Huang, Junwu Chen, Yuxing Fei, Zhuohan Li, Philippe Schwaller, and Gerbrand Ceder. CASCADE: Cumulative agentic skill creation through autonomous development and evolution. arXiv preprint arXiv:2512.23880, 2025b. URL https://arxiv.org/abs/2512.23880. Janner et al. [2019] Michael Janner, Justin Fu, Marvin Zhang, and Sergey Levine. When to trust your model: Model-based policy optimization. In Adv. Neural Inf. Process. Syst., volume 32, 2019. URL https://arxiv.org/abs/1906.08253. Jimenez et al. [2024] Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can language models resolve real-world GitHub issues? In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2310.06770. Jing et al. [2026] Bowen Jing, Ruiyang Hao, Weitao Zhou, and Haibao Yu. CounterScene: Counterfactual causal reasoning in generative world models for safety-critical closed-loop evaluation. arXiv preprint arXiv:2603.21104, 2026. URL https://arxiv.org/abs/2603.21104. Kaiser et al. [2020] Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Milos, Blazej Osinski, Roy H. Campbell, K. Czechowski, D. Erhan, Chelsea Finn, et al. Model-based reinforcement learning for Atari. In Int. Conf. Learn. Represent., 2020. URL https://arxiv.org/abs/1903.00374. Kang et al. [2025] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. In Int. Conf. Mach. Learn., 2025. URL https://arxiv.org/abs/2411.02385. Kerbl et al. [2023] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. URL https://arxiv.org/abs/2308.04079. Kim et al. [2024] Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, and Minjoon Seo. Prometheus: Inducing fine-grained evaluation capability in language models. In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2310.08491. Koh et al. [2021] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. PathDreamer: A world model for indoor navigation. In IEEE/CVF Int. Conf. Comput. Vis., pages 14738–14748, 2021. URL https://arxiv.org/abs/2105.08756. Koh et al. [2026] Woosung Koh, Sungjun Han, Segyu Lee, Se-Young Yun, and Jamin Shin. Generative visual code mobile world models. arXiv preprint arXiv:2602.01576, 2026. URL https://arxiv.org/abs/2602.01576. Kojima et al. [2022] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In Adv. Neural Inf. Process. Syst., 2022. URL https://arxiv.org/abs/2205.11916. Kong et al. [2025] Lingdong Kong, Wesley Yang, Jianbiao Mei, Youquan Liu, Ao Liang, Dekai Zhu, Dongyue Lu, Wei Yin, et al. 3D and 4D world modeling: A survey. arXiv preprint arXiv:2509.07996, 2025. URL https://arxiv.org/abs/2509.07996. Kong et al. [2026] Lingdong Kong, Ao Liang, Tianyi Yan, Hongsi Liu, Wesley Yang, Ziqi Huang, Xian Sun, Wei Yin, et al. Is your driving world model an all-around player? In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 36385–36399, 2026. URL https://arxiv.org/abs/2605.10858. LeCun [2022] Yann LeCun. A path towards autonomous machine intelligence. OpenReview preprint, 2022. URL https://openreview.net/pdf?id=BZ5a1r-kVsf. Leroy et al. [2024] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3D with MASt3R. In Eur. Conf. Comput. Vis., 2024. URL https://arxiv.org/abs/2406.09756. Li et al. [2024a] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martín-Martín, Chen Wang, Gabrael Levine, et al. BEHAVIOR-1K: A human-centered, embodied AI benchmark with 1,000 everyday activities and realistic simulation. arXiv preprint arXiv:2403.09227, 2024a. URL https://arxiv.org/abs/2403.09227. Li et al. [2025a] Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, et al. WorldModelBench: Judging video generation models as world models. arXiv preprint arXiv:2502.20694, 2025a. URL https://arxiv.org/abs/2502.20694. Li et al. [2024b] Haoran Li, Haolin Shi, Wenli Zhang, Wenjun Wu, Yong Liao, Lin Wang, Lik-Hang Lee, and Pengyuan Zhou. DreamScene: 3D gaussian-based text-to-3D scene generation via formation pattern sampling. In Eur. Conf. Comput. Vis., 2024b. URL https://arxiv.org/abs/2404.03575. Li et al. [2025b] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, Kazuki Kozuka, and Aditya Grover. MobileWorldBench: Towards semantic world modeling for mobile agents. arXiv preprint arXiv:2512.14014, 2025b. URL https://arxiv.org/abs/2512.14014. Li et al. [2026] Xiao Li, Xiang Zheng, Yifeng Gao, Xinyu Xia, Yixu Wang, Xin Wang, et al. Safety in embodied AI: A survey of risks, attacks, and defenses. arXiv preprint arXiv:2605.02900, 2026. URL https://arxiv.org/abs/2605.02900. Li et al. [2025c] Xinqing Li, Xin He, Le Zhang, Min Wu, Xiaoli Li, and Yun Liu. A comprehensive survey on world models for embodied AI. arXiv preprint arXiv:2510.16732, 2025c. URL https://arxiv.org/abs/2510.16732. Li et al. [2025d] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavely. MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2025d. URL https://arxiv.org/abs/2412.04463. Liang et al. [2026] Ao Liang, Youquan Liu, Yu Yang, Dongyue Lu, Linfeng Li, Lingdong Kong, Huaici Zhao, and Wei Tsang Ooi. LiDARCrafter: Dynamic 4D world modeling from LiDAR sequences. In AAAI Conf. Artif. Intell., volume 40, pages 18406–18414, 2026. URL https://arxiv.org/abs/2508.03692. Liang et al. [2023] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In IEEE Int. Conf. Robot. Autom., 2023. URL https://arxiv.org/abs/2209.07753. Lightman et al. [2024] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2305.20050. Lin et al. [2025] Kevin Qinghong Lin, Siyuan Hu, Linjie Li, Zhengyuan Yang, Lijuan Wang, Philip Torr, and Mike Zheng Shou. Computer-use agents as judges for generative user interface. arXiv preprint arXiv:2511.15567, 2025. URL https://arxiv.org/abs/2511.15567. Liu et al. [2026] Hanqing Liu, Songping Wang, Jiahuan Long, Jiacheng Hou, Jialiang Sun, Chao Li, Yang Yang, Wei Peng, Xu Liu, Tingsong Jiang, Wen Yao, and Yao Mu. JailWAM: Jailbreaking world action models in robot control. arXiv preprint arXiv:2604.05498, 2026. URL https://arxiv.org/abs/2604.05498. Liu et al. [2025] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise RingAttention. In Int. Conf. Learn. Represent., 2025. URL https://arxiv.org/abs/2402.08268. Liu et al. [2024] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, et al. AgentBench: Evaluating LLMs as agents. In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2308.03688. Liu et al. [2023] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-Eval: NLG evaluation using GPT-4 with better human alignment. In Proc. Conf. Empir. Methods Nat. Lang. Process., 2023. URL https://arxiv.org/abs/2303.16634. Lu et al. [2025a] Guanxing Lu, Baoxiong Jia, Puhao Li, Yixin Chen, Ziwei Wang, Yansong Tang, and Siyuan Huang. GWM: Towards scalable gaussian world models for robotic manipulation. In IEEE/CVF Int. Conf. Comput. Vis., pages 9263–9274, 2025a. URL https://arxiv.org/abs/2508.17600. Lu et al. [2025b] Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Felix Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, Zirui Wang, and Ruoming Pang. ToolSandbox: A stateful, conversational, interactive evaluation benchmark for LLM tool use capabilities. In Proc. Conf. N. Am. Chapter Assoc. Comput. Linguist., 2025b. URL https://arxiv.org/abs/2408.04682. Lu et al. [2024] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D gaussians for view-adaptive rendering. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2024. URL https://arxiv.org/abs/2312.00109. Luo et al. [2025] Dezhao Luo, Bohan Tang, Kang Li, Georgios Papoudakis, Jifei Song, Shaogang Gong, Jianye Hao, Jun Wang, and Kun Shao. ViMo: A generative visual GUI world model for app agents. arXiv preprint arXiv:2504.13936, 2025. URL https://arxiv.org/abs/2504.13936. Madaan et al. [2023] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, et al. Self-Refine: Iterative refinement with self-feedback. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2303.17651. Maimon et al. [2026] Gallil Maimon, Ori Yoran, Felix Kreuk, Michael Hassid, Gal Cohen, Pierre Chambon, and Yossi Adi. Self-execution simulation improves coding models. arXiv preprint arXiv:2604.03253, 2026. URL https://arxiv.org/abs/2604.03253. Makoviychuk et al. [2021] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac Gym: High performance GPU-based physics simulation for robot learning. In Adv. Neural Inf. Process. Syst., 2021. URL https://arxiv.org/abs/2108.10470. McAleese et al. [2024] Nat McAleese, Rai Michael Pokorny, Juan Felipe Ceron Uribe, Evgenia Nitishinskaya, Maja Trebacz, and Jan Leike. LLM critics help catch LLM bugs. arXiv preprint arXiv:2407.00215, 2024. URL https://arxiv.org/abs/2407.00215. Mei et al. [2024] Jianbiao Mei, Tao Hu, Xuemeng Yang, Licheng Wen, Yu Yang, Tiantian Wei, Yukai Ma, Min Dou, Botian Shi, and Yong Liu. Dreamforge: Motion-aware autoregressive video generation for multi-view driving scenes. arXiv preprint arXiv:2409.04003, 2024. URL https://arxiv.org/abs/2409.04003. Mei et al. [2025] Jianbiao Mei, Yu Yang, Xuemeng Yang, Licheng Wen, Jiajun Lv, Botian Shi, and Yong Liu. Vision-centric 4D occupancy forecasting and planning via implicit residual world models. arXiv preprint arXiv:2510.16729, 2025. URL https://arxiv.org/abs/2510.16729. Meng et al. [2024] Yu Meng, Mengzhou Xia, and Danqi Chen. SimPO: Simple preference optimization with a reference-free reward. In Adv. Neural Inf. Process. Syst., 2024. URL https://arxiv.org/abs/2405.14734. Micheli et al. [2023] Vincent Micheli, Eloi Alonso, and François Fleuret. Transformers are sample-efficient world models. In Int. Conf. Learn. Represent., 2023. URL https://arxiv.org/abs/2209.00588. Mildenhall et al. [2020] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. URL https://arxiv.org/abs/2003.08934. Min et al. [2024] Chen Min, Dawei Zhao, Liang Xiao, Jian Zhao, Xinli Xu, Zheng Zhu, Lei Jin, Jianshu Li, et al. DriveWorld: 4D pre-trained scene understanding via world models for autonomous driving. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 15522–15533, 2024. URL https://arxiv.org/abs/2405.04390. Moerland et al. [2023] Thomas M. Moerland, Joost Broekens, Aske Plaat, and Catholijn M. Jonker. Model-based reinforcement learning: A survey. Found. Trends Mach. Learn., 16(1), 2023. URL https://arxiv.org/abs/2006.16712. Müller et al. [2022] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), 2022. URL https://arxiv.org/abs/2201.05989. Nasiriany et al. [2024] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. RoboCasa: Large-scale simulation of everyday tasks for generalist robots. In Robot. Sci. Syst., 2024. URL https://arxiv.org/abs/2406.02523. Ouyang et al. [2022] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, et al. Training language models to follow instructions with human feedback. In Adv. Neural Inf. Process. Syst., 2022. URL https://arxiv.org/abs/2203.02155. Packer et al. [2023] Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems. arXiv preprint arXiv:2310.08560, 2023. URL https://arxiv.org/abs/2310.08560. Park et al. [2023] Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Annu. ACM Symp. User Interface Softw. Technol., pages 1–22, 2023. URL https://arxiv.org/abs/2304.03442. Qian et al. [2026] Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, and Heng Ji. Current agents fail to leverage world model as tool for foresight. arXiv preprint arXiv:2601.03905, 2026. URL https://arxiv.org/abs/2601.03905. Qiao et al. [2024] Shuofei Qiao, Runnan Fang, Ningyu Zhang, Yuqi Zhu, Xiang Chen, Shumin Deng, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. Agent planning with world knowledge model. In Adv. Neural Inf. Process. Syst., volume 37, 2024. URL https://arxiv.org/abs/2405.14205. Qin et al. [2024] Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, et al. WorldSimBench: Towards video generation models as world simulators. arXiv preprint arXiv:2410.18072, 2024. URL https://arxiv.org/abs/2410.18072. Rafailov et al. [2023] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2305.18290. Rahmani [2026] Babak Rahmani. Debugging code world models. arXiv preprint arXiv:2602.07672, 2026. URL https://arxiv.org/abs/2602.07672. Ren et al. [2026] Baochang Ren, Yunzhi Yao, Rui Sun, Shuofei Qiao, Ningyu Zhang, and Huajun Chen. Aligning agentic world models via knowledgeable experience learning. arXiv preprint arXiv:2601.13247, 2026. URL https://arxiv.org/abs/2601.13247. Ren et al. [2025] Zhenzhen Ren, Xinpeng Zhang, Zhenxing Qian, Yan Gao, Yu Shi, Shuxin Zheng, and Jiyan He. GTM: Simulating the world of tools for AI agents. arXiv preprint arXiv:2512.04535, 2025. URL https://arxiv.org/abs/2512.04535. Rivard et al. [2026] Luke Rivard, Sun Sun, Hongyu Guo, Wenhu Chen, and Yuntian Deng. NeuralOS: Towards simulating operating systems via neural generative models. In Int. Conf. Learn. Represent., 2026. URL https://arxiv.org/abs/2507.08800. Russell et al. [2025] Lloyd Russell, Anthony Hu, Lorenzo Bertoni, George Fedoseev, Jamie Shotton, Elahe Arani, and Gianluca Corrado. GAIA-2: A controllable multi-view generative world model for autonomous driving. arXiv preprint arXiv:2503.20523, 2025. URL https://arxiv.org/abs/2503.20523. Savva et al. [2019] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied AI research. In IEEE/CVF Int. Conf. Comput. Vis., pages 9339–9347, 2019. URL https://arxiv.org/abs/1904.01201. Schick et al. [2023] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2302.04761. Schrittwieser et al. [2020] Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy Lillicrap, and David Silver. Mastering Atari, Go, chess and shogi by planning with a learned model. Nature, 588(7839), 2020. URL https://arxiv.org/abs/1911.08265. Schulman et al. [2017] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. URL https://arxiv.org/abs/1707.06347. Shao et al. [2024] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. URL https://arxiv.org/abs/2402.03300. Shen et al. [2026] Zhouzhou Shen, Xueyu Hu, Xiyun Li, Tianqing Fang, Juncheng Li, and Shengyu Zhang. World-model-augmented web agents with action correction. arXiv preprint arXiv:2602.15384, 2026. URL https://arxiv.org/abs/2602.15384. Shinn et al. [2023] Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2303.11366. Shridhar et al. [2021] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CLIPort: What and where pathways for robotic manipulation. In Conf. Robot Learn., 2021. URL https://arxiv.org/abs/2109.12098. Shridhar et al. [2022] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conf. Robot Learn., 2022. URL https://arxiv.org/abs/2209.05451. Stiennon et al. [2020] Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback. In Adv. Neural Inf. Process. Syst., 2020. URL https://arxiv.org/abs/2009.01325. Sumers et al. [2024] Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive architectures for language agents. Trans. Mach. Learn. Res., 2024. URL https://arxiv.org/abs/2309.02427. Sun et al. [2024] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. DimensionX: Create any 3D and 4D scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. URL https://arxiv.org/abs/2411.04928. Sutton [1991] Richard S. Sutton. Dyna, an integrated architecture for learning, planning, and reacting. ACM SIGART Bull., 2(4):160–163, 1991. URL https://dl.acm.org/doi/10.1145/122344.122377. Tancik et al. [2023] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, et al. Nerfstudio: A modular framework for neural radiance field development. In Proc. ACM SIGGRAPH Asia Conf., 2023. URL https://arxiv.org/abs/2302.04264. Tang et al. [2024] Hao Tang, Darren Key, and Kevin Ellis. WorldCoder, a model-based LLM agent: Building world models by writing code and interacting with the environment. In Adv. Neural Inf. Process. Syst., volume 37, 2024. URL https://arxiv.org/abs/2402.12275. Tao et al. [2024] Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, et al. ManiSkill3: GPU parallelized robotics simulation and rendering for generalizable embodied AI. arXiv preprint arXiv:2410.00425, 2024. URL https://arxiv.org/abs/2410.00425. Todorov et al. [2012] Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo: A physics engine for model-based control. In IEEE/RSJ Int. Conf. Intell. Robots Syst., pages 5026–5033, 2012. URL https://ieeexplore.ieee.org/document/6386109. Tu et al. [2025] Sifan Tu, Xin Zhou, Dingkang Liang, Xingyu Jiang, Yumeng Zhang, Xiaofan Li, and Xiang Bai. The role of world models in shaping autonomous driving: A comprehensive survey. arXiv preprint arXiv:2502.10498, 2025. URL https://arxiv.org/abs/2502.10498. Uesato et al. [2022] Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. Solving math word problems with process- and outcome-based feedback. arXiv preprint arXiv:2211.14275, 2022. URL https://arxiv.org/abs/2211.14275. Vafa et al. [2024] Keyon Vafa, Justin Y. Chen, Ashesh Rambachan, Jon Kleinberg, and Sendhil Mullainathan. Evaluating the world model implicit in a generative model. In Adv. Neural Inf. Process. Syst., volume 37, 2024. URL https://arxiv.org/abs/2406.03689. Valevski et al. [2025] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. In Int. Conf. Learn. Represent., 2025. URL https://arxiv.org/abs/2408.14837. van de Ven et al. [2025] Gido M. van de Ven, Nicholas Soures, and Dhireesha Kudithipudi. Continual learning and catastrophic forgetting. In Learning and Memory: A Comprehensive Reference. Academic Press, 2025. URL https://arxiv.org/abs/2403.05175. Voleti et al. [2022] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In Adv. Neural Inf. Process. Syst., 2022. URL https://arxiv.org/abs/2205.09853. Wang et al. [2024a] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Trans. Mach. Learn. Res., 2024a. URL https://arxiv.org/abs/2305.16291. Wang and Agapito [2025] Hengyi Wang and Lourdes Agapito. 3D reconstruction with spatial memory. In Int. Conf. 3D Vis., 2025. URL https://arxiv.org/abs/2408.16061. Wang et al. [2025a] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. VGGT: Visual geometry grounded transformer. arXiv preprint arXiv:2503.11651, 2025a. URL https://arxiv.org/abs/2503.11651. Wang et al. [2026a] Jiaxu Wang, Yicheng Jiang, Tianlun He, Jingkai Sun, Qiang Zhang, Junhao He, Jiahang Cao, Zesen Gan, Mingyuan Sun, Qiming Shao, and Xiangyu Yue. MVISTA-4D: View-consistent 4D world model with test-time action inference for robotic manipulation. arXiv preprint arXiv:2602.09878, 2026a. URL https://arxiv.org/abs/2602.09878. Wang et al. [2025b] Kangrui Wang, Pingyue Zhang, Zihan Wang, Yaning Gao, Linjie Li, Qineng Wang, Hanyang Chen, Chi Wan, Yiping Lu, Zhengyuan Yang, et al. VAGEN: Reinforcing world model reasoning for multi-turn VLM agents. In Adv. Neural Inf. Process. Syst., 2025b. URL https://arxiv.org/abs/2510.16907. Wang et al. [2024b] Peiyi Wang, Lei Li, Zhihong Shao, R. X. Xu, Damai Dai, Yifei Li, Deli Chen, Y. Wu, and Zhifang Sui. Math-Shepherd: Verify and reinforce LLMs step-by-step without human annotations. In Annu. Meet. Assoc. Comput. Linguist., 2024b. URL https://arxiv.org/abs/2312.08935. Wang et al. [2025c] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3D perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025c. URL https://arxiv.org/abs/2501.12387. Wang et al. [2025d] Qineng Wang, Baiqiao Yin, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, et al. MindCube: Spatial mental modeling from limited views. arXiv preprint arXiv:2506.21458, 2025d. URL https://arxiv.org/abs/2506.21458. Wang et al. [2024c] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2024c. URL https://arxiv.org/abs/2312.14132. Wang et al. [2024d] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving. In Eur. Conf. Comput. Vis., 2024d. URL https://arxiv.org/abs/2309.09777. Wang et al. [2026b] Zhaoyang Wang, Canwen Xu, Boyi Liu, Yite Wang, Siwei Han, Zhewei Yao, Huaxiu Yao, and Yuxiong He. Agent world model: Infinity synthetic environments for agentic reinforcement learning. In Int. Conf. Mach. Learn., 2026b. URL https://arxiv.org/abs/2602.10090. Wang et al. [2025e] Zihan Wang, Kangrui Wang, Qineng Wang, Pingyue Zhang, Linjie Li, Zhengyuan Yang, Xing Jin, Kefan Yu, Minh Nhat Nguyen, Licheng Liu, et al. RAGEN: Understanding self-evolution in LLM agents via multi-turn reinforcement learning. arXiv preprint arXiv:2504.20073, 2025e. URL https://arxiv.org/abs/2504.20073. Wei et al. [2022] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Adv. Neural Inf. Process. Syst., volume 35, pages 24824–24837, 2022. URL https://arxiv.org/abs/2201.11903. World Labs Team [2025] World Labs Team. Marble: A multimodal world model. World Labs Technical Post, 2025. URL https://w.worldlabs.ai/blog/marble-world-model. Wu et al. [2024a] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D gaussian splatting for real-time dynamic scene rendering. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2024a. URL https://arxiv.org/abs/2310.08528. Wu et al. [2024b] Jialong Wu, Shaofeng Yin, Ningya Feng, Xu He, Dong Li, Jianye Hao, and Mingsheng Long. iVideoGPT: Interactive VideoGPTs are scalable world models. In Adv. Neural Inf. Process. Syst., volume 37, pages 68082–68119, 2024b. URL https://arxiv.org/abs/2405.15223. Wu et al. [2023] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Ken Goldberg, and Pieter Abbeel. DayDreamer: World models for physical robot learning. In Conf. Robot Learn., 2023. URL https://arxiv.org/abs/2206.14176. Wu et al. [2026] Rong Wu, Daocheng Fu, Licheng Wen, Xuemeng Yang, Shu Zou, Jianbiao Mei, Yuxin Wang, Hairong Zhang, Yu Yang, Tao Hu, et al. Memharness: Memory is reconstructed, not replayed. arXiv preprint arXiv:2607.28272, 2026. URL https://arxiv.org/abs/2607.28272. Xiao et al. [2026] Zikai Xiao, Jianhong Tu, Chuhang Zou, Yuxin Zuo, Zhi Li, Peng Wang, Bowen Yu, Fei Huang, Junyang Lin, and Zuozhu Liu. WebWorld: A large-scale world model for web agent training. arXiv preprint arXiv:2602.14721, 2026. URL https://arxiv.org/abs/2602.14721. Xie et al. [2024a] Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. CityDreamer: Compositional generative model of unbounded 3D cities. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2024a. URL https://arxiv.org/abs/2309.00610. Xie et al. [2024b] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, et al. OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments. In Adv. Neural Inf. Process. Syst., volume 37, 2024b. URL https://arxiv.org/abs/2404.07972. Xu et al. [2026a] Shuhan Xu, Siyuan Liang, Hongling Zheng, Yong Luo, Han Hu, Lefei Zhang, and Dacheng Tao. CtrlAttack: A unified attack on world-model control in diffusion models. arXiv preprint arXiv:2603.13435, 2026a. URL https://arxiv.org/abs/2603.13435. Xu et al. [2026b] Xiang Xu, Ao Liang, Youquan Liu, Linfeng Li, Lingdong Kong, Ziwei Liu, and Qingshan Liu. U4D: Uncertainty-aware 4D world modeling from LiDAR sequences. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10027–10039, 2026b. URL https://arxiv.org/abs/2512.02982. Yan et al. [2026a] Bo Yan, Weikai Lin, Yada Zhu, and Song Wang. SafeDream: Safety world model for proactive early jailbreak detection. arXiv preprint arXiv:2604.16824, 2026a. URL https://arxiv.org/abs/2604.16824. Yan et al. [2026b] Tianyi Yan, Tao Tang, Xingtai Gui, Yongkang Li, Jiasen Zheng, Weiyao Huang, Lingdong Kong, Wencheng Han, et al. AD-R1: Closed-loop reinforcement learning for end-to-end autonomous driving with impartial world models. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 1085–1095, 2026b. URL https://arxiv.org/abs/2511.20325. Yan et al. [2021] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video generation using VQ-VAE and transformers. arXiv preprint arXiv:2104.10157, 2021. URL https://arxiv.org/abs/2104.10157. Yang et al. [2025a] Jianing Yang, Alexander Sax, Kevin J. Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3R: Towards 3D reconstruction of 1000+ images in one forward pass. arXiv preprint arXiv:2501.13928, 2025a. URL https://arxiv.org/abs/2501.13928. Yang et al. [2024] Sherry Yang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2310.06114. Yang et al. [2025b] Yu Yang, Jianbiao Mei, Yukai Ma, Siliang Du, Wenqing Chen, Yijie Qian, Yuxiang Feng, and Yong Liu. Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving. In AAAI Conf. Artif. Intell., volume 39, pages 9327–9335, 2025b. URL https://arxiv.org/abs/2408.14197. Yang et al. [2026a] Yu Yang, Alan Liang, Jianbiao Mei, Yukai Ma, Yong Liu, and Gim Hee Lee. X-scene: Large-scale driving scene generation with high fidelity and flexible controllability. In Adv. Neural Inf. Process. Syst., volume 38, pages 104415–104451, 2026a. URL https://arxiv.org/abs/2506.13558. Yang et al. [2026b] Yu Yang, Yue Liao, Jianbiao Mei, Baisen Wang, Xuemeng Yang, Licheng Wen, Jiangning Zhang, Xiangtai Li, Liang Lv, Hanlin Chen, et al. SPIRAL: Self-evolving action-conditioned video generation via reflective planning agents. arXiv preprint arXiv:2603.08403, 2026b. URL https://arxiv.org/abs/2603.08403. Yao et al. [2023a] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Adv. Neural Inf. Process. Syst., volume 36, 2023a. URL https://arxiv.org/abs/2305.10601. Yao et al. [2023b] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. In Int. Conf. Learn. Represent., 2023b. URL https://arxiv.org/abs/2210.03629. Yu et al. [2022] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 5501–5510, 2022. URL https://arxiv.org/abs/2112.05131. Yu et al. [2025a] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T. Freeman, and Jiajun Wu. WonderWorld: Interactive 3D scene generation from a single image. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2025a. URL https://arxiv.org/abs/2406.09394. Yu et al. [2025b] Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. GameFactory: Creating new games with generative interactive videos. In IEEE/CVF Int. Conf. Comput. Vis., pages 11590–11599, 2025b. URL https://arxiv.org/abs/2501.08325. Yu et al. [2025c] Xinlei Yu, Chengming Xu, Guibin Zhang, Zhangquan Chen, Yudong Zhang, Yongbo He, Peng-Tao Jiang, Jiangning Zhang, Xiaobin Hu, and Shuicheng Yan. Vismem: Latent vision memory unlocks potential of vision-language models. arXiv preprint arXiv:2511.11007, 2025c. URL https://arxiv.org/abs/2511.11007. Yu et al. [2026] Xinlei Yu, Zhangquan Chen, Yongbo He, Tianyu Fu, Guanting Dong, Cheng Yang, Chengming Xu, Yue Ma, Xiaobin Hu, Zhe Cao, et al. The latent space: Foundation, evolution, mechanism, ability, and outlook. arXiv preprint arXiv:2604.02029, 2026. URL https://arxiv.org/abs/2604.02029. Yu et al. [2024] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D gaussian splatting. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 19447–19456, 2024. URL https://arxiv.org/abs/2311.16493. Yuan et al. [2023] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. RRHF: Rank responses to align language models with human feedback without tears. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2304.05302. Zeng et al. [2020] Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Ayzaan Wahid, Vikas Sindhwani, and Johnny Lee. Transporter networks: Rearranging the visual world for robotic manipulation. In Conf. Robot Learn., 2020. URL https://arxiv.org/abs/2010.14406. Zeng et al. [2024] Zifan Zeng, Chongzhe Zhang, Feng Liu, Joseph Sifakis, Qunli Zhang, Shiming Liu, and Peng Wang. World models: The safety perspective. arXiv preprint arXiv:2411.07690, 2024. URL https://arxiv.org/abs/2411.07690. Zha et al. [2025] Jirong Zha, Yuxuan Fan, Xiao Yang, Chen Gao, and Xinlei Chen. How to enable LLM with 3D capacity? a survey of spatial reasoning in LLM. In Int. Joint Conf. Artif. Intell., pages 10817–10825, 2025. URL https://arxiv.org/abs/2504.05786. Zhang et al. [2025a] Di Zhang, Junxian Li, Jingdi Lei, Xunzhi Wang, Yujie Liu, Zonglin Yang, Jiatong Li, Weida Wang, et al. Critic-V: VLM critics help catch VLM errors in multimodal reasoning. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 9050–9061, 2025a. URL https://arxiv.org/abs/2411.18203. Zhang et al. [2024a] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2NeRF: Text-driven 3D scene generation with neural radiance fields. IEEE Trans. Vis. Comput. Graph., 2024a. URL https://arxiv.org/abs/2305.11588. Zhang et al. [2025b] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. MonST3R: A simple approach for estimating geometry in the presence of motion. In Int. Conf. Learn. Represent., 2025b. URL https://arxiv.org/abs/2410.03825. Zhang et al. [2026a] Kai Zhang, Xiangchao Chen, Bo Liu, Tianci Xue, Zeyi Liao, Zhihan Liu, Xiyao Wang, Yuting Ning, et al. Agent learning via early experience. In Int. Conf. Mach. Learn., 2026a. URL https://arxiv.org/abs/2510.08558. Zhang et al. [2024b] Lunjun Zhang, Yuwen Xiong, Ze Yang, Sergio Casas, Rui Hu, and Raquel Urtasun. Copilot4D: Learning unsupervised world models for autonomous driving via discrete diffusion. In Int. Conf. Learn. Represent., 2024b. URL https://arxiv.org/abs/2311.01017. Zhang et al. [2023] Weipu Zhang, Gang Wang, Jian Sun, Yetian Yuan, and Gao Huang. STORM: Efficient stochastic transformer based world models for reinforcement learning. In Adv. Neural Inf. Process. Syst., volume 36, 2023. URL https://arxiv.org/abs/2310.09615. Zhang et al. [2026b] Xichen Zhang, Ziyi He, Yinghao Zhu, Sitong Wu, Shaozuo Yu, Meng Chu, Wenhu Zhang, Haoru Tan, and Jiaya Jia. SearchGym: Bootstrapping real-world search agents via cost-effective and high-fidelity environment simulation. arXiv preprint arXiv:2601.14615, 2026b. URL https://arxiv.org/abs/2601.14615. Zhang et al. [2025c] Yifan Zhang, Chunli Peng, Boyang Wang, Puyi Wang, Qingcheng Zhu, Fei Kang, Biao Jiang, Zedong Gao, Eric Li, Yang Liu, and Yahui Zhou. Matrix-Game: Interactive world foundation model. arXiv preprint arXiv:2506.18701, 2025c. URL https://arxiv.org/abs/2506.18701. Zhang et al. [2025d] Zhang Zhang, Qiang Zhang, Wei Cui, Shuai Shi, Yijie Guo, Gang Han, Wen Zhao, Jingkai Sun, et al. Occupancy world model for robots. arXiv preprint arXiv:2505.05512, 2025d. URL https://arxiv.org/abs/2505.05512. Zhen et al. [2025] Haoyu Zhen, Qiao Sun, Hongxin Zhang, Junyan Li, Siyuan Zhou, Yilun Du, and Chuang Gan. TesserAct: Learning 4D embodied world models. arXiv preprint arXiv:2504.20995, 2025. URL https://arxiv.org/abs/2504.20995. Zheng et al. [2023] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, et al. Judging LLM-as-a-judge with MT-Bench and chatbot arena. In Adv. Neural Inf. Process. Syst., 2023. URL https://arxiv.org/abs/2306.05685. Zheng et al. [2024] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. OccWorld: Learning a 3D occupancy world model for autonomous driving. In Eur. Conf. Comput. Vis., 2024. URL https://arxiv.org/abs/2311.16038. Zheng et al. [2026] Yuhao Zheng, Li’an Zhong, Yi Wang, Rui Dai, Kaikui Liu, Xiangxiang Chu, Linyuan Lv, Philip Torr, and Kevin Qinghong Lin. Code2World: A GUI world model via renderable code generation. arXiv preprint arXiv:2602.09856, 2026. URL https://arxiv.org/abs/2602.09856. Zhou et al. [2024] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A realistic web environment for building autonomous agents. In Int. Conf. Learn. Represent., 2024. URL https://arxiv.org/abs/2307.13854. Zhu et al. [2025] Haoyi Zhu, Yifan Wang, Jianjun Zhou, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Chunhua Shen, Jiangmiao Pang, and Tong He. Aether: Geometric-aware unified world modeling. In IEEE/CVF Int. Conf. Comput. Vis., pages 8535–8546, 2025. URL https://arxiv.org/abs/2503.18945. Zhu et al. [2024] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Bohan Li, Nianchen Deng, Min Dou, Yuqi Wang, et al. Is Sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. URL https://arxiv.org/abs/2405.03520. Ziegler et al. [2019] Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. URL https://arxiv.org/abs/1909.08593.