Ctrl-World Models

• Ctrl-World Models are generative world models that decompose simulation into a deterministic physics layer and a stochastic, schema-constrained LLM imagination layer.
• They maintain object permanence and simulation consistency using deterministic seeding and typed JSON contracts that enforce strict invariants.
• This hybrid approach enables scalable applications in robotics, navigation, and open-ended simulations, with empirical gains in simulation fidelity and control.

A Ctrl-World Model is a class of generative world models that prioritize controllability by enforcing explicit, physically meaningful structure within the learned or generated environment dynamics. Specifically, these models employ a decomposition between a strictly code-defined, deterministic "physics" layer and a flexible, model-driven "imagination" layer, typically powered by LLMs or high-capacity generative backbones. This hybrid architecture enables robust, consistent, and interpretable simulation of open-ended, interactive environments, supporting agents in tasks that demand both logical consistency and rich contextual generation (Feng et al., 29 Dec 2025, He et al., 1 Dec 2025, Guo et al., 11 Oct 2025). The concept is formalized through the lens of persistent state decomposition, schema-constrained imagination, and deterministic content generation for object permanence and fidelity control.

1. Formal Definition and Mathematical Structure

A Ctrl-World Model is formally defined as a tuple

$M = (S, A, T, G, \pi_\theta)$

where:

• $S = S^\phi \times S^\psi$ denotes the latent state space, with:
  • $S^\phi$: physics/symbolic state—deterministic, code-defined (e.g., object locations, resource inventories, topology).
  • $S^\psi$: imagination/perceptual state—LLM- or generator-produced, high-dimensional (e.g., natural language descriptions, dialogue, images).
• $A$: the action space (including agent commands, user input, navigation).
• $T$: the transition operator, factorized as:

$S^\phi_{t+1} = f_\text{code}(S^\phi_t, a_t)$

$$S^\psi_{t+1} \sim \pi_\theta( \cdot \mid S^\phi_{t+1} )$$

i.e., the physics layer is updated deterministically by code; the imagination layer is stochastically generated conditioned on new physics.

• $G: S \to \text{Text}$ or $\text{JSON}$ is a structured generation function (often a template or rendering overlay).

A distinctive engineering innovation is deterministic seeding: for any world location or key $x$, a hash function $h(x)$ is used to seed the generative model, yielding

$S^\psi(x) = S^\psi(x)$

across repeated visits, thus ensuring object permanence without explicit database storage.

2. Architectural Breakdown and Implementation Pipeline

Ctrl-World Models are implemented as a two-layer pipeline:

• Layer 1 (Physics/Code): Hosted in strongly typed languages (e.g., TypeScript). Enumerates all valid entities, enforces invariants, and encodes symbolic transitions for perfect logical consistency. Exposes strict, observable interfaces to higher layers.
• Layer 2 (Imagination/LLM): Invoked after physics updates. Prompts an LLM or generator to populate or re-texturize the perceptual state, constrained by a strict JSON or schema interface. This layer is stateless and can degrade gracefully (e.g., templated fallback) under latency or reliability constraints.

The contract between layers is a typed interface (e.g., JSON Schema or TypeScript interface). All outputs from the LLM must validate against this schema, with immediate enforcement.

Pseudocode:

The following summarization illustrates the step sequence:

// Physics update (deterministic) function physicsUpdate(state: PhysicsState, action: Action): PhysicsState { ... } // Generate deterministic seed for location/object function computeSeed(key: string): number { return hash32(key); } // Imagination update (LLM call with schema validation) async function imaginationUpdate(phi: PhysicsState): Promise<ImaginationState> { ... } // Full pipeline step async function step(state: FullState, action: Action): Promise<FullState> { const phiNext = physicsUpdate(state.phi, action); const psiNext = await imaginationUpdate(phiNext); return { phi: phiNext, psi: psiNext }; }

(Feng et al., 29 Dec 2025)

3. Design Principles: Controllability, Fidelity, and Contracts

Several engineering principles distinguish Ctrl-World Models from classical or fully generative approaches:

• Separation of Concerns: All core invariants and hard constraints are maintained in the physics layer (e.g., collision, inventory management), while the imagination layer handles only creative elaboration under contract.
• Typed Interfaces: JSON or strongly typed schemas serve as contracts for all AI-generated outputs, enabling rigorous enforcement and reliable downstream integration.
• Infinite Worlds and Object Permanence: Deterministic seeding via hashing supports infinite open-ended environments with strict object permanence, eliminating the requirement for persistent storage on-the-fly content.
• Fidelity Slider/Graceful Degradation: Supports multiple fidelity levels—ranging from direct generative model output, through cached based-on-seed resolutions, to fixed template fallback—without compromising deterministic code-level behavior.
• Microservice Protocols: The LLM layer is orchestrated as a microservice with strict input/output requirements (REST/gRPC, JSON payloads), facilitating distributed deployment and robust scaling. (Feng et al., 29 Dec 2025)

4. Instantiations and Concrete Examples

Ctrl-World Models support a variety of research and production settings. Representative instantiations include:

• Infinite Travel Atlas: The physics layer encodes geospatial coordinates, environment class, and available themes; the imagination layer yields a JSON travel guide with descriptions, itineraries, and tips. Deterministic seeding by (lat, lon) enforces object permanence.
  • $$\displaystyle S^\psi(\text{lat},\text{lon}) = \pi_\theta(\cdot|S^\phi(\text{lat},\text{lon}); \text{seed}=h(\text{lat},\text{lon}))$$
• Fictional Galaxy Explorer: The physics layer is a procedural graph of star systems and planets, each with code-defined seeds and topologies; the LLM layer generates descriptive narratives for individual planets, enforcing both navigational consistency and rich contextual flavor.
  • $$\displaystyle S^\psi_\text{planet} \sim \pi_\theta(\cdot | \text{biome}, \text{hazard}, \dots; \text{seed}=h(\text{planetID}))$$
  • (Feng et al., 29 Dec 2025)

These examples clarify that all unbounded content is generated on demand, driven by hash-derived seeds. The LLM is strictly contract-constrained and never allowed to override symbolic invariants.

5. Control and Grounding: Post-Training via Reward Alignment

For Ctrl-World Models that underpin embodied navigation, manipulation, or multimodal embodied reasoning, additional stages may be necessary to ensure groundedness and robust control. GrndCtrl [Editor's term: Grounded Ctrl-World Model] instantiates Reinforcement Learning with World Grounding (RLWG), which post-aligns a pretrained diffusion world model to physical reality via self-supervised, verifiable geometric/perceptual rewards (e.g., pose cycle-consistency, depth reprojection, temporal coherence). The alignment is performed by Group Relative Policy Optimization (GRPO), treating the generative model as a policy optimized via PPO-style clipped objectives, group-normalized multi-reward advantage, and stochastic gradient ascent.

Explicitly, reward alignment drives the model to:

• Penalize pose drift and rotational error by enforcing groupwise cycle-consistency in translation and orientation,
• Maximize frame-to-frame depth reprojection inliers,
• Encourage temporal coherence via penalizing abrupt transitions in feature space.

GRPO provides stable, high-variance-tolerant optimization through rollouts and group-normalized advantage estimation, yielding world models with up to 64% lower counterfactual translation error in navigation benchmarks compared to nominally pretrained models (He et al., 1 Dec 2025).

6. Applications and Empirical Performance

Areas of deployment span:

• Robust manipulation policy evaluation and improvement: Ctrl-World models with multi-view perception, long-horizon pose anchoring, and fine-grained action control support realistic, consistent 20 s rollouts for generalist robot policies, facilitating both imagination-driven evaluation and policy improvement through synthetic data (44.7 percentage point improvement in OOD tasks) (Guo et al., 11 Oct 2025).
• Large-scale, open-ended simulation: Infinite atlas or galaxy worlds scalable to web scale, leveraging schema-based imagination layered atop code-generated symbolic scaffolding (Feng et al., 29 Dec 2025).
• Grounded navigation and planning: Object permanence, cycle-consistent geometry, and contract-based LLM outputs enable long-horizon, spatially coherent planning with error and drift bounds through RLWG/GRPO alignment (He et al., 1 Dec 2025).

Empirical studies emphasize that separation of concerns and schema contracts are essential for both reliability and extensibility. Fully generative models, when not grounded or contractually limited, exhibit drift, hallucination, or inconsistent control over long horizons.

---

In summary, Ctrl-World Models represent a rigorously engineered, hybrid class of world models, uniquely characterized by strict symbolic scaffolding, schema-driven stochastic imagination, and deterministic generation for infinite, persistent environments. Grounding via self-supervised reward alignment further extends these models to physically reliable embodied settings. This framework is advocated as a foundation for scalable, controllable, and extensible world simulation in both research and production domains (Feng et al., 29 Dec 2025, He et al., 1 Dec 2025, Guo et al., 11 Oct 2025).