Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond

TL;DR

A position-driven survey from a 49-author HKUST/CUHK/NUS/Oxford-led group that proposes a two-axis taxonomy for world modeling: a three-level capability hierarchy (L1 Predictor → L2 Simulator → L3 Evolver) crossed with four governing-law regimes (physical, digital, social, scientific). L1 is one-step local transition learning; L2 composes multi-step action-conditioned rollouts that must satisfy domain constraints; L3 autonomously revises the model itself when predictions fail against new evidence. The framework is used to synthesize 400+ works and 100+ representative systems across model-based RL, video generation, web/GUI agents, multi-agent simulation, and AI-driven scientific discovery, with the explicit goal of giving previously isolated communities a shared vocabulary and a set of testable boundary conditions.

Key claims

• The L1/L2/L3 hierarchy is runtime-dispatched, not a static system classification: the same deployed agent can invoke L1 (fast reactive prediction), L2 (multi-step rollout for planning / counterfactuals / constraint checks), or L3 (model revision) depending on task demand [§1.2].
• L2 is distinguished from L1 by three complementary boundary conditions — long-horizon coherence under composition, intervention sensitivity, and constraint consistency — and these can trade off against one another (e.g. aggressive constraint enforcement can reduce intervention responsiveness) [§2.4].
• “Closed-loop use” (planning inside a world model) is orthogonal to L1/L2/L3 and must not be conflated with L3: a weather emulator can be an L2 WM with no embedded planner; only revising the WM stack itself from deployment evidence counts as L3 [§2.4].
• The four governing-law regimes are partitioned by how constraints are accessed: physical/digital regimes admit analytic or simulator-based verification; social regimes are constrained by beliefs/goals/norms; scientific regimes require empirical validation because mechanisms are only partially known [§1.2].
• Existing surveys are modality- or domain-partitioned (Ding et al. 2025a’s understanding-vs-predicting split, Zhu et al. 2024’s Sora-centric video framing, Yue et al. 2025’s G1–G4 visual taxonomy); this paper explicitly positions L1–L3 as capability-centric and modality-abstract and complementary to Yue’s G-levels [§1.1].
• The survey takes a normative stance on representation: latent dynamics suffice as a scaffold for L1/L2, but genuine L3 model revision requires a symbolic substrate because implicitly-encoded invariances (translation equivariance, shape bias) are not directly inspectable or modifiable [§2.2].
• Philosophical anchors are explicit: L1 ≈ Hume’s constant conjunction (statistical succession), L2 ≈ Lewis’s closest-possible-worlds counterfactuals, L3 ≈ Lakatos’s hard-core/protective-belt revision plus Duhem–Quine holism for failure-blame assignment [§2.1].
• A fourth “meta-world modeling” stage where the governing laws themselves become learnable is named as an open frontier beyond L3, and listed as a Section 8 open-problems direction [§1.3, §8].
• The agent–world-model coupling is bidirectional: agents rely on WMs for look-ahead planning and sample-efficient learning, and WMs benefit from agent-generated experience that produces task-relevant trajectories in decision-critical state-space regions [§1].

Method

The paper is a position-driven survey, not a benchmark or new model. Its method is conceptual cartography backed by a comprehensive literature scan. It (1) motivates the three capability stages from epistemological intuition (Hume → Lewis → Lakatos/Duhem–Quine), (2) fixes a unified POMDP-grounded notation (E = ⟨S, A, O, T, Ω, R, γ⟩ with learned latents z, encoder φ, dynamics f̂, decoder g, inverse dynamics, and explicit ^ marks for composed/approximate trajectory-level objects), (3) gives each level a formal definition with testable boundary conditions rather than informal descriptions, (4) re-organizes 400+ surveyed works and 100+ representative systems into level–regime pairs (3 levels × 4 regimes = 12-cell matrix), and (5) proposes decision-centric evaluation principles plus a “minimal reproducible evaluation package”. Two diagnostic separations are emphasized: world modeling vs. generic prediction, and world model vs. planner. The paper explicitly does not introduce a new benchmark or leaderboard.

Results

No experimental results. The substantive deliverables are:

• A 12-cell coverage map of the L1/L2/L3 × {physical, digital, social, scientific} taxonomy populated with 100+ representative systems and a per-cell failure-mode analysis.
• A vocabulary contribution: precise testable boundary conditions for L2 (long-horizon coherence + intervention sensitivity + constraint consistency) and L3 (deployment-evidence-driven stack revision with held-out probes), explicitly distinguished from closed-loop deployment and from policy/planner work.
• A representation-layer argument that L3 revision is structurally easier in symbolic substrates than in latent ones — anchored in the Newton → Maxwell → Standard-Model lineage as the human prototype of L3 systems.
• A roadmap section (Section 8) introducing meta-world modeling (learnable governing laws) as a stage beyond L3, with open problems in architectural guidance, evaluation, and governance.

Why it’s interesting

This is the first filed survey on the wiki that proposes a cross-modality capability taxonomy for the WFM cluster — directly addressing the open question on World Foundation Models of whether the right axis of “WFM-ness” is generation quality, predictive latent quality, or something else. The L1/L2/L3 framing is orthogonal to A Functional Taxonomy of World Models‘s renderer/simulator/planner functional split: Fei-Fei partitions by output channel (pixels vs. state vs. actions), this paper partitions by capability level under composition. Both can be applied to the same system simultaneously, and together they give the cluster a 2D vocabulary it previously lacked. The paper also touches the Tool-Use Agents and Reasoning RL clusters by explicitly including web/GUI/social/multi-agent systems as world-modeling targets — a framing that complements Qwen-AgentWorld: Language World Models for General Agents‘s “language world models for agentic environments” as a concrete L2-in-the-digital-regime instance. The L3-requires-symbolic-substrate stance is a sharp counterpoint to The flavor of the bitter lesson for computer vision‘s “explicit-3D should dissolve” position: Sitzmann argues for end-to-end latent generative rollouts; this paper argues that without inspectable structure, the system caps out at L2.

See also

• World Foundation Models — primary concept page; this survey’s L1/L2/L3 axis is the cleanest capability-centric vocabulary filed for the cluster so far.
• A Functional Taxonomy of World Models — Fei-Fei Li’s renderer/simulator/planner taxonomy; orthogonal axis (output channel vs. capability level under composition), compatible and pairable.
• The flavor of the bitter lesson for computer vision — sharp counterposition to the L3-needs-symbolic-substrate stance; argues latent generative rollouts should replace explicit structure.
• Awesome World Models — domain-partitioned awesome-list; this survey is the capability-partitioned counterpart that catalogs many of the same systems under a different vocabulary.
• Qwen-AgentWorld: Language World Models for General Agents — concrete instance of an L2-class WM in the digital-world regime (browsers, code interpreters, terminals).
• The DAWN of World-Action Interactive Models — concrete instance of L2 with action-conditional rollout in the physical-world regime; recursive coupling formalism complements the paper’s ^p̂(z_{t+1:t+h} | z_t, a_{1:h}) rollout query.
• Learning Causal World Models: A Closed-Loop Recipe for Exploration, Representation, and Decision Making — operationalizes an L2→L3 path via causality-guided exploration and a dynamics-consistency critic; closest filed instance of the L3 boundary the paper formalizes.
• Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation — pixel-side language-conditioned WM; concrete L2 instance in the physical-world regime.