Summary (Overview)

• This paper introduces a "levels × laws" taxonomy for world modeling in agentic AI, organizing the field along two orthogonal axes: three capability levels (L1 Predictor, L2 Simulator, L3 Evolver) and four governing-law regimes (physical, digital, social, and scientific worlds).
• The L1→L2→L3 hierarchy formalizes a capability progression from one-step local prediction, through decision-usable multi-step simulation with constraint satisfaction, to evidence-driven autonomous model revision—with testable boundary conditions for each level.
• The paper synthesizes over 400 works and 100+ representative systems spanning model-based reinforcement learning, video generation, web/GUI agents, multi-agent social simulation, and AI-driven scientific discovery, connecting previously isolated research communities.
• The authors propose decision-centric evaluation principles and a minimal reproducible evaluation package, along with architectural guidance, open problems, and governance challenges.
• A key contribution is the formalization of L3 (Evolver) as a distinct capability level—evidence-driven model growth where systems autonomously collect new evidence and revise their own dynamics models.

Introduction and Theoretical Foundation

Background and Motivation

The paper addresses the conceptual fragmentation of the term "world model" across research communities:

• In reinforcement learning, world models learn transition structure to imagine futures before acting (Sutton, 1991; Ha and Schmidhuber, 2018; Hafner et al., 2020)
• In computer vision, world models denote video or 3D generators maintaining visual dynamics (Brooks et al., 2024; Bruce et al., 2024)
• In language modeling, the term refers to text-grounded simulation for planning and social environments (Wang et al., 2024d; Park et al., 2023)
• In robotics, learned dynamics serve safe planning and sim-to-real transfer (Wu et al., 2023a)

Philosophical Foundations

The capability hierarchy draws on epistemological traditions:

• L1 (Predictor): Grounded in Hume's constant conjunction—extracting statistical co-occurrences without certifying causal necessity
• L2 (Simulator): Grounded in Lewis's theory of closest possible worlds—counterfactual reasoning exploring worlds maximally similar to our own
• L3 (Evolver): Grounded in Lakatos's hard core vs. protective belt distinction and Duhem-Quine holism for blame assignment in model revision

Formal Framework

The environment is formalized as a Partially Observable Markov Decision Process (POMDP):

$$\mathcal{E} = (\mathcal{X}, \mathcal{A}, \Omega, T, O, R, \gamma)$$

with transitions following:

$$x_{t+1} \sim T(x_{t+1} \mid x_t, a_t), \qquad o_t \sim O(o_t \mid x_t)$$

Methodology

L1 Predictor: Local Markov Prediction

L1 factorizes into four local operators:

Representative methods include:

• State inference: CPC, SimCLR, MoCo, CURL, SPR, I-JEPA, V-JEPA, DINOv2, RSSM
• Forward dynamics: PILCO, MuZero, Dreamer family, PETS, TD-MPC2, IRIS, DIAMOND
• Observation decoding: VAE, β-VAE, VQ-VAE, latent diffusion models
• Inverse dynamics: Pathak et al. (2017) curiosity-driven exploration, Baker et al. (2022) Video PreTraining

L2 Simulator: Decision-Usable Multi-Step Simulation

L2 extends L1 to trajectory-level queries:

$$\hat{p}(\tau \mid z_0, a_{1:H}, c), \qquad \tau = (z_1, \ldots, z_H)$$

Three boundary conditions mark L1→L2:

1. Long-horizon coherence: rollouts remain usable over H steps without degradation from compounding error
2. Intervention sensitivity: counterfactual edits induce stable, directionally meaningful trajectory changes
3. Constraint consistency: generated futures respect governing laws of the target regime

L3 Evolver: Evidence-Driven Model Revision

L3 extends L2 through an explicit update loop:

$$(\mathcal{M}_t, d_t) \xrightarrow{\text{diagnose + distill + validate}} \mathcal{M}_{t+1}$$

Three boundary conditions mark L2→L3:

1. Evidence-grounded diagnosis: failures attributed to actionable causes using replayable evidence
2. Persistent asset update: fixes promoted as reusable assets (skills, rules, parsers, tests)
3. Governed validation: updates pass regression and robustness gates before default enablement

Governing-Law Regimes

Empirical Validation / Results

Key Findings Across Levels

L1 systems demonstrate that:

• Latent-space "imagination" rollouts in model-based RL can match or exceed model-free baselines (Hafner et al., 2025; Schrittwieser et al., 2020)
• DreamerV3 achieves state-of-the-art performance across diverse domains including Atari, continuous control, and Minecraft with a single configuration
• EfficientZero achieves superhuman performance in Atari with just two hours of experience

L2 systems reveal that:

• Video generation models like Sora (Brooks et al., 2024) produce photorealistic frames but face challenges in long-horizon coherence and intervention sensitivity
• Web and GUI agents require deterministic program semantics for reliable planning (Gu et al., 2025b)
• Multi-agent social simulation requires modeling reflexive transitions where beliefs change the state itself (Park et al., 2023)

L3 systems demonstrate:

• Scientific discovery pipelines pair surrogate models with hypothesis-driven experimentation (Lu et al., 2024a; Boiko et al., 2023)
• Autonomous experimentation systems can design, execute, and reflect on experiments to revise models

Historical Development Timeline

The paper traces world modeling through four eras:

• Mathematical Principles (–1956)
• Symbolic Intelligence (1956–1986)
• Connectionist Resurgence (1986–2020)
• Generative Revolution (2020–present)

with two AI winters (1974–1980, 1987–1993) marking paradigm transitions.

Theoretical and Practical Implications

Theoretical Contributions

1. Unified capability framework: The L1/L2/L3 hierarchy provides testable boundary conditions that cut across modalities and application domains, addressing the conceptual fragmentation in the field.

2. Representation question: The paper argues that the endpoint of world modeling may require symbolic discovery, with neural latents as scaffold—drawing parallels to how Newton's laws and Maxwell's equations represent successful human L3 systems.

3. L3 as distinct capability: Evidence-driven model growth is qualitatively different from L2 rollout, requiring the model itself to become an object of revision.

Practical Implications

• Decision-centric evaluation: The paper proposes evaluating world models by the quality of decisions they enable, not just prediction accuracy or visual fidelity
• Cross-domain transfer: The framework reveals shared principles across communities (e.g., constraint consistency applies to physics engines and social norms alike)
• Architectural guidance: Provides direction for building systems that can escalate from L1 to L2 to L3 based on task demands

Conclusion

Main Takeaways

The paper establishes a capability-based roadmap for world modeling in agentic AI, formalizing the progression from local prediction (L1) through decision-usable simulation (L2) to evidence-driven model revision (L3). The "levels × laws" taxonomy connects previously isolated communities—computer vision, reinforcement learning, robotics, NLP, and AI for science—under a common conceptual framework.

Future Directions

The paper identifies several open problems:

• Meta-world modeling: where governing laws themselves become learnable
• Partial L3 loops: systems that implement some but not all L3 boundary conditions
• Evaluation methodology: the need for decision-centric evaluation principles and reproducible evaluation packages
• Governance challenges: ensuring model revisions pass regression and robustness gates
• Architectural design: balancing latent dynamics (scaffold) with symbolic representations (revisable substrate)

The roadmap charts a path from passive next-step prediction toward world models that can simulate and ultimately reshape the environments in which agents operate.