'World in World' Systems: Definition, Applications, and Key Components

• World in World systems provide internal, generative, spatial, or symbolic models that predict, simulate, or plan actions within environments. These systems construct robust models through observation-recreation of world details, structural consistency, and physical causality, including but not limited to video generators, spatial models, and life simulation worlds.
• Propelled by video generators, spatial models, and scene generators, World in World systems push the frontier of AI development. Systems like TeleWorld and HoloWorld facilitate improved decision-making, with TeleWorld integrating generated video sequences into dynamic 4D scenes for navigation. Additionally, AstraNav-World couples a video generator and a vision-language policy for future-oriented, actionable predictions.
• The evaluation of World in World systems extends beyond visual quality to include structure, behavior, and memory, requiring multi-representation approaches. Comprehensive evaluation considers factors like world persistence, causal coherence, and action-conditional responses, demonstrating predictive accuracy, interactional responsiveness and holistic consistency, even as systems face challenges such as errors, inference delays, and biases from external estimator dependencies.

“World in World” is an interpretive term for systems that construct an internal, executable, generative, spatial, or symbolic representation of an external environment and use that representation to predict, simulate, reason, plan, or act. The term does not denote a single formal architecture. Across contemporary research, it encompasses conditional driving-world generators, programmatic simulators, spatial extrapolation models, multi-view video systems, closed-loop embodied benchmarks, latent dynamics models, code-governed environments, persistent 3D reconstructions, and cross-scale urban generators. Its common subject is the construction of an internally maintained world whose usefulness depends not only on visual plausibility, but also on structural consistency, action controllability, memory, causal persistence, and interaction with an external environment.

1. Conceptual scope and definitions

The central distinction is between a model that reproduces observations and a model that maintains an internal account of the processes producing those observations. A video generator may synthesize realistic frames, but a stronger world model should represent or approximate entities, spatial relations, temporal evolution, physical regularities, uncertainty, possible futures, and action consequences. “World in World” therefore refers to an internal model that is sufficiently structured to support counterfactual prediction or interaction.

A classical formalization uses a latent state $s_t$, an action $a_t$, an observation $o_t$, and an optional reward $r_t$:

$$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$

This formulation appears explicitly in OpenWorldLib’s definition of advanced world models (Team et al., 6 Apr 2026). The latent state is treated as a compact internal representation incorporating memory for long-horizon dependencies. The action may be a motor command, camera movement, navigation command, language instruction, generation control, or manipulation signal.

The stronger architectural interpretation adds a closed interaction loop:

$$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$

This loop distinguishes world modeling from isolated world-knowledge injection. A system may use physical knowledge to improve image generation, 3D estimation, or spatial reasoning without maintaining a reusable, persistent, actionable model. The distinction is emphasized by the normative analysis in “Research on World Models Is Not Merely Injecting World Knowledge into Specific Tasks” (Zeng et al., 2 Feb 2026).

The internal world may be:

• Latent: a learned state or feature sequence used for future prediction.
• Explicit: a point cloud, mesh, 3D Gaussian Splatting representation, occupancy field, or simulator-compatible scene.
• Executable: source code implementing transition rules and persistent state.
• Visual: a generative model of future images or videos.
• Persistent: a memory of entities, relations, events, actions, and scene states.
• Compositional: a collection of specialized models or programs that jointly define the environment.

These forms are not mutually exclusive. TeleWorld, for example, uses generated video to update a dynamic 4D representation, which is rendered back into subsequent video generation (Chen et al., 31 Dec 2025). Code World Model separates executable state evolution from visual realization by using code-governed state and proxy-conditioned video synthesis (Chen et al., 26 Aug 2026). HY-World 2.0 converts panoramic hypotheses, generated views, and reconstructed geometry into navigable 3D Gaussian Splatting scenes and collision meshes (HY-World et al., 15 Apr 2026).

The term should not be interpreted as implying literal recursive universes. Most systems do not contain independently complete nested worlds. Their internal components are approximations, conditional predictors, local causal mechanisms, or structured memories. The phrase is most precise when used to describe an internal model that can be queried, updated, rolled forward, rendered, or acted within.

2. Representational forms of internal worlds

Latent dynamical states

World Machine defines an internal latent sequence $ws_t\in\mathbb{R}^{d_{ws}}$ that is learned to reconstruct observations and advance into future states. Its State Discovery procedure iteratively replaces stored latent states with shifted predictions, while sensory masking and sequence breaking train the model to continue with incomplete observations. In the Toy1D experiments, the latent dimensionality is $128$, and the model is tested on state-only continuation and shallow prediction from a compressed final state (Nascimento et al., 21 May 2026).

The model’s intended operation is:

$$\widehat{ws}_{t+1}=f_\theta(\widehat{ws}_{\leq t},s_{\leq t},m_{\leq t}),$$

followed by decoding:

$\widehat{y}_{t+1}=g_\phi(\widehat{ws}_{t+1}).$

The latent state is not directly supervised as a physical state. It is pressured to become useful through sensory reconstruction and future prediction. This supports an interpretation of $a_t$0 as an internal simulated state, although the experiments establish this capability only on the controlled Toy1D synthetic system. Prediction Shallow remains imperfect, and the paper does not establish causal identification of the latent coordinates.

Structured and executable worlds

PoE-World represents its internal environment as an exponentially weighted product of programmatic experts. Each expert is a Python program describing a local mechanism, such as ladder climbing, player motion, object contact, or collision. The product combines expert predictive distributions, while learned weights suppress irrelevant or contradictory programs (2505.10819).

The model is compositional rather than monolithic. One program need not explain the entire environment. Object attributes such as position, velocity, contact, visibility, and deletion status are represented separately, and hard constraints remove physically implausible states. New observations can trigger synthesis of additional experts, reweighting, pruning, or graph repair. The resulting world is executable, probabilistic, compositional, and updatable, but it remains dependent on object-centric perception, LLM-generated code, manually adapted Atari environments, and computationally expensive planning.

Code World Model gives executable representation a stronger causal role. It decomposes state into:

$a_t$1

where $a_t$2 contains persistent entities, relations, rules, events, schedules, and numerical variables, while $a_t$3 contains visual realization. The executable state is updated by an agent–code transition, and a video model generates observations conditioned on the resulting executable state through a proxy representation (Chen et al., 26 Aug 2026).

This separation addresses a limitation of purely visual world models: an event may disappear from the screen while its consequences should persist. In the proposed framework, code can update succession, faction relations, trade, patrols, goals, or information propagation after an event such as an assassination. The prototype, however, relies heavily on existing game-engine scenes and mechanisms; it does not demonstrate autonomous construction of a complete open-world simulator.

Explicit spatial and 4D worlds

HY-World 2.0 constructs navigable 3D worlds from text, single images, multi-view images, or video. Its pipeline includes panorama generation, trajectory planning, world expansion, world composition, and 3DGS optimization. WorldNav converts a panorama into semantic landmarks, a point cloud, a mesh, and a navigation mesh. WorldStereo 2.0 generates camera-controlled keyframes using global-geometric and spatial-stereo memory. WorldMirror 2.0 reconstructs point maps, depth, normals, camera parameters, and Gaussian attributes (HY-World et al., 15 Apr 2026).

The resulting representation is a 3D scene that can be rendered, navigated, lit, and used for collision detection. 3DGS serves as the primary visual representation, while a mesh extracted through TSDF fusion supplies collision and physical structure. The system therefore realizes an internal computational environment rather than only a sequence of images. Its main limitations are uncertainty in unseen regions, depth-scale ambiguity, dependence on generated views, heuristic trajectory planning, primarily static-scene assumptions, and the absence of a complete physical dynamics model.

TeleWorld extends this spatial approach temporally. It reconstructs generated video segments into a persistent dynamic 4D representation containing static scene structure and moving-object components. Macro-from-Micro Planning predicts sparse anchor frames, while key-frame reconstruction and dynamic saliency reduce the cost of long-horizon memory updates. Rendered 4D memory guides subsequent generation (Chen et al., 31 Dec 2025).

TeleWorld is therefore a closed generation–reconstruction–guidance loop:

$a_t$4

It maintains a visual and geometric world memory, but does not provide formal rigid-body dynamics, collision mechanics, object-interaction laws, or a complete action-conditioned simulator.

Relational and cross-scale worlds

HoloWorld models urban environments through a continuously updated cross-scale context. A city is represented as an $a_t$5 grid of blocks, each containing persistent building instances, while selected buildings receive associated interiors. The context hierarchy includes world-level, block-level, and building-level information (Huang et al., 6 Aug 2026).

The system preserves relations in both directions:

• city descriptions define functional organization, architectural style, and block plans;
• neighboring blocks condition subsequent exterior generation;
• exterior building instances provide identity, appearance, and footprints;
• footprints constrain indoor geometry;
• interior and exterior artifacts update the relevant context.

This produces a nested city–block–building–interior structure. Unlike a generic scene generator, HoloWorld attempts to ensure that an interior belongs to a specific exterior building in a specific block. Its reported Shape IoU is $a_t$6, while removing dynamic localization reduces Shape IoU to $a_t$7. The framework currently focuses on single-floor interiors and does not explicitly model multi-story connectivity or complete vertical urban structure.

3. World models as generative predictors and simulators

A major family of “World in World” systems uses generative models to imagine observations that have not yet occurred.

DriveDreamer constructs a conditional driving world model from real-world driving data rather than games or simulators. Its Auto-DM diffusion model generates images and videos conditioned on HD maps, 3D bounding boxes, object categories, text, reference images, and actions. ActionFormer predicts future structured traffic conditions recursively, and an action decoder predicts future ego actions (Wang et al., 2023).

The core rollout is:

$a_t$8

Maps constrain lane boundaries, lane dividers, and pedestrian crossings; boxes constrain object placement; text modulates weather and time of day; actions influence recurrent structural prediction. On nuScenes, DriveDreamer reports FID $a_t$9 and FVD $o_t$0 with first-stage training and ActionFormer. Its average open-loop trajectory error is $o_t$1 meters, with a collision rate of $o_t$2. These results demonstrate action-conditioned predictive correlation, but not closed-loop safety or complete causal simulation.

RemoteBAGEL applies the same broad principle spatially rather than temporally. Given a central remote-sensing tile and a direction, it generates an adjacent tile. The task is direction-conditioned spatial extrapolation:

$o_t$3

The central tile is a local observation of a larger geographic world, and the generated tile is a hypothesis about its surrounding context (Lu et al., 22 Sep 2025). RSWISE evaluates both visual fidelity and spatial reasoning, using FID and GPT-4o-based judgment of directional compliance, continuity, semantic transitions, and geographic coherence. RemoteBAGEL achieves an average RSWISE score of $o_t$4, compared with $o_t$5 for BAGEL. The task remains local and static: the model does not maintain a persistent global map, calibrated geographic state, or long-range spatial rollout.

IC-World addresses shared world modeling across camera views. It concatenates multiple input images into a grid and generates a single grid video so that all views share a denoising process and cross-panel computation. Reinforcement fine-tuning with GRPO uses geometry and motion rewards derived from point-cloud reconstruction and 3D point tracking (Wu et al., 1 Dec 2025).

The method improves registered point-cloud consistency and cross-view motion consistency, but does not explicitly construct a persistent 3D scene. Its consistency is statistically encouraged through joint generation and rewards based on imperfect external estimators such as Pi3, Lepard, and SpatialTrackerV2.

DreamWorld jointly predicts video latents, optical-flow-derived temporal features, DINOv2 semantic features, and VGGT spatial features. Consistent Constraint Annealing reduces heterogeneous world-feature weights during training, while Multi-Source Inner-Guidance combines text, temporal, semantic, and spatial predictions during inference (Tan et al., 28 Feb 2026).

DreamWorld reports an overall VBench score of $o_t$6, VideoPhy semantic adherence of $o_t$7, and Physical Commonsense of $o_t$8. Its world representations are proxies rather than explicit causal variables: optical flow does not encode physical laws, and DINOv2 and VGGT do not provide symbolic dynamics. The system improves world-consistent video but does not demonstrate persistent memory, action-conditioned intervention, or a general-purpose simulator.

4. Interaction, planning, and closed-loop utility

Open-loop visual quality is not sufficient to establish that a world model is useful for embodied agents. World-in-World introduces a platform that evaluates generative world models inside closed-loop environments using a common proposal–simulation–revision architecture (Zhang et al., 20 Oct 2025).

At each cycle, a proposal policy generates candidate action sequences. The world model simulates each candidate, a revision policy selects or refines a decision, and the selected action is executed in the real environment. A new observation then initiates the next cycle. The platform evaluates Active Recognition, Image-Goal Navigation, Active Embodied Question Answering, and robotic manipulation.

The reported findings are:

• visual quality alone does not guarantee task success;
• action controllability is more predictive of embodied utility than appearance metrics;
• action–observation post-training is more effective than merely scaling generic video pretraining;
• additional inference-time rollouts can improve closed-loop performance.

For SVD with action–observation post-training, increasing inference from three to eleven rollouts raises Active Recognition success from $o_t$9 to $r_t$0. In manipulation, improvements are modest for strong VLM policies and larger when the base policy is weak. These results show that world models are evaluated not only by the images they generate, but by whether those images support action selection under repeated environmental feedback.

AstraNav-World couples a diffusion-based video generator and a vision-language policy through synchronized rollouts and Multimodal Fusion Cross-Attention. Actions influence visual futures, and visual futures constrain executable actions. The system predicts future front-view frames and trajectories jointly, using a shared Qwen2.5-VL-3B context (Hu et al., 25 Dec 2025).

The architecture implements two directions of coupling:

$r_t$1

Sparse Foresight Scheduling activates expensive visual generation periodically while allowing policy-only intermediate control. On R2R-CE, the diffusion policy reaches a success rate of $r_t$2 and SPL of $r_t$3; on RxR-CE, it reaches a success rate of $r_t$4 and SPL of $r_t$5. Removing the video generator or disabling coupling degrades navigation reliability. The model is nevertheless an egocentric, task-conditioned predictor rather than a complete persistent 3D environment.

WPT uses a different strategy: the world model is used during training and then removed at deployment. A world-aware teacher evaluates candidate driving trajectories against predicted future occupancy or instance-level dynamics. A reward model supplies imitation and simulation rewards, including no collision, drivable-area compliance, time-to-collision, progress, and comfort. Query and reward distillation transfer the teacher’s future-aware planning into a lightweight student policy (Jiang et al., 25 Nov 2025).

On nuScenes, WPT-Teacher reports average displacement error $r_t$6 meters and collision rate $r_t$7; WPT-Student retains most of the gain with $r_t$8 meters and $r_t$9. On Bench2Drive, WPT-Teacher achieves driving score $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$0 and success rate $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$1. The student runs at $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$2 milliseconds, compared with $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$3 milliseconds for the teacher with online reward scoring, a reported approximately $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$4 speed advantage.

DWM decomposes latent transitions into an action-invariant world effect and a complementary action-driven effect. The world head is trained with a normalized contrastive objective under alternative current actions, while an orthogonality constraint encourages the residual to carry complementary information (Zhang et al., 21 Jul 2026).

The decomposition is:

$$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$5

with the predicted residual defined by:

$$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$6

DWM improves CEM planning success by a mean $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$7 percentage points across PushT-W, Reacher-W, and TwoRoom-W. Its purpose is not to create a separate simulator, but to make the internal transition model distinguish changes caused by the agent from changes that would occur under a null action, such as gravity, drift, inertia, or persistent oscillation. The decomposition is an inductive bias rather than a proof of unique causal identification.

WorldLens evaluates driving world models across Generation, Reconstruction, Action-Following, Downstream Task, and Human Preference (Liang et al., 11 Dec 2025). Its central result is that a model may perform well on open-loop appearance and still fail in closed-loop interaction. Reported route completion ranges from $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$8 to $$p(s_{t+1}\mid s_t,a_t),\qquad p(o_t\mid s_t),\qquad r_t\sim p(r_t\mid s_t,a_t).$$9, while Arena Driving Scores range from $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$0 to $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$1. This gap indicates that plausible short-term prediction does not establish a stable transition function under interventions.

WorldLens therefore evaluates whether generated videos can be reconstructed as 4D scenes, rendered from novel views, used by frozen perception systems, and acted upon by pretrained planners. DiST-4D achieves FVD $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$2, novel-view quality $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$3, and novel-view discrepancy $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$4, while OpenDWM leads several appearance-oriented metrics. The multidimensional profile rejects a single scalar notion of world quality.

5. Memory, causality, and consistency

Persistent memory distinguishes many “World in World” systems from next-frame predictors. Memory may be implicit in a latent state, explicit in a 3D scene, symbolic in executable code, or multimodal in a reusable context.

TeleWorld maintains a dynamic 4D memory reconstructed from generated video. Sparse keyframes are selected through Macro-from-Micro Planning, and static and dynamic scene components are updated separately. The memory is rendered under later camera poses and supplied as guidance to the generator (Chen et al., 31 Dec 2025). This mechanism helps preserve scene layout, object identity, camera-relative geometry, and temporal continuity, but it can also reinforce errors because the memory is inferred from generated content.

HoloWorld maintains a cross-scale context:

$$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$5

The context is updated only after task-specific validation. This lets city-level semantics, block-level layout, building identities, footprints, and interior programs remain connected across generation stages (Huang et al., 6 Aug 2026). In this setting, memory is not only temporal; it is hierarchical and relational.

World Machine uses a compact latent state to discard historical observations while retaining information needed for prediction. OpenWorldLib generalizes this requirement into an abstract Memory interface supporting recording, retrieval, compression, and management (Team et al., 6 Apr 2026). The interface is conceptually important but leaves retrieval scoring, compression objectives, capacity limits, and forgetting criteria unspecified.

Code World Model places persistence in executable state. The visual generator need not remember every off-screen consequence because the code world maintains entities, rules, event history, and relations (Chen et al., 26 Aug 2026). This provides a clearer causal substrate than visual memory, but the prototype depends on inherited game-engine logic and does not demonstrate autonomous long-horizon world construction.

Causal persistence is also central to World-Time Compute with Verified Code World Models. OpenWorld worlds are symbolic environments with deterministic transition programs, objectives, and invariants. Verified code generates exact trajectories, allowing training across independently instantiated worlds and evaluation on held-out worlds (Schwoebel et al., 7 Jul 2026).

In the paper’s diagnosis family, world-time fine-tuning yields an approximately $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$6 percentage-point gain at $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$7 billion parameters, with smaller gains for larger models. Verified code completes $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$8 twenty-step rollouts exactly and achieves $$\text{environment} \rightarrow \text{perception} \rightarrow \text{internal state} \rightarrow \text{reasoning and prediction} \rightarrow \text{action} \rightarrow \text{environment}.$$9 on ten-times out-of-distribution probes, whereas a per-step LLM predictor completes $ws_t\in\mathbb{R}^{d_{ws}}$0 exactly and achieves $ws_t\in\mathbb{R}^{d_{ws}}$1 on the probe suite. These results concern symbolic state and executable dynamics, not pixel-native environments. They support the claim that verified simulation can provide exact labels and cross-world generalization when the task family shares a learnable skill.

The distinction between visual, latent, explicit, and executable memory is consequential:

Memory form	Primary content	Main strength	Main limitation
Latent state	Compressed temporal dynamics	Efficient continuation	Weak interpretability
3D/4D representation	Geometry, appearance, motion	Rendering and spatial interaction	Often incomplete physics
Programmatic state	Rules, entities, events	Persistence and inspectability	Code synthesis and verification costs
Cross-scale context	City, block, building relations	Compositional coherence	Validation and representation complexity
External memory API	Stored observations and outputs	Modular reuse	Algorithmically under-specified

No single form is universally sufficient. Latent states are compact, explicit geometry supports spatial reasoning, executable code supports causal persistence, and contextual memory supports cross-task composition.

6. Evaluation, limitations, and future directions

The evaluation of world models has shifted from visual quality toward structural and behavioral validity. WorldLens argues that a generated world should be visually convincing, spatially and temporally coherent, reconstructable as a 3D or 4D scene, physically plausible, responsive to actions, useful for downstream perception, and aligned with human judgment (Liang et al., 11 Dec 2025).

This multidimensional perspective exposes recurring failure modes:

• Visual realism without geometry: sharp textures coexist with depth jumps, floaters, broken occlusion, or inconsistent novel views.
• Geometry without behavior: stable reconstructions fail to respond correctly to interventions.
• Short-term plausibility without persistence: objects or consequences disappear after leaving the field of view.
• Action-conditioned prediction without causal attribution: a model predicts what happens but not which changes are induced by the action.
• Symbolic correctness without visual realization: executable state is correct while the generated video violates it.
• Local continuity without global coherence: adjacent tiles or blocks look plausible individually but fail to form a consistent larger world.
• Open-loop success without closed-loop reliability: a planner follows generated observations but fails once its own actions alter the future.

Several papers provide quantitative evidence for these limitations. DriveDreamer’s planning results are open-loop; World-in-World shows that visual quality and task success are weakly aligned; WorldLens reports very low closed-loop route completion despite respectable open-loop PDMS; TeleWorld acknowledges that generated memory can reinforce hallucinated geometry; and DreamWorld improves world-consistent video without demonstrating persistent interaction or action-conditioned simulation.

Evaluation also remains dependent on external estimators. IC-World uses point-cloud reconstruction and point tracking; WorldLens uses frozen planners, perception models, and human judgments; RemoteBAGEL uses GPT-4o as a semantic judge; HoloWorld uses GPT-5.5 and human experts; WorldLens-Agent is trained from WorldLens-26K human rationales (Liang et al., 11 Dec 2025). These systems extend evaluation beyond FID or FVD, but each introduces estimator bias, prompt sensitivity, perception error, or metric incompleteness.

Future world models are likely to require the following capabilities:

1. Persistent state: object identity, event consequences, spatial memory, and revisitation consistency.
2. Action-conditioned prediction: futures that respond faithfully to the model’s actual action interface.
3. Explicit causal structure: mechanisms that distinguish autonomous evolution from agent-induced change.
4. Multi-representation integration: latent dynamics, geometry, executable state, semantic memory, and visual realization.
5. Closed-loop evaluation: repeated perception, planning, action, and replanning in an environment that responds to those actions.
6. Uncertainty quantification: alternative futures, calibrated confidence, and ambiguity estimates.
7. Physical validation: contact, collision, friction, object permanence, deformation, and long-horizon consistency.
8. Cross-scale composition: local observations that remain connected to global spatial, temporal, and semantic context.
9. Efficient inference: distilled policies, sparse foresight, hierarchical planning, or segment-level computation.
10. Verified or inspectable dynamics: code, constraints, or other mechanisms that make world evolution testable rather than merely plausible.

The central controversy is definitional. Some systems call a video generator a world model because it predicts future observations; others reserve the term for an embodied, persistent, action-conditioned simulator. OpenWorldLib and the conceptual analysis of world-model research favor a capability-oriented definition centered on perception, interaction, long-term memory, internal representation, and prediction (Team et al., 6 Apr 2026, Zeng et al., 2 Feb 2026). WorldLens similarly treats world quality as a multidimensional property rather than an image-quality score.

A cautious synthesis is therefore appropriate. “World in World” does not identify one model class or guarantee complete simulation. It names a design problem: construct an internal world that remains coherent when it is queried, rendered, reconstructed, compared with reality, subjected to alternative actions, and used for planning. Latent world models address compact state and prediction; programmatic models address inspectable mechanisms; 3D and 4D systems address geometry and persistence; closed-loop platforms address embodied utility; and cross-scale systems address compositional world structure.

The strongest systems combine several of these properties, but none of the cited approaches establishes a complete, general-purpose internal universe. Current models remain conditional, distribution-dependent, computationally constrained, and vulnerable to compounding error, perception failure, causal misattribution, or inconsistent rendering. The technically defensible meaning of “World in World” is consequently an internal, evolving model of an external environment—one whose credibility depends on its ability to preserve state, predict consequences, support action, and remain consistent across representations and time.