PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Meng Chu, Xuan Billy Zhang, Kevin Qinghong Lin, Lingdong Kong, Jize Zhang, Teng Tu, Weijian Ma, Ziqi Huang, Senqiao Yang, Wei Huang, Yeying Jin, Zhefan Rao, Jinhui Ye, Xinyu Lin, Xichen Zhang, Qisheng Hu, Shuai Yang, Leyang Shen, Wei Chow, Yifei Dong, Fengyi Wu, Quanyu Long, Bin Xia, Shaozuo Yu, Mingkang Zhu, Wenhu Zhang, Jiehui Huang, Haokun Gui, Runyi Li, Chenyu Tang, Dong Huang, Xuhang Chen, Rui Liu, Chengzu Li, Shiyi Du, Xu Huang, Haoxuan Che, Long Chen, Qifeng Chen, Wenya Wang, Wenxuan Zhang, Xiaojuan Qi, Yang Deng, Yanwei Li, Mike Zheng Shou, Zhi-Qi Cheng, See-Kiong Ng, Ziwei Liu, Philip Torr, Jiaya Jia

arXiv:2604.22748v3cs.AI

TL;DR

World-model research lacks a shared framework across communities using different meanings and evaluation criteria. This paper introduces a levels × laws taxonomy and synthesis showing that world-model capabilities range from prediction to simulation and autonomous revision across physical, digital, social, and scientific regimes.

  • Problem

    World-model research is conceptually fragmented across communities, limiting comparability between systems evaluated under different meanings and criteria.

    Source: §1 · p. 3, §1.1 · p. 4

  • Method

    The paper synthesizes over 400 works using a taxonomy of three capability levels and four governing-law regimes, with decision-centric evaluation principles.

    Source: §9 · p. 52

  • Results

    The synthesis finds that scientific L3 systems are established, whereas social simulation remains premature, with maturity varying across governing-law regimes.

    Source: §4.2.3 · p. 28, §5.4 · p. 38

  • Takeaways & Limitations

    The taxonomy provides a common language and roadmap for developing and evaluating world models from local prediction toward law-consistent simulation and autonomous revision.

    Source: §1 · p. 3, §9 · p. 52

    Evaluation remains constrained by benchmark saturation, evaluation gaming, costly human judgment, and limited systematic meta-evaluation.

    Source: §6.5 · p. 42

Abstract

from arXiv · show

As AI systems move from generating text to accomplishing goals through sustained interaction, the ability to model environment dynamics becomes a central bottleneck. Agents that manipulate objects, navigate software, coordinate with others, or design experiments require predictive environment models, yet the term world model carries different meanings across research communities. We introduce a "levels x laws" taxonomy organized along two axes. The first defines three capability levels: L1 Predictor, which learns one-step local transition operators; L2 Simulator, which composes them into multi-step, action-conditioned rollouts that respect domain laws; and L3 Evolver, which autonomously revises its own model when predictions fail against new evidence. The second identifies four governing-law regimes: physical, digital, social, and scientific. These regimes determine what constraints a world model must satisfy and where it is most likely to fail. Using this framework, we synthesize over 400 works and summarize more than 100 representative systems spanning model-based reinforcement learning, video generation, web and GUI agents, multi-agent social simulation, and AI-driven scientific discovery. We analyze methods, failure modes, and evaluation practices across level-regime pairs, propose decision-centric evaluation principles and a minimal reproducible evaluation package, and outline architectural guidance, open problems, and governance challenges. The resulting roadmap connects previously isolated communities and charts a path from passive next-step prediction toward world models that can simulate, and ultimately reshape, the environments in which agents operate. Code and resources are available at: https://github.com/matrix-agent/awesome-agentic-world-modeling.

1 Introduction

This survey addresses the fragmented meaning of world models by defining their value through downstream decision support and organizing them by capability level and governing-law regime. It proposes the L1/L2/L3 hierarchy, distinguishes evidence-driven model revision as L3, and synthesizes cross-domain research without introducing a new benchmark.

1 Introduction

  • The survey defines world models as learned or hybrid transition operators whose value lies in enabling informed agent decisions through prediction, planning, simulation, and scientific reasoning.Agents provide targeted experience that improves accuracy in decision-critical state-space regions, while world models support look-ahead planning and sample-efficient learning.Source: §1 · p. 3

1.1 Motivation

  • Existing surveys are fragmented by modality or application domain, so systems evaluated for visual fidelity and task performance remain difficult to compare.This survey supplies a common language that preserves domain-specific differences while aligning communities.Source: §1 · p. 3, §1.1 · p. 4

1.2 Scope and Organizing Principle

  • The four regimes encode distinct constraints: physical interaction, digital program semantics, social beliefs and norms, and scientific mechanisms, observables, and causal structure.Physical transitions may admit analytic or simulator verification, whereas scientific mechanisms typically require empirical validation because they are only partially known.Source: §1.2 · pp. 5–6
  • The taxonomy is a capability coordinate system rather than a model classification: L1 predicts locally, L2 rolls out constrained alternatives, and L3 revises the model when evidence exposes systematic failures.A deployed agent may invoke different levels at runtime; L3 governs revision while L1 and L2 remain necessary for reactive prediction and planning.Source: §1.2 · p. 6, Key Contributions · p. 7

1.1 Motivation

  • The proposed L1/L2/L3 hierarchy captures one-step prediction, decision-usable long-horizon simulation, and evidence-driven model revision across physical, digital, social, and scientific regimes.Its boundary conditions emphasize rollout, intervention sensitivity, constraint consistency, and closed-loop use rather than generic prediction.Source: §1.1 · p. 4, §1.2 · p. 5, Key Contributions · p. 7

Key Contributions

  • The paper’s principal contribution is a cross-domain capability-by-law framework connecting computer vision, language modeling, reinforcement learning, robotics, and AI for science.It argues that shared principles can be revealed across isolated communities while domain-specific constraints make direct transfer non-trivial.Source: §1.1 · p. 5, Key Contributions · p. 7
  • Unlike a pure survey or position paper, the work proposes testable boundaries and reviews existing systems, but introduces neither a new benchmark nor a leaderboard.Its roadmap covers foundations, capability levels, evaluation, architecture, computational considerations, and open problems including meta-world modeling.Source: Key Contributions · p. 7

2 Preliminaries

The preliminaries define world modeling as a three-level capability hierarchy—L1 Predictor, L2 Simulator, and L3 Evolver—formalized in a unified POMDP-based notation. They also argue that representation and governing-law regimes determine whether these capabilities can support reliable prediction, simulation, and evidence-driven revision.

2.1 Philosophical Motivations

  • Epistemological analogies motivate L1 prediction, L2 counterfactual simulation, and L3 evidence-driven model revision, while emphasizing that these analogies are heuristic rather than claims about philosophical implementation.L1 learns one-step regularities but is fragile under distribution shift; L2 produces model-relative rollouts that may drift; L3 revises model structure when persistent failures demand it.Source: §2.1 · pp. 8–9

2.2 Representation in World Modeling: Lessons from Scientific Theories

  • Representation is foundational: latent dynamics scaffold L1 and L2, whereas L3’s genuine revision of governing laws requires an inspectable and modifiable symbolic substrate.Scientific theories illustrate explicit, revisable, composable models, while implicit neural invariances are difficult to modify when the model structure itself must change.Source: §2.2 · pp. 9–11

2.3 Notations

  • The notation grounds the taxonomy in a POMDP, distinguishing hidden states, observations, actions, latent states, one-step dynamics pθ, composed trajectories ˆp, and planners or policies that query but do not constitute the world model.The convention defines ˆp for composed trajectory-level objects such as ˆp(τ | z0, a1:H, c), while ordinary one-step dynamics remain pθ(zt | zt−1, at).Source: §2.3 · pp. 11–12

2.4 Definitions of Capabilities

  • The capability definitions establish a containment hierarchy in which L2 composes L1 operators into rollouts and L3 uses L2 interaction to revise the model stack.Figure 7 scopes L1 to a single latent transition, L2 to trajectory rollout under a fixed model, and L3 to evidence-driven revision from Mt to Mt+1.Source: Figure 7 · p. 10, §2.4 · p. 12, L3 Evolver · p. 13

L1 Predictor

  • L1 Predictor provides local one-step or short-horizon operators optimized under the training distribution, without guaranteeing coherent multi-step composition.Its forward dynamics may be represented as pθ(zt | zt−1, at) or pθ(zt | zt−1).Source: L1 Predictor · p. 12

L2 Simulator

  • L2 Simulator is defined by decision-usable rollouts whose long-horizon coherence, intervention sensitivity, and constraint consistency distinguish rollout fidelity from one-step quality.These conditions are complementary: coherent trajectories can remain action-insensitive, while action-sensitive trajectories can violate domain constraints; closed-loop deployment is orthogonal to the capability level.Source: L2 Simulator · pp. 12–13

L3 Evolver

  • L3 Evolver treats the world-modeling stack as a revisable object, requiring evidence-grounded diagnosis, persistent asset updates, and governed validation through regression, robustness, rollback, and canary gates.Its meta-actions can include diagnostic probes, architecture modifications, and regression tests, enabling revision across the temporal and abstraction scales relevant to planning.Source: L3 Evolver · pp. 13–14

2.5 Scope of Laws

  • The four governing-law regimes are physical, digital, social, and scientific, differing in whether constraints are analytically verifiable, mechanically checkable, normatively reflexive, or learned and validated from empirical evidence.Real systems may combine regimes, such as autonomous driving’s physical and social constraints or drug design’s natural mechanisms and digital simulation pipelines.Source: §2.5 · pp. 14–15

3 L1 Predictor: Local Markov Prediction

L1 Predictor models one-step or short-horizon, action-conditioned transitions over a learned Markov-like internal state, using local operators that support decision-making. Its components enable prediction but do not guarantee coherent multi-step behavior, motivating L2.

3.1 Definition

  • L1 defines local prediction as maintaining an internal state z_t that anticipates the next state, observation, or action at one-step or short fixed-horizon granularity.The learned state approximates a latent environmental state or belief without asserting causal necessity.Source: §3 · p. 15, §3.1 · p. 15
  • L1 factorizes into state inference, forward dynamics, observation decoding, and inverse dynamics, with forward dynamics serving as the core transition operator.These operators map observations to states, states to next states, states to observations, and successive states to actions.Source: §3.1 · pp. 15–16, Table 2 · p. 16

3.2.1 State Inference

  • State inference compresses observations and history into decision-relevant latent beliefs, using contrastive, predictive, recurrent, auxiliary-objective, and structured representations.RSSM combines deterministic recurrence with stochastic inference to preserve uncertainty from arbitrary-length histories.Source: §3.2.1 · p. 17

3.2.2 Forward Dynamics: The Core L1 Operator

3.2.3 Observation Decoding

  • Observation decoding reconstructs or renders observations, supplying training signals, representation diagnostics, and imagined outputs while leaving coherent multi-step dynamics as an unresolved challenge.VAE-based decoders underpin World Models (Ha and Schmidhuber, 2018) and the Dreamer lineage (Hafner et al., 2020; 2021; 2025).Source: §3.2.3 · p. 19

3.2.4 Inverse Dynamics

  • Inverse dynamics predicts actions between latent states and can shape representations, support exploration, retrospectively label videos, and connect subgoals to low-level execution.Baker et al. (2022) used retrospective action labeling to enable Video PreTraining from unlabeled Minecraft videos.Source: §3.2.4 · p. 19
  • Inverse dynamics is optional and unreliable when inferred action labels are noisy, actions are nonunique, or environments are stochastic.Errors accumulate at distribution edges where reliable supervision is most needed.Source: §3.2.4 · p. 20

3.3 Discussion

  • L1 may model local transitions along temporal or refinement axes, but one-step accuracy alone does not ensure long-horizon consistency, intervention, counterfactual reasoning, or decision-usable behavior.L2 instead evaluates multi-step rollout accuracy and constraint adherence, while adaptation to new evidence is reserved for L3.Source: §3.3 · p. 20

4 L2 Simulator: Decision-Usable Multi-Step Simulation

This section examines L2 simulators that compose action-conditioned transitions into decision-relevant, multi-step rollouts across physical, digital, social, scientific, and cross-domain settings while preserving governing-law constraints. It surveys elevation criteria, applications, and failure modes, emphasizing decision-aware modeling, joint constraint satisfaction, explicit evidence, controllability, and frequent verification.

4 L2 Simulator: Decision-Usable Multi-Step Simulation

  • L2 enables agents to roll out candidate action sequences under constraints, compare likely trajectories, and select actions before real-environment interaction.Unlike L1, L2 requires coherent multistep rollout under governing laws rather than one-step predictive quality alone.Source: §4 · p. 20

4.1 Requirements for Elevation

  • An L2 simulator requires long-horizon coherence, intervention sensitivity, and constraint consistency, with evaluation focused on decision usability as rollout horizons increase.Diagnostics include success cliffs, action-insensitive rollouts, and constraint-violation rates; mitigation strategies include replanning, action-consistency checks, hard constraints, and verification gates.Source: §4.1 · pp. 21–22

4.2 Applications

  • Applications span four governing-law regimes, whose domain-specific constraints determine how L2 boundary conditions are instantiated and evaluated.This part surveys physical, digital, social, and scientific regimes, although the detailed applications presented here cover physical, digital, and social worlds.Source: §4.2 · p. 22

4.2.1 Laws of the Physical World

  • Physical-world simulators range from explicit physics engines to action-conditioned video worlds, but visual plausibility alone does not guarantee intervention sensitivity, long-horizon coherence, or constraint consistency.Standard video metrics capture realism, while VBench-style suites, VBench-2.0, and VChain assess controllability, physics and commonsense consistency, and causal coherence.Source: §4.2.1 · pp. 22–24

4.2.2 Laws of the Digital World

  • Digital-world models exploit mechanically verifiable program semantics, with code-based simulators supporting interpretable rollouts and SWE-World lifting a 32B coder from 6.2% to 52.0% on SWE-bench Verified.Web, GUI, game, and tool-calling systems predict structured state transitions, though agents may rarely invoke simulators, misuse rollouts, or degrade in performance.Source: §4.2.2 · pp. 25–27

4.2.3 Laws of the Social World

  • Social simulators must preserve partially observable mental states, commitments, norms, and persistent identities, yet current systems exhibit illusory theory of mind, role drift, goal forgetting, and weak commitment tracking.ExploreToM reports accuracy as low as 9% for GPT-4o (Chen et al., 2025d), while CICERO achieves more than 2× the average human score in Diplomacy.Source: §4.2.3 · pp. 27–28

4.2.4 Laws of the Scientific World

  • Scientific L2 simulators support forward system-dynamics forecasting and decision simulation over sequential experiment choices, but fixed-data systems cannot actively gather evidence to resolve uncertainty or accumulated bias.Forward models replace expensive solvers with learned operators, while decision simulators update beliefs over candidate experiments; L3 systems address the remaining limitation by actively gathering evidence.Source: §4.2.4 · pp. 28–29

4.2.5 Cross-Domain Analysis

  • Across regimes, effective simulators should model governing constraints—such as geometry, state machines, norms, or evidence chains—because explicit, loggable, replayable constraints can improve long-horizon stability more than perceptual fidelity.Cross-regime deployments require joint constraint satisfaction because violations can cascade across physical, digital, social, and scientific domains.Source: §4.2.5 · p. 30

4.3 Failure Modes

  • L2 systems face five recurring failures: compounding error, state aliasing and drift, controllability failure, simulator escape, and calibration failure under distribution shift.Mitigations include shorter verified planning segments with replanning, explicit state verification and retrieval, action-sensitive models, hardened evaluation, and treating confident errors under shift as evolution signals.Source: §4.3 · p. 30
  • These failures are system-level pathologies arising from representation, rollout horizon, control, and evidence quality, so average-case prediction gains are insufficient without failure localization and behavioral change under shift or exploitation.A constraints-first logging and regression strategy helps prevent agents from learning incorrect lessons; explicit stable constraints can also support reliable improvement through Evolver-style asset distillation.Source: §4.3 · pp. 30–31

5 L3 Evolver: Evidence-Driven Model Revision

L3 Evolvers close the design–execute–observe–reflect loop by actively acquiring evidence, revising persistent world-model updates, and validating them across deployment. The framework distinguishes L3 from fixed L2 simulation and shows strongest maturity in scientific settings, while digital, physical, and social regimes remain partial, emerging, or aspirational.

5.1 Formal Definition

  • L3 systems actively design interventions, execute them, observe outcomes, and reflect by revising persistent world-model parameters, structure, or assumptions to reduce uncertainty.Unlike passive data fitting or fixed-model planning, each iteration targets prediction–observation discrepancies and produces reusable updates validated through regression checks.Source: §5.1 · p. 31, §5 · p. 31
  • Revision policy maps small anomalies to online adaptation, persistent anomalies to distilled parameter updates, and epistemic gaps to structural changes, but attribution across model components remains ambiguous.Latent representations flexibly absorb anomalies through parameter updates, whereas resolving missing invariances may require symbolic representations that expose and manipulate them explicitly.Source: §5.1 · pp. 32–33

5.2 Distinction from L2

  • L3 differs from L2 because deployment makes the model plastic: systems expand information, autonomously execute experiments, and revise beliefs under challenge rather than merely consuming data.Growth can involve parameter updates, architecture changes, or hypothesis-space expansion; the last mode introduces explanations unavailable in the current model class.Source: §5.2 · p. 33

5.3 Examples and Applications

  • Scientific intelligence is the most mature L3 regime: autonomous laboratories close the full loop, including CAMEO’s information-gain experimentation and A-Lab’s 353 experiments producing 36 compounds from 57 targets in 17 days.Other systems span autonomous biology, materials discovery, and chemistry, but current biomedical benchmarks still expose a gap between literature retrieval and genuine model revision.Source: §5.3 · pp. 35–36, §5.4 · p. 38

5.4 L3 in Context: Maturity, Governance, and Outlook

  • Digital L3 benefits from deterministic replay and regression gates but often lacks active information expansion; physical systems are emerging, while social systems remain aspirational because attribution and experimentation are difficult.Digital examples include Romera-Paredes et al. (2024), Novikov et al. (2025), and Butt et al. (2024); social governance rules from Kumar et al. (2026) surpassed human-designed rules by 123%.Source: §5.3 · pp. 34–35, §5.4 · p. 38
  • Persistent automated revision creates governance risks from benchmark overfitting, contaminated evidence, and misattributed fixes; held-out probes, canary deployment, and causal ablations mitigate regressions.Reliable evolution also requires falsifiable, reversible, compressible evidence with privacy-preserving storage and audit trails.Source: §5.3 · p. 37, §5.4 · p. 38

6 Evaluations

The evaluation framework shifts from perceptual prediction metrics to decision-centric tests of long-horizon coherence, intervention sensitivity, and constraint consistency across physical, digital, social, and scientific regimes. It further shows that protocol design determines whether a benchmark measures L1 prediction, L2 simulation, or L3 evidence-driven model revision, while current evaluations remain incomplete and vulnerable to saturation, gaming, and weak meta-evaluation.

6.1 From prediction-centric to decision-centric evaluation

  • Decision-centric evaluation prioritizes long-horizon coherence, intervention sensitivity, and constraint consistency because perceptually convincing rollouts can still fail during planning.Action Success Rate measures whether model-guided planning achieves real-world goals, while Counterfactual Outcome Deviation measures meaningful responsiveness to action interventions.Source: §6 · p. 38, §6.1 · p. 39

6.2 Evaluating the three boundary conditions

  • Benchmark coverage spans physical, digital, social, scientific, and open-world regimes, but no single benchmark tests all governing laws or boundary conditions.Representative anchors include Atari 100k and RoboCasa for physical tasks, OSWorld and SWE-bench for digital tasks, Sotopia for social simulation, and ScienceWorld and DiscoveryBench for scientific reasoning.Source: §6.2 · pp. 39–40

6.3 Differentiating L1, L2, and L3 via evaluation protocol

  • The same benchmark measures L1, L2, or L3 depending on protocol: local prediction, perturbation-robust decision-usable simulation, or persistent model revision from evidence.RoboCasa, OSWorld, SWE-bench, Sotopia, ScienceWorld, and DiscoveryBench illustrate how increasingly demanding protocols expose different capability levels.Source: §6.3 · p. 41
  • Current evaluation is dominated by L1 tests, while L2 protocols are nonstandard and L3 infrastructure is essentially absent outside autonomous science.Needed infrastructure includes counterfactual injection, degradation curves, constraint-violation detection, regression suites, asset validation gates, and cross-episode improvement tracking.Source: §6.3 · p. 42

6.4 Benchmarks and coverage analysis

  • WorldSimBench and WorldModelBench show that perceptual realism can diverge from action-conditioned fidelity, motivating capability matrices and the Minimal Reproducible Evaluation Package.MREP specifies version locking, trace logging, failure taxonomies, tail statistics, and boundary-condition mapping as evidence infrastructure for credible L3 evaluation.Source: §6.4 · p. 42

6.5 Open Challenges in Evaluation

  • Evaluation remains limited by benchmark saturation, gaming, costly and variable human judgment, and the lack of systematic meta-evaluation.These problems reduce discriminative power, permit benchmark-specific optimization, and leave protocol validity insufficiently tested.Source: §6.5 · p. 42

7 Architectural and Computational Considerations

World-model design is governed by choices in representation, dynamics, and control interface, whose tradeoffs determine attainable capability levels and suitability across physical, digital, social, and scientific regimes. Effective systems therefore require regime-aware composition, explicit constraint enforcement, instrumentation, and planner-aligned representations.

7.1 Architectural building blocks: representation, dynamics, and control

  • World-model implementations must jointly select representation, dynamics, and control components because each combination determines reachable capability levels and governing-law applicability.Table 11 organizes these architectural choices by their strengths and dominant failure modes.Source: §7.1 · p. 43, Table 11 · p. 44
  • Representations trade interpretability and hard constraints against multimodal flexibility, while dynamics trade uncertainty modeling, control integration, modality unification, and visual fidelity against long-horizon consistency or action controllability.Control interfaces range from replanning MPC and tree search to imagined-rollout optimization, offline distillation, and replayable environments, each with distinct compute, robustness, or distribution-shift costs.Source: §7.1 · p. 43

7.2 Design tradeoffs across governing-law regimes

  • Governing-law regimes constrain viable architectural combinations: physical systems require geometry, stability, and fast closed-loop correction; digital systems prioritize explicit state consistency; social systems must preserve identity, beliefs, and norms; and scientific systems require traceable, falsifiable evidence chains.Generative simulation instead prioritizes action-response consistency and long-horizon stability over raw perceptual realism, while regime-specific latency budgets constrain dynamics and control choices.Source: §7.2 · pp. 44–45
  • Mature systems should be evaluated as composed stacks of symbolic planning, replayable validation, and short-horizon continuous control rather than as isolated modules.The relevant unit of analysis is the composed system because representation, dynamics, and control jointly determine constraints and available evidence.Source: §7.2 · p. 46

7.3 Implementation Roadmap

  • Across capability levels and regimes, hard constraints should be enforced at inference, logging and replay should precede iteration, and representations should expose the variables required by the planner.These principles support verifiable rollouts, governable L3 revision, and planner-relevant fidelity rather than realism alone.Source: §7.3 · pp. 46–47, Table 13 · p. 47

8 Trends & Open Problems

World models are advancing from one-step predictors toward agent-facing simulators and self-revising systems, but reliability depends on decision-usable rollouts that respect governing laws and real-world evidence. The section traces this trajectory, identifies capability- and regime-specific bottlenecks, and extends the roadmap toward security-aware systems and meta-world modeling.

8.1 Historical Development

  • World-modeling progressed from mathematical and symbolic models through learned representations to generative and emerging neuro-symbolic systems, with progress driven by changes in representation, compositionality, and revisability rather than scale alone.The historical arc spans Newtonian prediction, symbolic planning and its scalability failures, neural world models, and generative systems that still exhibit systematic physics violations.Source: §8.1 · pp. 47–48

8.2 Open Problems by Capability Level

  • World models must become decision-usable by preserving governing constraints, responding to interventions, and calibrating predictions against real evidence across physical, social, digital, and scientific regimes.Current systems remain visually plausible but physically inconsistent, socially biased, digitally over-simplified, and vulnerable to surrogate-to-reality degradation.Source: §8.2 · p. 48
  • The most urgent capability-level problems are physical faithfulness, metric-aware video modeling, programmable visual representations, partially observable software, behavioral alignment, continual revision, and surrogate-to-reality transfer.PhyWorldBench’s best frontier model reaches only 0.262 success on conservation-law and object-permanence probes, while scientific surrogates degrade on real measurements and require scarce-data calibration.Source: §8.2 · pp. 49–50
  • Shared bottlenecks are detecting deployment shift, enforcing constraints at inference time, and governing persistent updates across stability, plasticity, and auditability.The proposed MREP framework offers version-locked, reproducible evaluation packages for comparing these challenges across regimes.Source: §8.2 · p. 50

8.3 Security and Safety of World Models

  • World models are high-leverage attack surfaces because corrupting imagined rollouts can propagate violations of physical, digital, social, or scientific laws into agent decisions.Reported attacks include 55% targeted success for PhysCond-WMA and 84.2% attack success for JailWAM, but these preliminary preprint findings await independent replication.Source: §8.3 · p. 51
  • Security defenses combine safety-state monitoring, counterfactual scenario generation, inference-time constraint enforcement, symbolic layers, and verification gates, while requiring regime-specific threat models and benchmarks.CounterScene raises the long-horizon collision rate from 12.3% to 22.7% over the strongest baseline, illustrating the dual-use nature of generative safety machinery.Source: §8.3 · p. 51

8.4 Beyond L3

  • Beyond L3, meta-world models would explore and compare alternative transition functions and governing laws rather than refining a model of one underlying reality.Possible implementations include program synthesis, open-ended evolution, and procedural world generation; whether this is the endpoint of world modeling remains open.Source: §8.4 · pp. 51–52

9 Conclusion

The paper concludes that agentic world modeling should be organized by capability levels and governing-law regimes, while advancing from prediction to simulation and autonomous model revision. It identifies symbolic law discovery and safe closed-loop evolution as central challenges for future systems.

9 Conclusion

  • The proposed taxonomy combines three capability levels—Predictor, Simulator, and Evolver—with four governing-law perspectives: physical, digital, social, and scientific.The levels describe increasing modeling capabilities, while the perspectives capture domain-specific constraints.Source: §9 · p. 52
  • L1 learns local transitions, L2 produces law-consistent long-horizon rollouts, and L3 revises its dynamics model through experiments and new evidence.CAMEO and A-Lab demonstrate closed-loop revision in autonomous science, while FunSearch and AlphaEvolve (Novikov et al., 2025) support partial L3 loops through scoring and regression gates.Source: §9 · p. 52
  • The same L1→L2→L3 progression applies across regimes, but valid rollouts, law violations, and available evidence differ fundamentally by domain.Physical systems use geometric and conservation-law priors, digital systems exploit program semantics, social systems require Theory-of-Mind representations, and scientific systems couple models to experimental evidence.Source: §9 · p. 52
  • The paper proposes decision-centric evaluation, a minimal reproducible package, and an architectural roadmap, with open problems spanning causal representation, law-consistent rollout, and safe autonomous revision.These priorities correspond respectively to L1, L2, and L3 capabilities.Source: §9 · p. 52
  • A central open problem is developing world models that discover and manipulate explicit symbolic governing laws rather than encoding them only implicitly in latent representations.Latent continuous representations support L1 and L2, but symbolic substrates may be better suited to L3 because model structure itself must be revised across physical, digital, social, and scientific worlds.Source: §9 · p. 52

Key Takeaway

  • The future of agentic AI is framed as moving beyond larger predictors toward systems that internalize laws, simulate dynamics, and evolve through active trial-and-error loops to reshape their environments.This progression is intended to support navigation, interpretation, and intervention in the world.Source: Key Takeaway · p. 52

A Philosophical Motivations for Hierarchical World Modeling

This appendix expands the philosophical motivations underlying the L1/L2/L3 hierarchy introduced in Section 2.1.

A Philosophical Motivations for Hierarchical World Modeling

  • The appendix develops the philosophical rationale for the hierarchical L1/L2/L3 world-modeling framework.It specifically extends the motivations presented in Section 2.1.Source: A Philosophical Motivations for… · p. 81

A.1 L1: Inductive Priors and the No-Free-Lunch Theorem … C.2 Software, Web, and Game Systems

The section defines world modeling as action-conditioned prediction of stateful dynamics, organized by capability from inductive one-step prediction through simulation and evidence-driven model revision. It distinguishes models from planners and commonsense, compares the taxonomy with related frameworks, and surveys L2 systems across embodied, driving, game, web, and software domains.

A.1 L1: Inductive Priors and the No-Free-Lunch Theorem

  • Architectures and curricula act as inductive priors because, without structural constraints, no learner is universally superior; L1 systems therefore learn persistence-based next-step regularities from trajectories.Convolutions, equivariances, attention patterns, and JEPA-style targets constrain which regularities are expressible, while Dreamer-style systems operationalize short-horizon latent prediction.Source: §A.1 · p. 81

A.2 L2: Modal Semantics, Epistemic Drift, and Active Inference

  • L2 simulators use intervention-structured rollouts to analyze nearby counterfactual futures, but internally coherent trajectories can drift from reality and become brittle outside the training manifold.MuZero operationalizes this idea by searching learned dynamics with Monte Carlo Tree Search, whereas active inference treats prediction, planning, and exploration as a continuum even though the taxonomy retains an engineering boundary based on rollout reliability.Source: §A.2 · p. 81

A.3 L3: Falsifiability, Paradigm Shifts, and Abduction

  • L3 systems revise model structure when persistent errors implicate the architecture’s hard core rather than merely its learned parameters, using falsifiable probes and targeted diagnosis to guide updates.Abductive systems generally search structured spaces such as program sketches, simulator hooks, and experiment templates, while confirmation holism makes blame assignment across modules non-trivial.Source: §A.3 · p. 82

B.1 World Modeling versus Generic Prediction

  • World models differ from generic prediction by answering action-conditioned transition queries about state evolution, while closed-loop planning additionally requires exploration, reward specification, safety, and search.This distinction separates decision-relevant futures from indiscriminate modeling of observable variation and clarifies why some weather or video systems qualify as L2 models without embedded planners.Source: §B.1 · p. 82

B.2 World Model versus Planner

  • The world model is descriptive and answers forward or counterfactual queries, whereas the planner is normative and selects actions; separating them exposes failure sources and permits reuse across planners and dynamics.L1/L2/L3 classify query depth and reliability rather than whether a planner exists, and forecasting or video models can therefore provide world models without planners.Source: §B.2 · pp. 82–83

B.3 World Modeling versus Commonsense Modeling

  • World modeling predicts evolving states, whereas commonsense supplies persistence and invariants; reliable modeling requires both because fluent rollouts can violate physical constraints and static facts cannot predict trajectories.In L2, constraint consistency makes a testable subset of commonsense measurable against explicit rules or domain simulators, although full commonsense remains open-ended.Source: §B.3 · p. 83

B.4 Taxonomy Comparison

  • The taxonomy is architecture-agnostic and cross-domain, with partial correspondences to Pearl’s causal hierarchy, LeCun’s autonomous-machine architecture, Friston’s active inference, and Moerland et al.’s RL-specific survey, but only L3 explicitly revises the model.These frameworks organize related capabilities differently: association and intervention, reactive behavior and planning, perception and planning, or model-learning integration strategy.Source: §B.4 · p. 83

C Detailed L2 Systems by Different Domains; C.1 Embodied and 3D Systems; C.2 Software, Web, and Game Systems

  • Across detailed L2 systems, representations increasingly shift from passive pixels toward geometry-carrying, action-coupled, and simulation-ready scene states, while games, web, and software systems combine perceptual dynamics with deterministic rules and application state.Embodied methods use occupancy, point or particle, Gaussian, and asset-based representations with distinct tradeoffs in collision reasoning, contact dynamics, scalability, and fidelity; driving systems emphasize coherent counterfactual control, and game systems include examples such as NitroGen’s 52% improvement on unseen games and GameNGen’s 20 FPS interactive DOOM simulation.Source: §C.1 · p. 84, §C.2 · p. 85

C.3 Social Simulation and Multi-Agent Systems … D.1 Dynamics Model Revision from Grasp Failure

The reviewed systems span structured social reasoning, large-scale multi-agent simulation, scientific dynamics modeling, and illustrative L3 model-revision loops. Together, they show how agents can represent social and scientific dynamics while revising models from observed failures.

C.3 Social Simulation and Multi-Agent Systems

  • Social reasoning systems use explicit belief graphs, perspective filtering, recursive Level-K reasoning, or Bayesian-like hypothesis generation, indicating that reasoning structure is a central bottleneck.SymbolicToM supports up to third-order beliefs, while Thought-Tracing significantly outperforms reasoning models such as o3-mini.Source: §C.3 · p. 85
  • Large-scale sandboxes produce specialization, social networks, governance, cultural transmission, and persistent community norms across simulations with up to 10,000+ agents.Project Sid deployed up to 1,000 agents across six towns, while AgentSociety generated 5 million interactions among more than 10,000 agents.Source: §C.3 · p. 85
  • Social dynamics remain model-dependent: only 2 of 15 LLMs achieve sustainable commons-dilemma cooperation, although norms, conventions, group biases, leadership, and deception can emerge.Melting Pot provides more than 50 substrates for evaluating cooperation, competition, deception, and coordination, while Avalon studies emergent leadership and camouflage.Source: §C.3 · pp. 85–86

C.4 AI-for-Science Systems

  • Scientific world models combine interpretable latent dynamics with operator-learning and physics-informed methods to simulate, predict, and discover governing structure in neural and physical systems.DyNeMo models evolving neural network modes and intervention responses, while neural operators provide approximation theory and PINO incorporates PDE residual losses for sparse-data generalization.Source: §C.4 · p. 86
  • Scientific surrogate models extend from molecular properties and thermodynamic sampling to weather and climate forecasting, using architectures including SchNet, Boltzmann Generators, and ClimaX.These systems learn quantum-chemical properties, sample equilibrium states without sequential molecular dynamics, and support forecasting and climate projection after CMIP6 pretraining.Source: §C.4 · p. 87

D Illustrative L3 Evolution Loops

  • The illustrative L3 framework depicts evolution as anomaly detection, attribution, model revision, and validation rather than deployed-system evidence.The appendix explicitly presents these examples as constructive illustrations, not reports of deployed system behavior.Source: D Illustrative L3 Evolution Loops · p. 87

D.1 Dynamics Model Revision from Grasp Failure

  • A grasp-failure loop updates an underestimated friction prior, adds online squeeze diagnostics, tests five material variants, and replays RoboCasa tasks to avoid regressions.The scenario begins with repeated drops, uses force/torque logs and contact timestamps for attribution, and validates that other material classes do not degrade.Source: §D.1 · p. 87

D.2 Strategy Revision in a Service Environment · D.3 Installation Failure Producing Reusable Skill · D.4 Closed-Loop Materials Discovery

Across service, software-installation, and materials-discovery settings, agents revise their models or strategies after diagnosing prediction failures. The examples pair anomaly detection and attribution with targeted revisions, regression tests, and validation through interaction, replay, or new experiments.

D.2 Strategy Revision in a Service Environment

  • 20% of 50 cancellation dialogues retain users after discounts, versus the model’s predicted 70%, exposing a major service-strategy failure.The agent consistently fails to retain users who threaten cancellation despite the societal model predicting discounts would resolve the conversations.Source: §D.2 · p. 87
  • Conversation analysis attributes the retention failure to users’ service-quality frustration rather than price sensitivity, revealing systematically incorrect model beliefs.The social model assumes cancellation threats indicate price sensitivity, but most users are frustrated by service quality.Source: §D.2 · p. 87
  • The revised service strategy distinguishes user intent, escalates quality complaints instead of discounting them, preserves discounts for price-sensitive users, and is tested on 100 interactions.Validation measures retention, satisfaction, and new failures such as unnecessary escalation for genuinely price-sensitive users; a regression test protects the discount behavior.Source: §D.2 · p. 87

D.3 Installation Failure Producing Reusable Skill

  • Permission-denied installation failures are attributed to insufficient privileges rather than typos or network errors, after repeated dependency-installation retries fail.Error codes and permission status identify the failure class.Source: §D.3 · pp. 87–88
  • The agent converts the installation diagnosis into a reusable skill that checks privileges, chooses an alternative install path, re-verifies, and passes restricted-permission regression tests.Replay across OSWorld and macOSWorld variants tests transfer beyond the original failure case (Xie et al., 2024; Yang et al., 2025a).Source: §D.3 · p. 88

D.4 Closed-Loop Materials Discovery

  • An XRD pattern contradicts a Bayesian surrogate’s predicted single crystal phase because the model underweights synthesis temperature’s effect on phase stability.Synthesis instead yields an unexpected mixed phase.Source: §D.4 · p. 88
  • The next synthesis confirms the revised prediction, while calibration improves on held-out compositions after temperature-aware Bayesian updating and an intermediate-temperature experiment.The revision adds the new data point and expands the hypothesis space to include temperature-dependent phase boundaries.Source: §D.4 · p. 88

E Additional Details of Evaluation and Benchmarks … E.3 Constraint Consistency

The appendix operationalizes L2 evaluation through protocols for long-horizon coherence, intervention sensitivity, and constraint consistency, alongside domain-specific benchmarks and the Minimal Reproducible Evaluation Package. It emphasizes trajectory-level diagnostics for compounding error, counterfactual responsiveness, and governing-law violations across physical, digital, social, and scientific settings.

E Additional Details of Evaluation and Benchmarks

  • The appendix details evaluation protocols for three L2 boundary conditions, world-model-specific benchmarks, capability coverage analysis, and the Minimal Reproducible Evaluation Package.These materials extend the evaluation framework introduced in Section 6.Source: E Additional Details of Evaluation and… · p. 88

E.1 Long-Horizon Coherence

  • Long-horizon coherence evaluates whether multi-step predictions remain usable as horizon H grows, with compounding error and state aliasing causing trajectories to diverge from reality.Small per-step deviations can amplify into unreachable branches, while distinct real states may collapse into similar representations.Source: §E.1 · p. 88
  • Coherence should be measured through task-relevant degradation curves across horizons, spanning manipulation success, codebase consistency, social commitments, and causal ordering in laboratory actions.RoboCasa and ManiSkill3 track success-rate decline; SWE-bench tests multi-file resolution; Sotopia tracks relational stability; ScienceWorld tests ordered interventions.Source: §E.1 · p. 88
  • Most benchmarks report fixed-horizon success without characterizing performance versus H, despite open-world tasks such as Minecraft requiring skill composition across sequences exceeding a thousand actions.Crafter offers a more controlled compositional setting, while NetHack adds procedural generation and symbolic complexity.Source: §E.1 · p. 88

E.2 Intervention Sensitivity

  • Intervention sensitivity tests controllability by changing an action or initial condition and measuring whether the resulting rollout changes meaningfully rather than merely appearing realistic.Counterfactual divergence testing compares trajectories from the same z0 after a single-step action perturbation, using action sensitivity ratio and normalized outcome divergence.Source: §E.2 · p. 89
  • Current benchmarks under-test intervention sensitivity because they emphasize output quality; evaluation should explicitly vary actions and measure task-relevant outcome divergence.Examples include replanning after OSWorld interruptions, adapting manipulation to changed object placement in RoboCasa, and shifting negotiation outcomes in Sotopia.Source: §E.2 · p. 89

E.3 Constraint Consistency

  • Constraint consistency checks whether entire rollouts obey regime-specific governing laws, exposing trajectory-level violations, exploitability, and calibration failures that per-step metrics can miss.Verification includes physical compliance and stability, software receipts and contracts, social norms and commitments, scientific conservation and evidence chains, and strict chemical-plan validity.Source: §E.3 · p. 89

E.4 World-Model-Specific Evaluation · E.5 Capability Coverage Matrix

World-model evaluation must assess both perceptual quality and action-conditioned behavior, because strong local diagnostics can coexist with incoherent long-horizon states and weakly generalizable physical reasoning. No single benchmark spans all relevant capability boundaries and governing-law regimes, so evaluation suites should be mapped explicitly to avoid over-claiming generalization.

E.4 World-Model-Specific Evaluation

  • WorldSimBench evaluates video world simulators on both perceptual quality and manipulative capability, exposing sharp divergences between realism and action-conditioned fidelity.Source: §E.4 · p. 89
  • WorldModelBench covers 7 domains, 56 subdomains, and 350 prompts with 67K human labels, finding constraint-consistency gaps across all 14 tested frontier models.Source: §E.4 · p. 89
  • Per-step diagnostic success does not guarantee long-horizon coherence, because models may retain highly incoherent internal world states while performing well on standard tests.Source: §E.4 · p. 89
  • Video generation models can rely on “case-based” mimicry rather than learning generalizable physical principles, including at scale.Source: §E.4 · p. 90

E.5 Capability Coverage Matrix

  • No single benchmark covers all three boundary conditions and four governing-law regimes, limiting broad claims from any individual evaluation suite.Source: §E.5 · p. 90
  • Researchers should map selected evaluation suites to the three boundary conditions and four governing-law regimes to avoid over-claiming generalization.Source: §E.5 · p. 90

E.6 Minimal Reproducible Evaluation Package (MREP)

The Minimal Reproducible Evaluation Package (MREP) addresses non-comparable agent results and leaderboard hacking by standardizing evaluation artifacts and reporting. It connects reproducible evaluation to L3 governed validation, supporting decisions to promote or roll back model updates.

Motivation

  • MREP targets non-comparable agent results and “leaderboard hacking” by establishing a community evaluation standard.The proposal responds to the lack of standardization in agent evaluation (Henderson et al., 2018).Source: §E.6 · p. 90

MREP Components

  • The package requires version locking, replayable trace logs, automated failure taxonomy, tail statistics, and explicit boundary-condition mapping.These components specify exact environment and task commits, intermediate observations/actions/receipts, five L2 failure categories, stratified bootstrap confidence intervals, IQM, performance profiles, and tested boundaries.Source: §E.6 · p. 90

Implementation Status

  • Existing tools support trace logging, version locking, and tail statistics, whereas automated failure taxonomy and boundary-condition mapping require new infrastructure.Trace logging is supported through LangSmith/W&B Weave, version locking through containerization, and tail statistics through Agarwal et al. (2021).Source: §E.6 · p. 90

Governed Validation

  • MREP supplies the evaluation assets needed for L3 governed validation to determine whether model updates should be promoted or rolled back.The framework connects directly to the L3 gatekeeper’s validation decision.Source: §E.6 · p. 90

F Implementation Details and Efficient Deployment

This appendix extends the paper’s discussion of practical implementation considerations and efficiency techniques for world-model systems.

F Implementation Details and Efficient Deployment

F.1 Practical Implementation Considerations … F.4 Memory and KV Cache Compression

Practical world-model deployment requires matching training structure, latency, transfer strategy, and error handling to the environment, while generative dynamics and model compression address the costs of real-time simulation. Long-horizon operation remains constrained by compounding approximation errors, memory growth, and video-specific activation statistics.

F.1 Practical Implementation Considerations

  • End-to-end training shares gradients across representation, dynamics, decoding, and policy, whereas modular training improves debugging and recombination but risks mismatched interfaces.Dreamer (Hafner et al., 2020; 2021; 2025) exemplifies end-to-end learning, while modular systems assign separate objectives to each component; StarVLA (Community, 2026) supports composable ablations.Source: §F.1 · p. 90
  • Deployment latency determines model choice: robotics favors sub-100 ms lightweight latent rollouts, web agents tolerate seconds, and high-fidelity generative models may be restricted to offline planning.Learned forward passes scale as O(1) versus O(N) or worse for explicit simulation, but a large diffusion model’s constant factor can exceed a lightweight physics engine for modest N.Source: §F.1 · pp. 90–91
  • Sim-to-real transfer combines domain randomization, system identification, progressive transfer, or learned residual physics, while reliable systems shorten rollouts or trigger replanning when uncertainty rises.DayDreamer (Wu et al., 2023a) transfers latent imagination to robots using online real sensor data, whereas MBPO (Janner et al., 2019) falls back to real data beyond accurate horizons and ensemble disagreement signals replanning.Source: §F.1 · p. 91

F.2 Few-Step Distillation for Generative Dynamics

  • Few-step distillation converts iterative generative dynamics into faster mappings, and spatiotemporal acceleration makes video generation more viable for real-time planning.Consistency Models learn direct PF-ODE mappings, flow-map and distribution-matching methods provide alternatives, and video acceleration adapts these techniques to costly spatiotemporal sampling.Source: §F.2 · p. 91

F.3 Model Compression: Quantization and Pruning

  • PagedAttention improves large-model serving throughput by managing KV memory in non-contiguous blocks, while diffusion transformers require attention-aware calibration for their distinctive activation distributions.vLLM (Kwon et al., 2023) introduced the virtual-memory-like cache organization; Q-DiT (Chen et al., 2025c), PTQ4DiT (Wu et al., 2024b), SVDQuant (Li et al., 2024b), and ViDiTQ (Zhao et al., 2024) address diffusion-transformer calibration.Source: §F.3 · pp. 91–92
  • Quantization and pruning reduce memory and computation but can induce severe long-horizon semantic drift, motivating rollout-aware pruning that preserves parameters essential for temporal coherence.Diffusion quantization methods such as QDM and TerDiT (Lu et al., 2024b) maintain performance at 1-2 bit precision but require substantial training overhead.Source: §F.3 · pp. 91–92

F.4 Memory and KV Cache Compression

  • Autoregressive world-model rollouts are memory-bound because KV caches grow linearly, requiring token eviction, chunked generation, quantization, or spatiotemporal-aware compression.Chunked video generation is often limited to approximately 60 seconds, and LLM-oriented KV quantization can cause severe quality loss when transferred to video diffusion because activation statistics differ.Source: §F.4 · p. 92

Related papers

Loading 2604.22748v3…