Understanding Rollout Error in Graph World Models
Summary
This paper analyzes long-horizon rollout error in Graph World Models (GWMs), proposing a unified framework with dynamic edges and introducing Error-Aware GWM that uses spectral regularization, rollout consistency, and critical-node weighting to prevent divergence.
View Cached Full Text
Cached at: 06/29/26, 05:27 AM
# Understanding Rollout Error in Graph World Models
Source: [https://arxiv.org/html/2606.27780](https://arxiv.org/html/2606.27780)
Xinyuan Song1Zekun Cai2,3 1Emory University, Atlanta, GA, USA2The University of Tokyo, Tokyo, Japan 3LocationMind, Tokyo, Japan xinyuan\.song@emory\.edu, caizekun@csis\.u\-tokyo\.ac\.jp
###### Abstract
World models are often used for planning by rolling learned dynamics forward\. Many planning environments, however, are not vectors or images; they are graphs of agents, tools, skills, routes, and dependencies\. In these settings, a local prediction error may stay local or spread through the graph, and the failure mode changes again when edges are predicted rather than fixed\. This paper studies long\-horizon rollout error in Graph World Models \(GWMs\)\. We formulate a unified fixed\-edge and dynamic\-edge GWM framework with action nodes for node\-, edge\-, and graph\-level decisions\. We develop graph\-valued rollout bounds that separate topology\-induced amplification from model\-induced amplification, and we introduce a joint node\-edge operator for dynamic\-edge rollouts\. Guided by the analysis, we proposeError\-Aware GWM, which combines spectral regularization, rollout consistency, and critical\-node weighting\. Across synthetic topologies and heterogeneous agent\-graph testbeds, rollout error and planning regret grow with horizon, dynamic\-edge training is needed when structure evolves, and Error\-Aware GWM prevents long\-horizon divergence while preserving prediction accuracy\. Real\-world graph benchmarks clarify the scope of GWMs: they are most useful for dynamic graph rollout and agent planning, while specialized graph models remain strong on static or sparse prediction tasks\. An anonymized implementation and benchmark suite will be released with the submission\. Code is available at[https://github\.com/Hik289/graph\_world\_model\_accumulative\_error\.git](https://github.com/Hik289/graph_world_model_accumulative_error.git)\.
Understanding Rollout Error in Graph World Models
Xinyuan Song1Zekun Cai2,31Emory University, Atlanta, GA, USA2The University of Tokyo, Tokyo, Japan3LocationMind, Tokyo, Japanxinyuan\.song@emory\.edu, caizekun@csis\.u\-tokyo\.ac\.jp
## 1Introduction
A world modelHa and Schmidhuber \([2018](https://arxiv.org/html/2606.27780#bib.bib9)\)represents environment observations as states and predicts future states conditioned on actions\. Modern world models trained on large\-scale dataHafneret al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib20),[2020](https://arxiv.org/html/2606.27780#bib.bib11),[2023](https://arxiv.org/html/2606.27780#bib.bib22)\); Chuaet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib70)\); Wuet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib64)\); Bruceet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib63)\)have shown strong planning, prediction, and generation capabilities\. Most of them, however, are built for vector\- or image\-like states\. They do not directly capture environments whose state is a graph\. Such environments are common in agent planning: multi\-agent calling trees coordinate distributed task executionAnokhinet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib31)\), skill graphs encode tool dependencies, and communication networks route actions among interacting entities\. Planning in these settings requires reasoning about node states and about how effects travel through relations\. The same autoregressive rollout that enables long\-horizon planning also creates a risk: prediction errors can compound and degrade downstream decisionsAsadiet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib1)\); Janneret al\.\([2019](https://arxiv.org/html/2606.27780#bib.bib4)\)\.
Graph World Models \(GWMs\) address this setting by autoregressively predicting graph statesGt=\(V,Et,Xt,At\)G\_\{t\}=\(V,E\_\{t\},X\_\{t\},A\_\{t\}\)Fenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\)\. Given a learned transition modelℱθ\(G^k,ak\)\\mathcal\{F\}\_\{\\theta\}\(\\hat\{G\}\_\{k\},a\_\{k\}\), a GWM rolls outG^1,…,G^H\\hat\{G\}\_\{1\},\\ldots,\\hat\{G\}\_\{H\}and scores candidate action sequences\. Existing work has mostly studied graph foundation models for static graph tasksLiuet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib65)\); Chenet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib66)\)or domain\-specific graph planningZhanget al\.\([2021](https://arxiv.org/html/2606.27780#bib.bib26)\); Ammanabrolu and Riedl \([2021](https://arxiv.org/html/2606.27780#bib.bib33)\)\. Feng et al\.Fenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\)introduced a unified GWM with action nodes, but their setting focuses on fixed\-edge \(FE\) environments\. It remains unclear how GWMs behave under*long\-horizon rollout error*and in*dynamic\-edge \(DE\) regimes*\.
Graph structure makes compounding error harder to analyze than in vector\-state world models\. In vector settings, rollout error is often summarized by a scalar Lipschitz constant\. In graph settings, the same local error can behave very differently depending on topology\. An error on a chain may stay localized; an error on a hub, dense graph, or high\-spectral\-radius topology can spread widely, consistent with spectral analyses of graph propagation, graph\-network dynamics, and GNN stabilityBattagliaet al\.\([2016](https://arxiv.org/html/2606.27780#bib.bib71),[2018](https://arxiv.org/html/2606.27780#bib.bib36)\); Sanchez\-Gonzalezet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib72)\); Oono and Suzuki \([2020](https://arxiv.org/html/2606.27780#bib.bib34)\); Chamberlainet al\.\([2021](https://arxiv.org/html/2606.27780#bib.bib73)\)\. The problem is harder when edges are predicted\. A node\-state error can corrupt future edge prediction, and an edge error can change later message passing\. GWM rollout error is therefore not just scalar prediction error; it is a topology\-dependent, and sometimes coupled, node\-edge dynamical process\.
Existing rollout\-error theory is not designed for graph\-structured planning\. Standard model\-based RL analyses bound compounding error with a scalar Lipschitz constant over a fixed\-dimensional state vectorAsadiet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib1)\); Janneret al\.\([2019](https://arxiv.org/html/2606.27780#bib.bib4)\)\. This misses three aspects of graph rollouts\. First, topology is explicit: for a message\-passing GWM, node\-error growth separates into a topology termρ\(A\)\\rho\(A\)and a model term∏ℓ‖Wℓ‖2\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}, whereAAis the adjacency matrix,ρ\(A\)\\rho\(A\)is its spectral radius, andWℓW\_\{\\ell\}are layer weights\. Second, dynamic graphs require coupled error dynamics: node\-feature errorekX=‖X^k−Xk‖Fe^\{X\}\_\{k\}=\\\|\\hat\{X\}\_\{k\}\-X\_\{k\}\\\|\_\{F\}and edge errorekA=‖A^k−Ak‖Fe^\{A\}\_\{k\}=\\\|\\hat\{A\}\_\{k\}\-A\_\{k\}\\\|\_\{F\}evolve together rather than through a scalar recursion\. Third, rollout error must be tied to planning regret over horizonHH\. These gaps motivate our graph\-valued analysis, which replaces the scalar constant withGEAF^=ρ\(A\)∏ℓ‖Wℓ‖2\\widehat\{\\mathrm\{GEAF\}\}=\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}and uses a joint node\-edge operatorBBfor the recursion of\(ekX,ekA\)\(e^\{X\}\_\{k\},e^\{A\}\_\{k\}\)in dynamic\-edge GWMs\.
We analyze accumulative rollout error in GWMs both theoretically and empirically\. The graph error amplification factorGEAF^=ρ\(A\)∏ℓ‖Wℓ‖2\\widehat\{\\mathrm\{GEAF\}\}=\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}separates topology\-dependent amplification from the model\-dependent spectral norm product\. For dynamic\-edge GWMs, we introduce a joint node\-edge error operatorBBthat captures node\-to\-node, edge\-to\-node, node\-to\-edge, and edge\-to\-edge propagation\. This operator cleanly separates two regimes: fixed\-edge \(FE\) models have no edge\-prediction head, soMX=0M\_\{X\}=0and the dynamics reduce to node\-only prediction; dynamic\-edge \(DE\) models have an edge\-prediction head, soMX\>0M\_\{X\}\>0and node\-edge cross\-coupling becomes active\.
We evaluate the framework on a seven\-topology synthetic benchmark with existing GWM baselines, including GCNKipf and Welling \([2017](https://arxiv.org/html/2606.27780#bib.bib13)\), MPNNGilmeret al\.\([2017](https://arxiv.org/html/2606.27780#bib.bib14)\), GPSRampášeket al\.\([2022](https://arxiv.org/html/2606.27780#bib.bib15)\), and ActionNode GWMFenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\)\. The benchmark includes FE and DE simulators, plus heterogeneous agent\-graph testbeds\. The results support a regime\-dependent view of GWM rollout error\. Long rollouts turn small prediction errors into larger planning errors\. When graph structure evolves, dynamic\-edge training sharply outperforms fixed\-edge training, and the learned dynamics exhibit the node\-edge coupling predicted by the joint operator\. The central empirical point is that GWM stability is not an architecture property alone; it depends on topology, horizon, and rollout regime\.
We further proposeError\-Aware GWMto improve rollout stability on high\-spectral\-radius graphs\. It combines spectral regularization, rollout consistency, and critical\-node weighting to control topology\-induced amplification while preserving prediction accuracy\. Empirically, Error\-Aware GWM removes the divergence observed in vanilla GCN rollouts and keeps long\-horizon prediction error low\. Simpler stabilizers can reduce divergence, but often lose predictive fidelity\. This distinction matters: a useful graph world model must be stable enough to roll out and accurate enough for those rollouts to support planning\.
Our contributions are summarized as follows:
- •Unified GWM framework with FE/DE instantiations \(Section[3](https://arxiv.org/html/2606.27780#S3)\)\.We formulate graph\-structured agent planning as state\-action\-transition modeling and support both fixed\-edge and dynamic\-edge rollouts\.
- •Topology\-aware rollout error theory \(Section[4](https://arxiv.org/html/2606.27780#S4)\)\.We develop graph\-valued error bounds that account for topology\-dependent amplification and node\-edge coupling during autoregressive rollout\.
- •Error\-Aware GWM for stable planning \(Section[6](https://arxiv.org/html/2606.27780#S6)\)\.We introduce a training objective that combines spectral control, rollout consistency, and critical\-node weighting to improve long\-horizon stability\.
- •Evaluation across synthetic and agent\-graph settings \(Section[7](https://arxiv.org/html/2606.27780#S7)\)\.We test the framework across diverse graph topologies, rollout regimes, and agent\-graph environments to study when GWMs are stable or unstable\.
- •Scope analysis on real\-world benchmarks \(Section[7\.7\.1](https://arxiv.org/html/2606.27780#S7.SS7.SSS1)\)\.We evaluate GWMs beyond controlled synthetic dynamics and clarify their practical boundary relative to standard graph methods\.
## 2Related Work
##### World models for planning\.
World models have been widely studied in model\-based reinforcement learning, where a learned dynamics model is used to simulate future trajectories for planning or policy learning\. A central challenge is compounding error: small one\-step model errors can accumulate over long rollouts and degrade decision quality\. Asadi et al\.Asadiet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib1)\)analyze this effect through Lipschitz continuity of learned transition models\. PETS uses probabilistic ensembles and trajectory sampling to propagate uncertainty during model\-based planningChuaet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib70)\)\. Janner et al\.Janneret al\.\([2019](https://arxiv.org/html/2606.27780#bib.bib4)\)study model\-based policy optimization with short branched rollouts to reduce error accumulation\. Kakade and LangfordKakade and Langford \([2002](https://arxiv.org/html/2606.27780#bib.bib62)\)provide simulation\-lemma\-style tools that connect model error to policy performance\. Recent systems reduce rollout drift through retrieval\-augmented contextChenet al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib37)\), hierarchical latent planningZhanget al\.\([2026](https://arxiv.org/html/2606.27780#bib.bib38)\), uncertainty\-based rollout truncationNiet al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib44)\), modular dynamics expertsLiet al\.\([2025a](https://arxiv.org/html/2606.27780#bib.bib43)\), skill\-level abstractionGürtler and Martius \([2025](https://arxiv.org/html/2606.27780#bib.bib39)\), scheduled\-sampling\-style trainingRivardet al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib42)\), and controlled long\-horizon evaluation benchmarksLiet al\.\([2025b](https://arxiv.org/html/2606.27780#bib.bib40)\)\. These works motivate rollout\-error analysis, but they do not model graph topology, message passing, or coupled node\-edge prediction\.
##### Graph neural networks and graph\-structured prediction\.
Graph neural networks provide a standard framework for learning over graph\-structured data by propagating information along edges\. Interaction Networks and Graph Networks introduced object\- and relation\-centric neural dynamics for physical reasoningBattagliaet al\.\([2016](https://arxiv.org/html/2606.27780#bib.bib71),[2018](https://arxiv.org/html/2606.27780#bib.bib36)\), and graph\-network physics engines showed that such models can support inference and control in dynamical systemsSanchez\-Gonzalezet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib72)\)\. Graph convolutional networks aggregate normalized neighborhood informationKipf and Welling \([2017](https://arxiv.org/html/2606.27780#bib.bib13)\)\. Message\-passing neural networks generalize local edge\-based aggregationGilmeret al\.\([2017](https://arxiv.org/html/2606.27780#bib.bib14)\)\. Graph transformers such as GPS combine message passing with global attentionRampášeket al\.\([2022](https://arxiv.org/html/2606.27780#bib.bib15)\)\. Prior theory shows that graph propagation is strongly affected by spectral properties of the adjacency or normalized adjacency matrixOono and Suzuki \([2020](https://arxiv.org/html/2606.27780#bib.bib34)\), while diffusion\-based GNNs connect graph learning to stability of discretized diffusion dynamicsChamberlainet al\.\([2021](https://arxiv.org/html/2606.27780#bib.bib73)\)\. Related work on graph rewiring studies how modifying topology can improve information flow and reduce harmful propagation effectsToriet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib50)\)\. Physics\-informed rewiring further studies topology modification for long\-range interactions in graph neural networksYuet al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib51)\)\. These results suggest that topology can shape prediction behavior, but they mainly study representation learning or one\-step dynamics rather than autoregressive graph rollouts where prediction errors become future inputs\.
##### Graph world models\.
Graph world models extend world modeling to graph\-structured states\. Feng et al\.Fenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\)introduce a GWM framework that represents multimodal observations, graph states, action nodes, and target nodes within a unified graph interface\. WorldGraph uses latent landmarks as a graph\-structured abstraction for planningZhanget al\.\([2021](https://arxiv.org/html/2606.27780#bib.bib26)\)\. Value Memory Graph constructs graph\-structured memory for offline reinforcement learningZhuet al\.\([2022](https://arxiv.org/html/2606.27780#bib.bib27)\)\. Knowledge\-graph world models represent textual environments through structured relational memoryAmmanabrolu and Riedl \([2021](https://arxiv.org/html/2606.27780#bib.bib33)\)\. Contrastive structured world models learn object\-centric relational dynamicsKipfet al\.\([2020](https://arxiv.org/html/2606.27780#bib.bib32)\)\. Dyn\-O builds structured world models with object\-centric representationsWanget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib29)\)\. These works mainly focus on representation, architecture, and task performance\. Our work instead studies long\-horizon error behavior: how topology affects rollout error, how fixed\-edge and dynamic\-edge regimes differ, and how node and edge prediction errors interact through a joint operator\.
##### Agent graphs and skill\-graph systems\.
Multi\-agent and tool\-using systems increasingly organize computation as graphs, including agent calling\-trees, knowledge\-graph memories, tool graphs, and skill graphs\. AriGraph uses graph memory for LLM agentsAnokhinet al\.\([2024](https://arxiv.org/html/2606.27780#bib.bib31)\)\. Audited Skill\-Graph Self\-Improvement studies verifier\-backed skill\-graph maintenanceHuang and Huang \([2025](https://arxiv.org/html/2606.27780#bib.bib55)\)\. CASCADE studies cumulative agentic skill creation through autonomous development and evolutionHuanget al\.\([2026](https://arxiv.org/html/2606.27780#bib.bib56)\)\. Reinforcement learning for self\-improving agents with skill libraries studies policy learning over reusable skillsWanget al\.\([2026](https://arxiv.org/html/2606.27780#bib.bib57)\)\. GraSP compiles retrieved skills into a typed skill graph so failed subgraphs can be repaired locally rather than replanning the full trajectoryXiaet al\.\([2026](https://arxiv.org/html/2606.27780#bib.bib41)\)\. These works show the practical value of graph\-structured agent execution\. Our work is complementary: we analyze how prediction errors accumulate when such graph states are rolled out over time, and we propose Error\-Aware GWM to improve long\-horizon stability in graph\-based planning\.
## 3Graph World Model Framework
### 3\.1Preliminaries: Agent Graph World
AGraph World Model\(GWM\) represents an agent environment as an evolving graph and predicts future graph states for planning\. At timett, the graph state is
Gt=\(V,Et,Xt,At\)\.G\_\{t\}=\(V,E\_\{t\},X\_\{t\},A\_\{t\}\)\.\(3\.1\)Here,VVis the fixed set of agent, tool, skill, or task nodes;EtE\_\{t\}is the edge set;Xt∈ℝN×dX\_\{t\}\\in\\mathbb\{R\}^\{N\\times d\}is the node\-feature matrix; andAt∈\{0,1\}N×NA\_\{t\}\\in\\\{0,1\\\}^\{N\\times N\}is the adjacency matrix\. In an agent system, node features can encode agent status, tool outputs, skill availability, task progress, or failure signals\. Edges represent calling, dependency, routing, or communication relations\.
An actionata\_\{t\}is represented as an action node injected into the graph\. The learned transition model
ℱθ:\(Gt,at\)↦G^t\+1\\mathcal\{F\}\_\{\\theta\}:\(G\_\{t\},a\_\{t\}\)\\mapsto\\hat\{G\}\_\{t\+1\}\(3\.2\)predicts the next graph state\. During planning, an agent evaluates candidate action sequencesa1,…,aHa\_\{1\},\\ldots,a\_\{H\}by rolling out
G^k\+1=ℱθ\(G^k,ak\),k=0,…,H−1,\\hat\{G\}\_\{k\+1\}=\\mathcal\{F\}\_\{\\theta\}\(\\hat\{G\}\_\{k\},a\_\{k\}\),\\qquad k=0,\\ldots,H\-1,\(3\.3\)and scoring the predicted outcomes\. Thus, GWM planning reduces long\-horizon decision\-making to simulation over predicted graph trajectories\.
### 3\.2Action Node Mechanism
Following Feng et al\.Fenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\), actions are encoded as action nodes connected to the current graph\. This gives a unified representation for three kinds of planning actions:node\-level actions, which modify a specific agent, skill, tool, or task node;edge\-level actions, which add, remove, or update dependency or communication edges; andgraph\-level actions, which change global routing, coordination, or allocation patterns\.
We use two action\-node connection types\.Explicit action nodesconnect to target nodes through task\-defined structure, such as a planner calling a tool or assigning a subtask to an executor\.Retrieval\-based action nodesconnect through embedding similarity or retrieval, such as a query routed to related skill nodes or auxiliary tools\. The same transition model can therefore represent explicit decisions and soft retrieval\-based dependencies in one agent graph\.
### 3\.3Two GWM Instantiations
##### Fixed\-Edge GWM \(FE\-GWM\)\.
In the fixed\-edge setting, topology is fixed during rollout and only node states are predicted\. A representative GCN\-style transition is
X^k\+1=tanh\(A~X^kW1\+Uak\)W2\+ξk,\\hat\{X\}\_\{k\+1\}=\\tanh\(\\tilde\{A\}\\hat\{X\}\_\{k\}W\_\{1\}\+Ua\_\{k\}\)W\_\{2\}\+\\xi\_\{k\},\(3\.4\)where
A~=D−1/2AD−1/2\.\\tilde\{A\}=D^\{\-1/2\}AD^\{\-1/2\}\.\(3\.5\)Here,UakUa\_\{k\}injects the action signal andξk\\xi\_\{k\}denotes rollout noise\. This setting covers ActionNode GWMFenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\)and standard GNN\-based world models with fixed agent\-graph structure\. Since edges are not predicted, FE\-GWM propagates node\-state error through a fixed graph, with dynamics governed byρ\(A\)\\rho\(A\)and the model weight norms\.
##### Dynamic\-Edge GWM \(DE\-GWM\)\.
In the dynamic\-edge setting, both node features and graph structure are predicted autoregressively\. In addition to the node transition, DE\-GWM uses an edge\-prediction head such as
A^ij=σ\(hi⊤Qhj\),\\hat\{A\}\_\{ij\}=\\sigma\(h\_\{i\}^\{\\top\}Qh\_\{j\}\),\(3\.6\)wherehi,hjh\_\{i\},h\_\{j\}are node embeddings andQQparameterizes pairwise edge scores\. The model is trained with
ℒ=ℒnode\+λeBCE\(A^,A\)\+12‖W‖2‖Q‖2\.\\mathcal\{L\}=\\mathcal\{L\}\_\{\\rm node\}\+\\lambda\_\{e\}\\mathrm\{BCE\}\(\\hat\{A\},A\)\+\\tfrac\{1\}\{2\}\\\|W\\\|\_\{2\}\\\|Q\\\|\_\{2\}\.\(3\.7\)This setting models changing agent interactions, such as task routing, tool calls, validation links, or recovery paths\. DE\-GWM activates node\-edge cross\-coupling: node errors can affect future edge prediction, and edge errors can change future message passing\. In the joint\-operator view, FE\-GWM is the special case with no edge\-prediction head andMX=0M\_\{X\}=0; DE\-GWM hasMX\>0M\_\{X\}\>0and enables node\-edge feedback\.
## 4Error Amplification Theory
##### Framework overview\.
Figures[1](https://arxiv.org/html/2606.27780#S4.F1)–[3](https://arxiv.org/html/2606.27780#S4.F3)summarize the graph rollout\-error framework\. Figure[1](https://arxiv.org/html/2606.27780#S4.F1)shows topology\-dependent amplification: with the same model and initial error, low\-GEAF graphs keep errors more local, while high\-GEAF hub\-like graphs spread them faster\. Figure[2](https://arxiv.org/html/2606.27780#S4.F2)presents the autoregressive rollout pipeline and the joint node\-edge error operatorBB\. Figure[3](https://arxiv.org/html/2606.27780#S4.F3)contrasts FE and DE regimes, where edge prediction determines whetherBBreduces to node\-only dynamics or activates node\-edge coupling\.
Figure 1:Topology\-controlled rollout error\.The same initial node error can evolve differently across graph topologies: low\-GEAF graphs keep errors localized, while high\-GEAF hub\-like graphs amplify them throughGEAF^=ρ\(A\)∏ℓ‖Wℓ‖2\\widehat\{\\mathrm\{GEAF\}\}=\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\.Figure 2:GWM accumulative\-error framework\.A GWM rolls out graph states autoregressively; one\-step errors accumulate through the joint node\-edge operatorBBand affect long\-horizon planning\.Figure 3:FE versus DE rollout regimes\.FE models keep edges fixed and reduce error propagation to node\-only dynamics, while DE models predict edges and activate the full joint operatorBB\.
### 4\.1Assumptions
We state the assumptions used in the analysis\. They specify the graph\-state domain, the one\-step approximation error, the message\-passing form of node prediction, the Lipschitz structure of dynamic edge prediction, and the reward conditions needed to convert rollout error into planning regret\. These assumptions match our synthetic and heterogeneous agent\-graph simulators: node identities are fixed during rollout, node features and adjacency matrices are bounded, and FE models are recovered when the edge\-update map is absent\.
###### Assumption 1\(Compact graph state space\)\.
At each timett, the graph state isGt=\(Vt,Et,Xt,At\)G\_\{t\}=\(V\_\{t\},E\_\{t\},X\_\{t\},A\_\{t\}\)with fixed\|Vt\|=N\|V\_\{t\}\|=N, node featuresXt∈ℝN×DX\_\{t\}\\in\\mathbb\{R\}^\{N\\times D\}, and continuous adjacency matrixAt∈\[0,1\]N×NA\_\{t\}\\in\[0,1\]^\{N\\times N\}\. There exist constantsRX,RA<∞R\_\{X\},R\_\{A\}<\\inftysuch that
‖Xt‖F≤RX,‖At‖F≤RA\\\|X\_\{t\}\\\|\_\{F\}\\leq R\_\{X\},\\qquad\\\|A\_\{t\}\\\|\_\{F\}\\leq R\_\{A\}\(4\.1\)for all rollout stepst∈\[0,T\]t\\in\[0,T\]\.
Assumption[1](https://arxiv.org/html/2606.27780#Thmassumption1)holds by construction in our simulators\. The main synthetic graphs, agent calling\-tree testbed, and skill\-graph testbed all keep node identities fixed\. Bounded node features follow from bounded nonlinearities or feature clipping, and bounded adjacency follows fromAt∈\[0,1\]N×NA\_\{t\}\\in\[0,1\]^\{N\\times N\}\. For growing graphs, the same analysis can be extended by replacingBBwith a time\-varying operatorBtB\_\{t\}and assumingsuptρ\(Bt\)<∞\\sup\_\{t\}\\rho\(B\_\{t\}\)<\\infty\.
###### Assumption 2\(Uniform one\-step model error\)\.
LetF⋆=\(FX⋆,FA⋆\)F^\{\\star\}=\(F^\{X\\star\},F^\{A\\star\}\)be the true graph transition andFθ=\(FθX,FθA\)F\_\{\\theta\}=\(F\_\{\\theta\}^\{X\},F\_\{\\theta\}^\{A\}\)be the learned GWM transition\. On the compact domain in Assumption[1](https://arxiv.org/html/2606.27780#Thmassumption1), there exist constantsϵX,ϵA≥0\\epsilon\_\{X\},\\epsilon\_\{A\}\\geq 0such that
‖FθX\(X,A,a\)−FX⋆\(X,A,a\)‖F\\displaystyle\\\|F\_\{\\theta\}^\{X\}\(X,A,a\)\-F^\{X\\star\}\(X,A,a\)\\\|\_\{F\}≤ϵX,\\displaystyle\\leq\\epsilon\_\{X\},\(4\.2\)‖FθA\(X,A,a\)−FA⋆\(X,A,a\)‖F\\displaystyle\\\|F\_\{\\theta\}^\{A\}\(X,A,a\)\-F^\{A\\star\}\(X,A,a\)\\\|\_\{F\}≤ϵA\.\\displaystyle\\leq\\epsilon\_\{A\}\.
Assumption[2](https://arxiv.org/html/2606.27780#Thmassumption2)is the graph analogue of the standard one\-step model\-error assumption in model\-based reinforcement learning\. The constantsϵX\\epsilon\_\{X\}andϵA\\epsilon\_\{A\}are worst\-case one\-step node and edge prediction errors\. If only expected one\-step error is available, the same recursion gives an expected rollout\-error bound with the same operator form\.
###### Assumption 3\(Message\-passing node transition\)\.
The learned node update has the message\-passing form
FθX\(X,A,a\)=σ\(A^XW1\+Ua𝟏⊤\)W2,F\_\{\\theta\}^\{X\}\(X,A,a\)=\\sigma\(\\hat\{A\}XW\_\{1\}\+Ua\\mathbf\{1\}^\{\\top\}\)W\_\{2\},\(4\.3\)whereσ\\sigmais coordinate\-wise Lipschitz with constantLσ≤1L\_\{\\sigma\}\\leq 1,A^\\hat\{A\}is the propagation matrix induced byAA,W1,W2W\_\{1\},W\_\{2\}are learnable weight matrices, andUUmaps the action embedding into the node\-update space\.
Assumption[3](https://arxiv.org/html/2606.27780#Thmassumption3)covers standard GNN\-style node predictors, including GCN, MPNN\-style message passing, graph transformers, and ActionNode GWM variants after choosing the propagation matrixA^\\hat\{A\}\. Multi\-layer models are handled by composing layer\-wise Lipschitz constants, which gives the weight product∏ℓ‖Wℓ‖2\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\. Non\-graph MLP world models are the degenerate case with no topology\-dependent propagation\.
###### Assumption 4\(Lipschitz dynamic edge update\)\.
For DE models, the learned edge update is Lipschitz in both node features and the current adjacency\. In particular, there exist constantsMX,MA,LA<∞M\_\{X\},M\_\{A\},L\_\{A\}<\\inftysuch that node and edge errors satisfy the coupled bounds
ek\+1X\\displaystyle e^\{X\}\_\{k\+1\}≤LXekX\+LAekA\+ϵX,\\displaystyle\\leq L\_\{X\}e^\{X\}\_\{k\}\+L\_\{A\}e^\{A\}\_\{k\}\+\\epsilon\_\{X\},\(4\.4\)ek\+1A\\displaystyle e^\{A\}\_\{k\+1\}≤MXekX\+MAekA\+ϵA\.\\displaystyle\\leq M\_\{X\}e^\{X\}\_\{k\}\+M\_\{A\}e^\{A\}\_\{k\}\+\\epsilon\_\{A\}\.
Assumption[4](https://arxiv.org/html/2606.27780#Thmassumption4)captures differentiable or relaxed differentiable edge predictors, such as a bilinear edge head followed by a bounded projection\. In FE models, there is no edge\-prediction head, soMX=0M\_\{X\}=0andMA=0M\_\{A\}=0\. The joint node\-edge operator then reduces to fixed\-edge node\-error recursion, which is the main regime distinction in our theory\.
###### Assumption 5\(Lipschitz reward and bounded reward approximation error\)\.
The true rewardR⋆\(G,a\)R^\{\\star\}\(G,a\)is Lipschitz under a graph distancedGd\_\{G\}: for all graph statesG,G′G,G^\{\\prime\}and actionsaa,
\|R⋆\(G,a\)−R⋆\(G′,a\)\|≤LRdG\(G,G′\)\.\|R^\{\\star\}\(G,a\)\-R^\{\\star\}\(G^\{\\prime\},a\)\|\\leq L\_\{R\}d\_\{G\}\(G,G^\{\\prime\}\)\.\(4\.5\)The learned reward modelRθR\_\{\\theta\}has uniformly bounded error
\|Rθ\(G,a\)−R⋆\(G,a\)\|≤ϵR\.\|R\_\{\\theta\}\(G,a\)\-R^\{\\star\}\(G,a\)\|\\leq\\epsilon\_\{R\}\.\(4\.6\)
Assumption[5](https://arxiv.org/html/2606.27780#Thmassumption5)is used only to connect rollout error to planning regret\. For sparse or discontinuous task rewards, the bound can be applied to a smoothed or shaped reward, with the smoothing error included in the final regret bound\.
###### Assumption 6\(Discounted finite\-horizon return\)\.
Planning uses the finite\-horizon discounted return
J⋆=∑k=0H−1γkrt\+k,γ∈\[0,1\]\.J^\{\\star\}=\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}r\_\{t\+k\},\\qquad\\gamma\\in\[0,1\]\.\(4\.7\)
Assumption[6](https://arxiv.org/html/2606.27780#Thmassumption6)follows the standard finite\-horizon planning setup\. The caseγ=1\\gamma=1is allowed for finiteHH\. For infinite\-horizon analysis, one needsγ<1\\gamma<1and a stable enough error operator; otherwise, the regret bound correctly diverges when rollout errors are amplified over time\.
###### Assumption 7\(Linearized graph distance\)\.
The graph distance used in the regret analysis is bounded by a linear combination of node and edge errors:
dG\(Gk,G^k\)≤α‖Xk−X^k‖F\+β‖Ak−A^k‖F,d\_\{G\}\(G\_\{k\},\\hat\{G\}\_\{k\}\)\\leq\\alpha\\\|X\_\{k\}\-\\hat\{X\}\_\{k\}\\\|\_\{F\}\+\\beta\\\|A\_\{k\}\-\\hat\{A\}\_\{k\}\\\|\_\{F\},\(4\.8\)whereα,β≥0\\alpha,\\beta\\geq 0\.
Assumption[7](https://arxiv.org/html/2606.27780#Thmassumption7)translates the vector error recursion into a reward\-error bound\. It is natural when node identities are fixed across rollout steps, as in our simulators\. Spectral adjacency distances and edge\-set distances can be upper\-bounded by Frobenius\-norm edge error, so their effects can be absorbed intoβ\\beta\.
Together, Assumptions[1](https://arxiv.org/html/2606.27780#Thmassumption1)–[7](https://arxiv.org/html/2606.27780#Thmassumption7)define the scope of the theory\. Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1)uses compact state, one\-step error, and message passing\. Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2)adds the dynamic\-edge Lipschitz condition\. Lemma[3](https://arxiv.org/html/2606.27780#Thmtheorem3)uses the same nonnegative operator structure with spectral norm bounds\. Lemma[4](https://arxiv.org/html/2606.27780#Thmtheorem4)uses the reward and graph\-distance assumptions to convert rollout error into planning regret\.
### 4\.2Problem Formulation
We study how prediction errors accumulate when a GWM is rolled out for planning\. The graph stateGtG\_\{t\}and transition modelℱθ\\mathcal\{F\}\_\{\\theta\}are defined in Equations \([3\.1](https://arxiv.org/html/2606.27780#S3.E1)\)–\([3\.2](https://arxiv.org/html/2606.27780#S3.E2)\)\. Starting from an initial predicted graphG^0\\hat\{G\}\_\{0\}, the model generates anHH\-step predicted trajectory according to Equation \([3\.3](https://arxiv.org/html/2606.27780#S3.E3)\)\. Because each predicted state becomes the input to the next step, one\-step errors can accumulate and affect downstream decisions\.
Let\(Xk,Ak\)\(X\_\{k\},A\_\{k\}\)denote the true node features and adjacency at rollout stepkk, and let\(X^k,A^k\)\(\\hat\{X\}\_\{k\},\\hat\{A\}\_\{k\}\)denote their GWM predictions\. We measure node and edge rollout errors by
ekX=‖X^k−Xk‖F,ekA=‖A^k−Ak‖F\.e^\{X\}\_\{k\}=\\\|\\hat\{X\}\_\{k\}\-X\_\{k\}\\\|\_\{F\},\\qquad e^\{A\}\_\{k\}=\\\|\\hat\{A\}\_\{k\}\-A\_\{k\}\\\|\_\{F\}\.\(4\.9\)We collect them into the joint error vector
uk=\(ekX,ekA\)⊤\.u\_\{k\}=\(e^\{X\}\_\{k\},e^\{A\}\_\{k\}\)^\{\\top\}\.\(4\.10\)The central question is howuku\_\{k\}evolves with horizon, topology, and rollout regime\. In fixed\-edge \(FE\) GWMs, the adjacency is fixed and only node states are predicted\. In dynamic\-edge \(DE\) GWMs, both node states and edges are predicted, so node and edge errors can influence each other\.
Our analysis uses the assumptions stated in Section[4\.1](https://arxiv.org/html/2606.27780#S4.SS1): compact graph states, bounded one\-step node and edge prediction errors, a message\-passing node transition, a Lipschitz dynamic edge update for DE models, and a Lipschitz reward used to translate rollout error into planning regret\.
### 4\.3Fixed\-Edge Rollout Error
We first analyze the fixed\-edge \(FE\) regime, where topology is held fixed and only node features are predicted\. Rollout error then propagates through a fixed message\-passing operator\. The bound separates the topology\-dependent part of amplification from the model\-dependent part\.
###### Lemma 1\(Fixed\-edge node error bound\)\.
Assume fixed adjacencyAt≡AA\_\{t\}\\equiv Aand a message\-passing node transition with activation Lipschitz constantLσL\_\{\\sigma\}\. Suppose the one\-step node prediction error is bounded byϵX\\epsilon\_\{X\}\. Then, for any rollout stepk≥0k\\geq 0,
ekX\\displaystyle e^\{X\}\_\{k\}≤LXke0X\+ϵXLXk−1LX−1,\\displaystyle\\leq L\_\{X\}^\{k\}e^\{X\}\_\{0\}\+\\epsilon\_\{X\}\\frac\{L\_\{X\}^\{k\}\-1\}\{L\_\{X\}\-1\},\(4\.11\)LX\\displaystyle L\_\{X\}=Lσρ\(A\)⏟topology∏ℓ‖Wℓ‖2⏟model\.\\displaystyle=L\_\{\\sigma\}\\underbrace\{\\rho\(A\)\}\_\{\\text\{topology\}\}\\underbrace\{\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\}\_\{\\text\{model\}\}\.
Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1)shows that FE\-GWM rollout error is controlled by two factors\. The topology factorρ\(A\)\\rho\(A\)determines how strongly errors can spread through the graph\. The model factor∏ℓ‖Wℓ‖2\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}captures layer\-wise amplification induced by the learned transition model\. This motivates theGraph Error Amplification Factor
GEAF^\(G\)=ρ\(A\)∏ℓ‖Wℓ‖2\.\\widehat\{\\mathrm\{GEAF\}\}\(G\)=\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\.\(4\.12\)Thus, with the same model weights, different graph topologies can induce different long\-horizon error growth rates\. The proof of Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1)is provided in Appendix[A\.1](https://arxiv.org/html/2606.27780#A1.SS1)\.
### 4\.4Dynamic\-Edge Rollout Error
We next analyze the dynamic\-edge \(DE\) regime, where the GWM predicts both node states and graph structure\. Here, rollout errors are coupled: node\-feature errors can affect future edge prediction, and edge errors can change future message passing\. We capture this interaction with a joint node\-edge error operator\.
###### Lemma 2\(Joint node\-edge error bound\)\.
Assume a DE GWM with bounded one\-step errorsϵX\\epsilon\_\{X\}andϵA\\epsilon\_\{A\}\. Suppose that the node and edge errors satisfy the coupled Lipschitz bounds
ek\+1X\\displaystyle e^\{X\}\_\{k\+1\}≤LXekX\+LAekA\+ϵX,\\displaystyle\\leq L\_\{X\}e^\{X\}\_\{k\}\+L\_\{A\}e^\{A\}\_\{k\}\+\\epsilon\_\{X\},\(4\.13\)ek\+1A\\displaystyle e^\{A\}\_\{k\+1\}≤MXekX\+MAekA\+ϵA\.\\displaystyle\\leq M\_\{X\}e^\{X\}\_\{k\}\+M\_\{A\}e^\{A\}\_\{k\}\+\\epsilon\_\{A\}\.Letuk=\(ekX,ekA\)⊤u\_\{k\}=\(e^\{X\}\_\{k\},e^\{A\}\_\{k\}\)^\{\\top\}\. Then
uk\+1\\displaystyle u\_\{k\+1\}⪯Buk\+ϵ,\\displaystyle\\preceq Bu\_\{k\}\+\\epsilon,\(4\.14\)B\\displaystyle B=\(LXLAMXMA\),\\displaystyle=\\begin\{pmatrix\}L\_\{X\}&L\_\{A\}\\\\ M\_\{X\}&M\_\{A\}\\end\{pmatrix\},ϵ\\displaystyle\\epsilon=\(ϵX,ϵA\)⊤\.\\displaystyle=\(\\epsilon\_\{X\},\\epsilon\_\{A\}\)^\{\\top\}\.The spectral radius ofBBis
ρ\(B\)\\displaystyle\\rho\(B\)=12\[\(LX\+MA\)\\displaystyle=\\frac\{1\}\{2\}\\Big\[\(L\_\{X\}\+M\_\{A\}\)\(4\.15\)\+\(LX−MA\)2\+4LAMX\]\.\\displaystyle\\quad\+\\sqrt\{\(L\_\{X\}\-M\_\{A\}\)^\{2\}\+4L\_\{A\}M\_\{X\}\}\\Big\]\.IfLAMX\>0L\_\{A\}M\_\{X\}\>0, then
ρ\(B\)\>max\(LX,MA\)\.\\rho\(B\)\>\\max\(L\_\{X\},M\_\{A\}\)\.\(4\.16\)
Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2)shows that DE\-GWM rollout error is governed by a coupled operator rather than a scalar node\-error recursion\. The off\-diagonal termsLAL\_\{A\}andMXM\_\{X\}encode edge\-to\-node and node\-to\-edge error propagation\. In FE models, there is no edge\-prediction head, soMX=0M\_\{X\}=0and the recursion reduces to node\-only prediction\. In DE models,MX\>0M\_\{X\}\>0activates node\-edge feedback, and Equation \([4\.16](https://arxiv.org/html/2606.27780#S4.E16)\) shows that the coupled spectral radius can exceed both uncoupled components\. The proof is provided in Appendix[A\.2](https://arxiv.org/html/2606.27780#A1.SS2)\.
### 4\.5GEAF as a Proxy for the Joint Operator
The joint operatorBBcaptures coupled node\-edge error propagation, but estimating all of its entries can be difficult\. We therefore useGEAF^\\widehat\{\\mathrm\{GEAF\}\}as a topology\-aware proxy for the dominant rollout amplification scale\.
###### Lemma 3\(GEAF proxy bound\)\.
Assume the signal\-dominant regime‖A‖2≥RX\\\|A\\\|\_\{2\}\\geq R\_\{X\}, whereRXR\_\{X\}bounds the non\-topological contribution to the first\-row error term\. Then
ρ\(B\)\\displaystyle\\rho\(B\)≤GEAF^\(G\)\(1\+RX‖A‖2\)\\displaystyle\\leq\\widehat\{\\mathrm\{GEAF\}\}\(G\)\\left\(1\+\\frac\{R\_\{X\}\}\{\\\|A\\\|\_\{2\}\}\\right\)\(4\.17\)≤2GEAF^\(G\)\.\\displaystyle\\leq 2\\widehat\{\\mathrm\{GEAF\}\}\(G\)\.Moreover, when model weights are fixed across topologies, the rank order ofGEAF^\(G\)\\widehat\{\\mathrm\{GEAF\}\}\(G\)follows the rank order ofρ\(A\)\\rho\(A\)\.
Lemma[3](https://arxiv.org/html/2606.27780#Thmtheorem3)shows thatGEAF^\\widehat\{\\mathrm\{GEAF\}\}upper\-bounds the joint operator up to a constant factor in the signal\-dominant regime\. In FE models,MX=0M\_\{X\}=0, soρ\(B\)=LX=GEAF^\\rho\(B\)=L\_\{X\}=\\widehat\{\\mathrm\{GEAF\}\}by construction; this is an identity rather than empirical validation\. In DE models,MX\>0M\_\{X\}\>0breaks this identity, so comparingρ\(B\)\\rho\(B\)andGEAF^\\widehat\{\\mathrm\{GEAF\}\}becomes meaningful\. The proof is provided in Appendix[A\.3](https://arxiv.org/html/2606.27780#A1.SS3)\.
### 4\.6Planning Regret Under Graph Rollout Error
Rollout error affects planning because actions are selected from predicted graph trajectories\. We connect the joint node\-edge error recursion to planning regret through a graph\-state simulation bound\.
###### Lemma 4\(Planning regret bound\)\.
Assume the reward is Lipschitz with constantLRL\_\{R\}under a graph distance
dG\(G^k,Gk\)=c⊤uk\.d\_\{G\}\(\\hat\{G\}\_\{k\},G\_\{k\}\)=c^\{\\top\}u\_\{k\}\.\(4\.18\)Letπ^θ\\hat\{\\pi\}\_\{\\theta\}be the policy selected by planning under the learned GWM, and letπ∗\\pi^\{\*\}be the policy selected under the true dynamics\. Then
J∗\(π∗\)−J∗\(π^θ\)\\displaystyle J^\{\*\}\(\\pi^\{\*\}\)\-J^\{\*\}\(\\hat\{\\pi\}\_\{\\theta\}\)≤2LR‖c‖2κ\(B\)ϵ~\\displaystyle\\leq 2L\_\{R\}\\\|c\\\|\_\{2\}\\kappa\(B\)\\tilde\{\\epsilon\}\(4\.19\)×ΦH\(γ,ρ\(B\)\)\\displaystyle\\quad\\times\\Phi\_\{H\}\(\\gamma,\\rho\(B\)\)\+2ϵR1−γH1−γ,\\displaystyle\\quad\+2\\epsilon\_\{R\}\\frac\{1\-\\gamma^\{H\}\}\{1\-\\gamma\},where
ΦH\(γ,ρ\)=1ρ−1\[1−\(γρ\)H1−γρ−1−γH1−γ\]\.\\Phi\_\{H\}\(\\gamma,\\rho\)=\\frac\{1\}\{\\rho\-1\}\\left\[\\frac\{1\-\(\\gamma\\rho\)^\{H\}\}\{1\-\\gamma\\rho\}\-\\frac\{1\-\\gamma^\{H\}\}\{1\-\\gamma\}\\right\]\.\(4\.20\)Ifγρ\(B\)\>1\\gamma\\rho\(B\)\>1, the rollout\-error contribution grows super\-linearly with horizonHH\.
Lemma[4](https://arxiv.org/html/2606.27780#Thmtheorem4)shows that small one\-step prediction errors can still cause large planning error when the graph rollout operator is expansive\. The factorρ\(B\)\\rho\(B\)controls how node and edge errors accumulate, whileγ\\gammacontrols how future errors are weighted in planning\. The proof is provided in Appendix[A\.4](https://arxiv.org/html/2606.27780#A1.SS4)\.
###### Corollary 5\(Error level versus error growth slope\)\.
Under the FE recursion in Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1), when the rollout is not dominated by the one\-step error floor, the asymptotic log\-error growth rate satisfies
∂logekX∂k→logLX\.\\frac\{\\partial\\log e^\{X\}\_\{k\}\}\{\\partial k\}\\to\\log L\_\{X\}\.\(4\.21\)Thus, the slope of the log\-error curve can be a more stable indicator of topology\-dependent amplification than the final error level\.
Corollary[5](https://arxiv.org/html/2606.27780#Thmtheorem5)explains why level\-based metrics can be weak even when topology affects rollout dynamics\. In practice, trained models may reach an error floor of orderϵX/\(1−LX\)\\epsilon\_\{X\}/\(1\-L\_\{X\}\), which can mask the relation betweenGEAF^\\widehat\{\\mathrm\{GEAF\}\}and final NodeMSE\. Therefore, we report GrowthSlope as an exploratory diagnostic rather than as a pre\-registered confirmatory claim\. The proof is provided in Appendix[A\.5](https://arxiv.org/html/2606.27780#A1.SS5)\.
## 5Dataset and Evaluation Protocol
### 5\.1Graph Topology Generation
We generate seven canonical graph families with distinct spectral and structural properties atN=50N=50nodes, using three outer seeds for each topology\. The topology statistics are summarized in Table[1](https://arxiv.org/html/2606.27780#S5.T1)\.
Table 1:Topology statistics for the seven synthetic graph families atN=50N=50\. The table shows the spectral range used to stress\-test GWM rollouts:ρ\(A\)\\rho\(A\)is the adjacency spectral radius,GEAF^\\widehat\{\\mathrm\{GEAF\}\}is the mean graph error amplification factor across successful baseline runs, NodeMSE@32 is the median 32\-step prediction error, GrowthSlope is the median log\-error slope from horizon 4 to horizon 32, and Diverged counts rollout explosions\.The seven topology families span a broad spectral range, fromρ\(A\)=2\.0\\rho\(A\)=2\.0for chain graphs toρ\(A\)=49\\rho\(A\)=49for complete graphs\. They cover three structural regimes: low\-connectivity or tree\-like graphs \(chain, tree\), lattice or clustered graphs \(grid, small\-world\), and hub\-dominated graphs \(scale\-free, star, complete\)\. Scale\-free graphs are generated by preferential attachment withm=2m=2edges per new nodeBarabási and Albert \([1999](https://arxiv.org/html/2606.27780#bib.bib2)\)\. Small\-world graphs use the Watts–Strogatz model withk=4k=4and rewiring probabilityp=0\.3p=0\.3Watts and Strogatz \([1998](https://arxiv.org/html/2606.27780#bib.bib3)\)\. All other topologies are deterministic for fixedNN\. Each instance storesAA, initial node featuresX0∈ℝN×8X\_\{0\}\\in\\mathbb\{R\}^\{N\\times 8\}, and the rollout trajectory\{Xt,At\}t=1T\\\{X\_\{t\},A\_\{t\}\\\}\_\{t=1\}^\{T\}\.
### 5\.2Simulators
##### Fixed\-Edge \(FE\) simulator\.
In the FE simulator, the adjacency matrixAAis fixed throughout rollout, matching Section[4\.3](https://arxiv.org/html/2606.27780#S4.SS3)\. Node features evolve according to
Xt=ftrue\(Xt−1,A\)\+εt,εt∼𝒩\(0,σ2I\)\.X\_\{t\}=f\_\{\\rm true\}\(X\_\{t\-1\},A\)\+\\varepsilon\_\{t\},\\qquad\\varepsilon\_\{t\}\\sim\\mathcal\{N\}\(0,\\sigma^\{2\}I\)\.\(5\.1\)Hereσ=0\.01\\sigma=0\.01\. Node featuresX∈ℝN×8X\\in\\mathbb\{R\}^\{N\\times 8\}encode state variables such as congestion or utilization\. Rollout follows Equation \([3\.3](https://arxiv.org/html/2606.27780#S3.E3)\)\. The rollout horizon isT=50T=50, and models are evaluated atH∈\{1,2,4,8,16,32\}H\\in\\\{1,2,4,8,16,32\\\}\.
##### Dynamic\-Edge \(DE\) simulator\.
In the DE simulator, both node features and edges evolve during rollout, matching Section[4\.4](https://arxiv.org/html/2606.27780#S4.SS4)\. The edge head follows the dynamic\-edge prediction form in Section[3\.3](https://arxiv.org/html/2606.27780#S3.SS3):
A^t=σ\(HtQHt⊤\)\.\\hat\{A\}\_\{t\}=\\sigma\(H\_\{t\}QH\_\{t\}^\{\\top\}\)\.\(5\.2\)We use‖Q‖2=1\.0\\\|Q\\\|\_\{2\}=1\.0by default\. The joint node\-edge error follows the operator recursion in Equation \([4\.14](https://arxiv.org/html/2606.27780#S4.E14)\), with the spectral radius given by Equation \([4\.15](https://arxiv.org/html/2606.27780#S4.E15)\)\. Empirically, this coupling is activated in the DE regime:ρ\(B\)=132\\rho\(B\)=132–368368, compared with FE values in the range22–5050\. The rollout horizon isT=32T=32\.
##### Agent calling\-tree testbed\.
The agent calling\-tree testbed evaluates GWM rollouts on a heterogeneous task\-execution graph\. Each instance is a directed acyclic graph withN≈22N\\approx 22–3030nodes, 9 node types, and 6 edge types\. Node types includeplanner,validator,executor,checker,aggregator,reporter,logger,error\_handler, andfinal\_answer\. Edge types includecalls,validates,reports,errors,triggers, andlogs\. Each node has an 8\-dimensional execution\-status vector\. The primary metric is sink\-node success rate,srsink\\mathrm\{sr\}\_\{\\rm sink\}, which measures whether the final\-answer node reaches a successful terminal state\. As calibration, a random policy givessrsink=0\.262\\mathrm\{sr\}\_\{\\rm sink\}=0\.262, while an oracle policy exceeds0\.400\.40\.
##### Platform skill\-graph testbed\.
The platform skill\-graph testbed evaluates GWM rollouts on a heterogeneous skill\-management graph\. Each instance containsN=40N=40nodes across 8 node types:skill,validator,adapter,artifact,failure\_case,patch,task, andtool\. The graph models dependencies among skills, validators, adapters, artifacts, and failure\-recovery components\. The primary metric is skill success rate,srskill\\mathrm\{sr\}\_\{\\rm skill\}\. Under a no\-op policy,srskill\\mathrm\{sr\}\_\{\\rm skill\}decays from1\.01\.0to about0\.3170\.317atT=10T=10, showing that skill quality degrades without maintenance actions\. R\-GCN calibration against the ground\-truth simulator gives Pearsonr=0\.231r=0\.231and MAE=0\.029=0\.029\.
## 6Error\-Aware GWM
##### Graph world model setup\.
We study GWMs that predict graph\-structured states using the transition model in Equation \([3\.2](https://arxiv.org/html/2606.27780#S3.E2)\) and the rollout procedure in Equation \([3\.3](https://arxiv.org/html/2606.27780#S3.E3)\)\. In the fixed\-edge setting, the node\-state update follows the GCN\-style transition:
X^k\+1\\displaystyle\\hat\{X\}\_\{k\+1\}=tanh\(A~X^kW1\+Uak\)W2\+ξk,\\displaystyle=\\tanh\(\\tilde\{A\}\\hat\{X\}\_\{k\}W\_\{1\}\+Ua\_\{k\}\)W\_\{2\}\+\\xi\_\{k\},\(6\.1\)A~\\displaystyle\\tilde\{A\}=D−1/2AD−1/2,\\displaystyle=D^\{\-1/2\}AD^\{\-1/2\},whereA~\\tilde\{A\}is the normalized adjacency matrix andξk\\xi\_\{k\}is i\.i\.d\. rollout noise\. This update follows the ActionNode\-style GWM formulation of Feng et al\.Fenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\), where actions are injected into the graph state during prediction\.
##### Error\-aware training objective\.
Our Error\-Aware GWM augments the standard prediction objective with three stability terms:
ℒEA=ℒpred\+λspecRspec\+λrollRroll\+λcritRcrit\.\\mathcal\{L\}\_\{\\rm EA\}=\\mathcal\{L\}\_\{\\rm pred\}\+\\lambda\_\{\\rm spec\}R\_\{\\rm spec\}\+\\lambda\_\{\\rm roll\}R\_\{\\rm roll\}\+\\lambda\_\{\\rm crit\}R\_\{\\rm crit\}\.\(6\.2\)The first term,ℒpred\\mathcal\{L\}\_\{\\rm pred\}, is the supervised one\-step prediction loss for node states and, in the dynamic\-edge setting, edge states\. The remaining terms reduce long\-horizon rollout instability\.
##### Spectral regularization\.
The spectral regularizer controls the model\-dependent part of graph error amplification\. Motivated by the FE bound in Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1), we penalize large transition\-layer spectral norms:
Rspec=∑ℓ‖Wℓ‖22\.R\_\{\\rm spec\}=\\sum\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}^\{2\}\.\(6\.3\)This term discourages large values of∏ℓ‖Wℓ‖2\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}, reducing the model factor inGEAF^\(G\)=ρ\(A\)∏ℓ‖Wℓ‖2\\widehat\{\\mathrm\{GEAF\}\}\(G\)=\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\. It is especially important on high\-ρ\(A\)\\rho\(A\)graphs, where topology already amplifies rollout errors\.
##### Rollout consistency regularization\.
The rollout consistency term directly penalizes multi\-step prediction drift\. Given a rollout horizon setℋ\\mathcal\{H\}, we define
Rroll=∑h∈ℋ‖X^t\+h\(h\)−Xt\+h‖F2,R\_\{\\rm roll\}=\\sum\_\{h\\in\\mathcal\{H\}\}\\\|\\hat\{X\}\_\{t\+h\}^\{\(h\)\}\-X\_\{t\+h\}\\\|\_\{F\}^\{2\},\(6\.4\)whereX^t\+h\(h\)\\hat\{X\}\_\{t\+h\}^\{\(h\)\}is thehh\-step autoregressive prediction produced by repeatedly applying the learned transition model\. Unlike one\-step training, this term exposes the model to its own rollout distribution and reduces compounding error over long horizons\.
##### Critical\-node weighting\.
The critical\-node term gives larger weight to errors on structurally important nodes\. Letwvw\_\{v\}be a normalized centrality\-based weight for nodevv, such as degree, PageRank, or betweenness centrality\. We define
Rcrit=∑v∈Vwv‖X^t\+1,v−Xt\+1,v‖22\.R\_\{\\rm crit\}=\\sum\_\{v\\in V\}w\_\{v\}\\\|\\hat\{X\}\_\{t\+1,v\}\-X\_\{t\+1,v\}\\\|\_\{2\}^\{2\}\.\(6\.5\)This term reflects that errors on hubs or high\-centrality nodes can propagate to many downstream nodes during message passing\. By emphasizing critical nodes, Error\-Aware GWM reduces the risk that local errors trigger global rollout instability\.
##### Dynamic\-edge extension\.
For dynamic\-edge models, we add the bilinear edge\-prediction head from Equation \([5\.2](https://arxiv.org/html/2606.27780#S5.E2)\)\. The edge loss is binary cross\-entropy with weightλe=1\.0\\lambda\_\{e\}=1\.0, and the graph regularization term is
Lg=12‖W‖2‖Q‖2\.L\_\{g\}=\\frac\{1\}\{2\}\\\|W\\\|\_\{2\}\\\|Q\\\|\_\{2\}\.\(6\.6\)This term controls the interaction between node\-transition weightsWWand edge\-prediction weightsQQ, which is important because DE rollouts activate the coupled node\-edge operator in Equation \([4\.14](https://arxiv.org/html/2606.27780#S4.E14)\)\.
## 7Experiments
### 7\.1Experimental Setup
##### Baselines\.
We compare six world\-model baselines\. B1 is an MLP world model \(MLP\-WM\) that does not use graph structure\. B2 is a vanilla GCN world model based on graph convolutionKipf and Welling \([2017](https://arxiv.org/html/2606.27780#bib.bib13)\)\. B3 is an MPNN world model based on neural message passingGilmeret al\.\([2017](https://arxiv.org/html/2606.27780#bib.bib14)\)\. B4 is a GPS\-style graph transformerRampášeket al\.\([2022](https://arxiv.org/html/2606.27780#bib.bib15)\)\. B5 is an ActionNode GWM based on the fixed\-edge architecture of Feng et al\.Fenget al\.\([2025](https://arxiv.org/html/2606.27780#bib.bib24)\)\. B6 is our Error\-Aware GWM\.
##### Topologies\.
We evaluate all baselines on seven synthetic graph families: chain, tree, grid, small\-world, scale\-free, star, and complete graphs\. Each graph hasN=50N=50nodes\. Scale\-free graphs follow the preferential\-attachment modelBarabási and Albert \([1999](https://arxiv.org/html/2606.27780#bib.bib2)\), and small\-world graphs follow the Watts–Strogatz modelWatts and Strogatz \([1998](https://arxiv.org/html/2606.27780#bib.bib3)\)\. These topologies span a broad spectral range, fromρ\(A\)=2\.0\\rho\(A\)=2\.0for chain graphs toρ\(A\)=49\.0\\rho\(A\)=49\.0for complete graphs, as summarized in Table[1](https://arxiv.org/html/2606.27780#S5.T1)\.
##### Fixed\-edge training\.
In the fixed\-edge setting, the graph topology is fixed during rollout and each model predicts node states only\. We train all six baselines on the FE simulator described in Section[5\.2](https://arxiv.org/html/2606.27780#S5.SS2)\. Evaluation uses rollout horizonsH∈\{1,2,4,8,16,32\}H\\in\\\{1,2,4,8,16,32\\\}, allowing us to measure both short\-horizon prediction quality and long\-horizon error accumulation, following the standard concern of compounding error in model\-based rolloutAsadiet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib1)\); Janneret al\.\([2019](https://arxiv.org/html/2606.27780#bib.bib4)\)\.
##### Dynamic\-edge training\.
In the dynamic\-edge setting, each baseline is extended with the bilinear edge\-prediction head in Equation \([5\.2](https://arxiv.org/html/2606.27780#S5.E2)\)\. The model predicts both node features and edges, and is trained with the node loss, edge BCE loss, and the graph regularizer in Equation \([6\.6](https://arxiv.org/html/2606.27780#S6.E6)\)\. This setting is related to dynamic graph representation learningYouet al\.\([2022](https://arxiv.org/html/2606.27780#bib.bib45)\)but differs in that the graph is used as an autoregressive world\-model state for rollout prediction\. All six baselines are retrained in Stream A across six topologies and three seeds, yielding6×6×3=1086\\times 6\\times 3=108runs\.
##### Heterogeneous agent testbeds\.
For the agent calling\-tree and platform skill\-graph testbeds, we evaluate six R\-GCN variants: small, medium, deep, wide, action\-node, and critical\-node variants\. These models handle typed nodes and typed edges in heterogeneous graphs, following the general message\-passing view of graph networksBattagliaet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib36)\)\. Each model is trained for 50 epochs with T\-batched forward passes\.
### 7\.2RQ1: Does Topology Predict Rollout\-Error Growth?
We first ask whether topology explains the*rate*at which rollout error grows\. The theory predicts thatρ\(A\)\\rho\(A\)andGEAF^\\widehat\{\\mathrm\{GEAF\}\}should appear most clearly in temporal growth, not necessarily in the final error level after optimization and error floors intervene\. We therefore test whether higher\-GEAF graphs have faster GrowthSlope, and whether this behavior differs from standard GNN over\-smoothing\.
#### 7\.2\.1GEAF tracks error\-growth slope
Figure[4](https://arxiv.org/html/2606.27780#S7.F4)relateslogGEAF^\\log\\widehat\{\\mathrm\{GEAF\}\}to GrowthSlope4→32, the slope of the log\-NodeMSE curve from horizon44to horizon3232\. The trend is positive: runs with largerlogGEAF^\\log\\widehat\{\\mathrm\{GEAF\}\}tend to accumulate error faster\. This matches Corollary[5](https://arxiv.org/html/2606.27780#Thmtheorem5), which predicts that topology\-dependent amplification should be more visible in the log\-error slope than in a single terminal NodeMSE value\. The takeaway is not that GEAF alone determines absolute error, but that it tracks the temporal failure mode: high\-GEAF graphs are more prone to long\-horizon error growth\.
Figure 4:GEAF correlates with rollout\-error growth\.Each point is a graph\-model run; the vertical axis is GrowthSlope4→32, the slope of the log\-NodeMSE curve from horizon44to horizon3232\. The positive trend supports the interpretation that GEAF is a growth\-rate diagnostic rather than a terminal\-error predictor\.
#### 7\.2\.2Rollout amplification is distinct from over\-smoothing
We next compare GWM rollout amplification with classical GNN over\-smoothing\. The two effects use different operators\. Over\-smoothing is governed by repeated normalized propagation, which pushes node representations toward a low\-variation domain\. GWM rollout error is governed by the autoregressive amplification factorGEAF^\(ℓ\)=ρ\(A\)ℓcℓ\\widehat\{\\mathrm\{GEAF\}\}\(\\ell\)=\\rho\(A\)^\{\\ell\}c^\{\\ell\}, where depth and topology jointly affect error growth\.
Figure[5](https://arxiv.org/html/2606.27780#S7.F5)shows this distinction\. The over\-smoothing operator remains near the normalized propagation domain, whileGEAF^\(ℓ\)\\widehat\{\\mathrm\{GEAF\}\}\(\\ell\)grows rapidly with message\-passing depth and graph spectral radius\. Thus, GWM error amplification is not a restatement of standard over\-smoothing; it describes a separate rollout\-error mechanism\.
Figure 5:Over\-smoothing and GWM error amplification use different operators\.The over\-smoothing operator remains near the normalized propagation domain, whileGEAF^\(ℓ\)\\widehat\{\\mathrm\{GEAF\}\}\(\\ell\)grows with message\-passing depth and graph spectral radius\. This separates representation collapse from autoregressive rollout amplification\.
#### 7\.2\.3Global spectral amplification differs from local hub failure
We further separate global spectral amplification from local hub\-mediated failure\. Figure[6](https://arxiv.org/html/2606.27780#S7.F6)shows that complete graphs have the largestGEAF^\\widehat\{\\mathrm\{GEAF\}\}, consistent with aρ\(A\)\\rho\(A\)\-dominated global error ceiling\. However, star graphs show the largest affected\-node fraction under hub injection, despite having lower GEAF than complete graphs\. This indicates that global rollout amplification and local failure spread are related but distinct mechanisms:ρ\(A\)\\rho\(A\)controls the global ceiling, while hub concentration controls localized propagation\.
Figure 6:Global GEAF and local hub failure are distinct\.Left: complete graphs have the largestGEAF^\\widehat\{\\mathrm\{GEAF\}\}, confirming aρ\(A\)\\rho\(A\)\-dominated global error ceiling\. Right: star graphs show the largest affected\-node fraction under hub injection\.
#### 7\.2\.4Topology effects strengthen with rollout horizon and edge density
We study how rollout error changes with horizon\. The experiment evaluatesH∈\{1,2,4,8,16,32,48\}H\\in\\\{1,2,4,8,16,32,48\\\}across six baselines, seven topologies, and three seeds\. Figure[7](https://arxiv.org/html/2606.27780#S7.F7)reports NodeMSE@HHas a function of rollout horizon\.
Low\-risk topologies such as chain and tree remain stable over long horizons\. High\-risk topologies such as scale\-free, star, and grid show sharp error growth at larger horizons\. This complements the regret analysis: long\-horizon behavior is strongly topology\-dependent, and short\-horizon performance does not fully characterize rollout stability\.
Figure 7:Rollout length scaling\.NodeMSE@HHis plotted as a function of rollout horizon across graph topologies\. Shaded regions indicate variation across baselines and seeds\.We next vary edge density using Erdős–Rényi graphs with edge probabilityp∈\{0\.02,0\.05,0\.10,0\.20\}p\\in\\\{0\.02,0\.05,0\.10,0\.20\\\}\. The experiment compares the vanilla GCN world model, ActionNode GWM, and Error\-Aware GWM over three seeds\. Figure[8](https://arxiv.org/html/2606.27780#S7.F8)reports the resulting spectral radius and NodeMSE@32\.
As edge density increases,ρ\(A\)\\rho\(A\)increases\. The vanilla GCN world model shows rapidly increasing NodeMSE at high densities, while ActionNode GWM and Error\-Aware GWM remain near the low\-error floor\. This supports the interpretation that edge density affects rollout difficulty through the topology\-dependent amplification factor\.
Figure 8:Edge density sweep\.Left: adjacency spectral radiusρ\(A\)\\rho\(A\)as a function of Erdős–Rényi edge probabilitypp\. Right: NodeMSE@32 as a function of edge probabilitypp\. Results are averaged over three seeds\.
### 7\.3RQ2: Does Rollout Error Affect Planning Regret?
We next test whether prediction error affects downstream planning\. GWMs select actions through imagined graph rollouts, so small rollout errors become harmful when they change predicted returns\. We measure whether planning regret grows with horizon and whether high\-GEAF settings lead to faster regret accumulation\.
Figure[9](https://arxiv.org/html/2606.27780#S7.F9)connects rollout error to downstream planning\. Runs with largerGEAF^\\widehat\{\\mathrm\{GEAF\}\}tend to have larger regret GrowthSlope from horizon44to horizon3232, consistent with Lemma[4](https://arxiv.org/html/2606.27780#Thmtheorem4): when the effective graph error operator is larger, imagined trajectories lose reliability faster\. The horizon\-ratio analysis tells the same story from another angle\. Planning regret at horizon3232is substantially larger than one\-step regret across every topology, showing that the planning failure is a long\-horizon rollout effect rather than a one\-step artifact\.
Figure 9:Planning\-regret growth increases with GEAF\.Regret GrowthSlope from horizon44to horizon3232is plotted againstlog10GEAF^\\log\_\{10\}\\widehat\{\\mathrm\{GEAF\}\}\. Higher\-GEAF runs tend to accumulate planning regret faster, linking the rollout\-error operator to downstream decision quality\.
### 7\.4RQ3: Does Error\-Aware GWM Improve Long\-Horizon Stability?
We evaluate whether Error\-Aware GWM stabilizes long\-horizon rollout without sacrificing accuracy\. This RQ compares the full objective with vanilla GCN rollouts and simpler stabilizers such as gradient clipping, weight decay, and spectral\-norm projection\. We also test robustness under graph rewiring, noisy observations, and graph\-size scaling\.
#### 7\.4\.1Error\-Aware GWM prevents divergence while preserving accuracy
Figure[10](https://arxiv.org/html/2606.27780#S7.F10)gives the central FE comparison\. On high\-risk scale\-free and star topologies, the vanilla GCN world model enters an exploding\-error regime as the horizon grows, while Error\-Aware GWM stays near the low\-error floor\. Across all paired topology\-seed cells, the full objective eliminates the divergences observed in the vanilla GCN setting\. The result supports the design principle behind Error\-Aware GWM: spectral control is most useful when paired with rollout consistency and critical\-node weighting, so the model remains both stable and predictive\.
Figure 10:Error\-Aware GWM prevents long\-horizon divergence\.NodeMSE@HHis shown for the vanilla GCN world model and Error\-Aware GWM on scale\-free and star topologies\. Error\-Aware GWM stays near the low\-error floor, while the vanilla model exhibits horizon\-dependent error explosion\.
#### 7\.4\.2Ablation: Stability alone is not enough
Table[2](https://arxiv.org/html/2606.27780#S7.T2)shows that simple stabilizers reduce divergence, but they neither eliminate it nor preserve prediction accuracy\. Weight decay and spectral\-norm projection lower divergence to 4/21 cells, yet their median NodeMSE@32 remains orders of magnitude larger than Error\-Aware GWM\. Spectral\-norm projection is1690×1690\\timesworse than Error\-Aware GWM, while weight decay collapses toward a near\-constant predictor\. The advantage of Error\-Aware GWM is therefore not only divergence reduction; it is the combination of stability and predictive accuracy\.
Table 2:Regularization ablation over 21 FE cells\. Simple stabilizers reduce divergence but leave large long\-horizon prediction error; Error\-Aware GWM is the only variant that combines zero divergence with a low NodeMSE@32 floor\.VariantRegularizerDivergedMedian NodeMSE@32Ratio vs\. Error\-Aware GWMVanilla GCN world modelnone10/210\.9956934×6934\\timesGCN \+ gradient clippinggradient clipping8/210\.9875790×5790\\timesGCN \+ weight decayweight decay4/211\.0911\.0914172×4172\\timesGCN \+ spectral\-norm projectionspectral norm projection4/210\.3641690×1690\\timesError\-Aware GWMRspec\+Rroll\+RcritR\_\{\\text\{spec\}\}\+R\_\{\\text\{roll\}\}\+R\_\{\\text\{crit\}\}0/212\.67×𝟏𝟎−𝟒\\mathbf\{2\.67\\times 10^\{\-4\}\}1×\\times
#### 7\.4\.3Robustness to topology perturbation, noise, and graph size\.
##### Graph Rewiring Robustness\.
We evaluate how topology rewiring affects long\-horizon rollout error on the scale\-free topology\. We compare six rewiring operations: identity, PageRank\-based edge removal, random edge addition, random edge flipping, sparsification, and densification\. The evaluation covers the vanilla GCN world model, ActionNode GWM, and Error\-Aware GWM over three seeds\.
Figure[11](https://arxiv.org/html/2606.27780#S7.F11)shows that the vanilla GCN world model remains in a high\-error regime across all rewiring methods\. ActionNode GWM is stable under most rewiring operations but shows higher error under densification and random edge addition\. Error\-Aware GWM remains stable across all rewiring types, with NodeMSE@H=20H=20in the range of approximately2\.02\.0–5\.0×10−45\.0\\times 10^\{\-4\}\. This supports the role of spectral regularization in controlling topology\-induced rollout error under graph perturbations\.
Figure 11:Graph rewiring comparison\.NodeMSE@H=20H=20is reported for six rewiring methods on the scale\-free topology across three world\-model variants and three seeds\. Identity denotes the original graph without rewiring\. Error\-Aware GWM remains stable across rewiring operations, supporting the role of spectral regularization under topology perturbation\.
##### Robustness to Noisy Graph Observations\.
We test whether observation noise is the main source of rollout instability\. Gaussian noise withσ∈\{0,10−3,10−2,0\.05,0\.1,0\.2\}\\sigma\\in\\\{0,10^\{\-3\},10^\{\-2\},0\.05,0\.1,0\.2\\\}is added to graph observations before rollout evaluation\. Figure[12](https://arxiv.org/html/2606.27780#S7.F12)reports NodeMSE@H=20H=20across baselines\.
Most baselines remain nearly unchanged as observation noise increases, while the vanilla GCN world model is already in a high\-error regime atσ=0\\sigma=0\. Thus, the observed rollout instability is mainly driven by model\-topology dynamics rather than small Gaussian observation noise\.
Figure 12:Robustness to noisy graph observations\.NodeMSE@H=20H=20is plotted as a function of Gaussian observation noise levelσ\\sigmaacross baselines\. The weak dependence onσ\\sigmashows that the main instability is model\-topology amplification rather than additive observation noise\.
##### Graph Size Scaling\.
We evaluate graph\-size scaling overN∈\{20,50,100,200,500\}N\\in\\\{20,50,100,200,500\\\}across five topologies, six baselines, and three seeds\. Figure[13](https://arxiv.org/html/2606.27780#S7.F13)reportsGEAF^\\widehat\{\\mathrm\{GEAF\}\}and NodeMSE@32 as functions of graph size\.
Across graph sizes, ActionNode GWM, GPS, and Error\-Aware GWM remain near the low\-error regime, while the vanilla GCN world model exhibits large rollout errors on high\-risk cells\. MPNN runs encounter out\-of\-memory failures at larger graph sizes due to dense message tensors\. These results show that the stability patterns observed atN=50N=50persist across a broader graph\-size range\.
Figure 13:Graph size scaling\.Left:GEAF^\\widehat\{\\mathrm\{GEAF\}\}as a function of graph sizeNN\. Right: NodeMSE@32 as a function of graph sizeNN\. Curves report means across seeds; shaded regions indicate variation across seeds or runs\. The low\-error regime of ActionNode GWM, GPS, and Error\-Aware GWM persists as graph size grows, while the vanilla GCN baseline remains vulnerable on high\-risk cells\.
### 7\.5RQ4: When Does Dynamic\-Edge Coupling Matter?
We study when node\-edge coupling matters\. In FE models, edges are fixed and the coupling term is dormant\. In DE models, edge prediction allows node and edge errors to affect each other\. We test whether DE training improves DE rollout accuracy and whether the spectral witnessρ\(B\)\>max\(LX,MA\)\\rho\(B\)\>\\max\(L\_\{X\},M\_\{A\}\)appears in trained models\.
#### 7\.5\.1DE training improves DE rollout accuracy
We evaluate whether training on dynamic\-edge trajectories improves rollout accuracy in DE environments\. The paired comparison usesn=36n=36paired cells: ActionNode GWM and Error\-Aware GWM across six topologies and three seeds\. Each pair compares a model trained only on fixed\-edge trajectories but evaluated on DE rollouts against the same architecture trained directly on DE trajectories\.
Dynamic\-edge training substantially improves DE rollout accuracy\. The improvement holds for both ActionNode GWM and Error\-Aware GWM, and every paired comparison favors DE training\. This supports Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2): in DE environments, the model must learn node dynamics and the edge\-prediction pathway through which node and edge errors interact\. FE\-trained models lack this pathway and are systematically mismatched to evolving graph structure\.
Table 3:Dynamic\-edge training improves DE rollout accuracy\. The ratio reports FE\-trained NodeMSE@32 divided by DE\-trained NodeMSE@32, so larger values indicate stronger gains from DE training\. The paired comparison covers ActionNode GWM and Error\-Aware GWM across six topologies and three seeds; the bottom row reports the spectral witness for non\-degenerate node\-edge coupling\.
#### 7\.5\.2DE models activate node\-edge coupling
We test whether the joint node\-edge operator in Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2)becomes active when the model is trained with an edge\-prediction head\. In the DE regime, the edge predictor creates a node\-to\-edge pathway withMX\>0M\_\{X\}\>0, while message passing provides an edge\-to\-node pathway withLA\>0L\_\{A\}\>0\. Therefore, the lemma predicts the spectral witness
ρ\(B\)\>max\(LX,MA\)\\rho\(B\)\>\\max\(L\_\{X\},M\_\{A\}\)\(7\.1\)whenever the coupling is non\-degenerate\.
Across all 108 DE\-trained cells, the witness holds in 108/108 cases, with sign\-testp<10−30p<10^\{\-30\}and mean excess approximately155155\(Table[3](https://arxiv.org/html/2606.27780#S7.T3)\)\. Since the inequality follows mathematically onceLAMX\>0L\_\{A\}M\_\{X\}\>0, the empirical content is the magnitude rather than only the direction\. The large excess shows that learned DE models do not collapse to a nearly fixed\-edge solution; they learn substantive node\-edge coupling\.
#### 7\.5\.3FE models show dormant coupling and perturbation contraction
In the FE regime, the edge\-prediction head is absent, soMX=0M\_\{X\}=0and the joint operator reduces to node\-only dynamics\. The super\-additive witness in Equation \([7\.1](https://arxiv.org/html/2606.27780#S7.E1)\) therefore cannot activate\. Figure[14](https://arxiv.org/html/2606.27780#S7.F14)supports this distinction\. The left panel shows injected perturbations decaying rapidly to the numerical floor, indicating contractive FE dynamics\. The right panel shows that Error\-Aware GWM substantially lowers GEAF compared with ActionNode GWM, consistent with spectral regularization reducing the rollout amplification ceiling\.
Figure 14:FE\-regime contraction and GEAF reduction\.Left: injected perturbations decay rapidly to the numerical floor in the FE regime, indicating dormant node\-edge coupling\. Right: Error\-Aware GWM achieves substantially lower GEAF than ActionNode GWM across topologies, showing how spectral regularization reduces the amplification ceiling\.Figure[15](https://arxiv.org/html/2606.27780#S7.F15)decomposes FE\-regime GrowthSlope under node\-only, edge\-only, and joint perturbation injection\. Node\-only perturbations have consistently negative slopes, indicating self\-correction under contractive FE dynamics\. Edge\-only slopes are close to zero, and joint slopes closely follow edge\-only slopes\. This is consistent with the FE structureMX=0M\_\{X\}=0, where the joint node\-edge operator does not activate cross\-coupling\.
Figure 15:Perturbation decomposition in the FE regime\.Node\-only perturbations are absorbed by contractive dynamics, edge\-only perturbations remain near zero growth, and joint perturbations closely follow edge\-only behavior\. This supports the interpretation that node\-edge cross\-coupling is dormant in the FE regime whereMX=0M\_\{X\}=0\.
#### 7\.5\.4OOD perturbation confirms the FE contraction boundary
We perform an OOD amplitude retest by training on chain graphs and evaluating on scale\-free graphs\. Perturbation amplitudes are swept overδ∈\{0\.1,0\.5,1\.0,2\.0\}\\delta\\in\\\{0\.1,0\.5,1\.0,2\.0\\\}, and injections are applied to six positions\. Figure[16](https://arxiv.org/html/2606.27780#S7.F16)reports NodeMSE@H=20H=20for the vanilla GCN world model, which is the only baseline with a measurable nonzero signal in this setting\.
NodeMSE increases with perturbation amplitude across all injection positions\. However, the pre\-registered high\-risk position tests do not reject after Holm correction\. For the other baselines, perturbations remain absorbed with NodeMSE below the measurement floor\. These results support the scope interpretation that FE dynamics are strongly contractive for most architectures, while measurable OOD sensitivity appears mainly in the vanilla GCN world model\.
Figure 16:OOD amplitude and injection\-position sweep\.The heatmap reportslog10\\log\_\{10\}NodeMSE@H=20H=20for the vanilla GCN world model under different perturbation amplitudes and injection positions\. Rows correspond to injection positions and columns correspond to perturbation amplitudes\.
#### 7\.5\.5Regime interpretation
##### Candidate mechanisms\.
The regime\-conditional behavior of the joint operator can be explained by four mechanisms\. First, in the FE regime, the edge\-prediction head is absent, soMX=0M\_\{X\}=0by construction and the coupling productLAMXL\_\{A\}M\_\{X\}is structurally zero\. Second, Error\-Aware GWM reduces the effective node\-transition gain by spectral regularization, making FE perturbations more likely to contract\. Third, if the coupling product is positive but small, longer horizons may be required for the asymptotic rate to appear; this explanation does not apply whenMX=0M\_\{X\}=0\. Fourth, small perturbations may remain in a local linear regime whose Jacobian is much smaller than the global Lipschitz bound\. The multi\-axis sweep over perturbation size and horizon separates these cases and supports the FE contraction interpretation\.
##### Regime interpretation\.
These results support a regime\-conditional view of graph rollout error\. Fixed\-edge GWM corresponds to theMX=0M\_\{X\}=0corner of the joint\-operator framework, where node\-edge cross\-coupling is dormant\. Dynamic\-edge training activates the off\-diagonal coupling terms and changes the effective rollout\-error operator\. The framework therefore does not replace fixed\-edge GWM; it explains when fixed\-edge dynamics are sufficient and when dynamic\-edge training is required\.
### 7\.6RQ5: How Do GWMs Behave on Heterogeneous Agent Graphs?
We evaluate GWMs on heterogeneous agent and skill graphs, where nodes have roles such as planner, executor, validator, tool, and failure handler\. This RQ tests whether rollout failures depend on role structure, whether R\-GCN rollouts are calibrated against simulators, and whether correction policies can reduce agent\-level errors\.
#### 7\.6\.1Calibration against heterogeneous simulators
We calibrate the R\-GCN rollout models against the ground\-truth heterogeneous simulators on the agent calling\-tree and platform skill\-graph testbeds\. Figure[17](https://arxiv.org/html/2606.27780#S7.F17)compares R\-GCN predictions with simulator outputs after the sigmoid\-head fix\. On the agent calling\-tree testbed, the R\-GCN mean is0\.4420\.442versus ground\-truth mean0\.4370\.437, with Pearsonr=0\.311r=0\.311and MAE=0\.051=0\.051\. On the skill\-graph testbed, the R\-GCN mean is0\.3870\.387versus ground\-truth mean0\.4060\.406, with Pearsonr=0\.231r=0\.231and MAE=0\.029=0\.029\.
These results show that the R\-GCN models are reasonably calibrated in mean and absolute error, but their instance\-level correlations are weak to moderate\. We therefore use the heterogeneous R\-GCN results as agent\-graph stress tests and calibration evidence, not as a high\-precision replacement for the ground\-truth simulator\.
Figure 17:R\-GCN calibration against heterogeneous simulators\.Left: agent calling\-tree testbed\. Right: platform skill\-graph testbed\. The predictions are calibrated in mean and absolute error, but instance\-level correlations remain modest, so these results are used as stress\-test evidence rather than high\-precision simulator replacement\.
#### 7\.6\.2Role\-dependent failure sensitivity
We evaluate whether different functional subgraphs contribute differently to agent\-task success\. We mask role\-specific subgraphs in the agent calling\-tree testbed and measure both success\-rate drop and trajectory\-level NodeMSE\. Figure[18](https://arxiv.org/html/2606.27780#S7.F18)shows that full\-graph masking gives the largest degradation, as expected\. Among partial masks, hub and validator\-related components produce larger changes than leaf masking, while leaf masking has almost no effect\.
These results support role\-dependent error amplification\. Not all nodes contribute equally to rollout quality: masking structurally or functionally central components has a larger effect than masking peripheral leaves\.
Figure 18:Critical subgraph masking on the agent calling\-tree testbed\.Left: success\-rate drop after masking different subgraphs\. Right: NodeMSE@TTunder the same masking conditions\. Full masking gives the largest degradation, while leaf masking has minimal effect\.We evaluate how injected errors propagate across different functional roles in the agent calling\-tree testbed\. We inject perturbations into planner, validator, hub, bridge, and random nodes across 20 instances\. Figure[19](https://arxiv.org/html/2606.27780#S7.F19)reports failure propagation depth \(FPD\) and the number of error nodes at the final timestep\.
Planner and validator injections produce median FPD around 3, while bridge and hub injections have smaller median FPD\. Random injection has the largest median FPD at 4\. The final error\-node counts are broadly similar across injection types, but the distributions vary by role\. These results show that failure propagation depends on functional role, not only graph position\.
Figure 19:Agent\-role injection analysis\.Left: failure propagation depth distributions by injection kind\. Right: number of error nodes at the final timestep by injection kind\.
#### 7\.6\.3Correction policies
We also test whether a local GEAF proxy can guide node\-level correction policies in heterogeneous agent graphs\. In the current FE formulation,
GEAF^local\(v\)=degree\(v\)∏ℓ‖Wℓ‖2\.\\widehat\{\\mathrm\{GEAF\}\}\_\{\\rm local\}\(v\)=\\mathrm\{degree\}\(v\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\.\(7\.2\)Since∏ℓ‖Wℓ‖2\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}is global for a fixed model,GEAF^local\(v\)\\widehat\{\\mathrm\{GEAF\}\}\_\{\\rm local\}\(v\)induces the same node ranking as degree centrality\. Figure[20](https://arxiv.org/html/2606.27780#S7.F20)confirms this equivalence: the GEAF\-proxy policy and the degree policy achieve the same mean error reduction of 12\.9%, outperforming random correction but remaining below the oracle policy at 22\.4%\. Thus, local GEAF is useful as a non\-random structural heuristic, but in this FE setting it does not provide a distinct correction signal beyond degree centrality\.
Figure 20:Local GEAF correction reduces to degree centrality in the FE setting\.Correction policies are compared on the heterogeneous agent calling\-tree testbed under a 10% node\-correction budget\. GEAF\-proxy and degree policies coincide because the FE local\-GEAF score is degree times a model\-global norm product\.We next compare node correction policies under a 10% correction budget\. Figure[21](https://arxiv.org/html/2606.27780#S7.F21)reports NodeMSE reduction across six baselines, six topologies, and three seeds\. Oracle correction gives the strongest upper\-bound performance\. Among practical policies, uncertainty\-based correction gives the largest average gain over random correction\. GEAF\-proxy correction improves over random but is indistinguishable from degree\-based correction\.
This result is consistent with the FE local\-GEAF identity\. When model weights are fixed, the local GEAF score reduces to a degree\-proportional ranking:
GEAF^local\(v\)=degree\(v\)∏ℓ‖Wℓ‖2\.\\widehat\{\\mathrm\{GEAF\}\}\_\{\\rm local\}\(v\)=\\mathrm\{degree\}\(v\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\.\(7\.3\)Thus, GEAF is useful as a non\-random structural heuristic, but in this FE correction setting it does not provide a distinct node\-ranking signal beyond degree centrality\.
Figure 21:Node correction policy comparison\.NodeMSE reduction is reported under a 10% correction budget\. Oracle correction provides the upper bound, uncertainty gives the strongest practical improvement, and GEAF\-proxy correction matches degree\-based correction\.We also compare correction policies on the agent calling\-tree testbed under a 10% node\-correction budget\. The policies include random, degree, GEAF\-proxy, and oracle correction\. Figure[22](https://arxiv.org/html/2606.27780#S7.F22)reports the reduction in error\-flag mean\.
The oracle policy gives the largest reduction, while GEAF\-proxy and degree correction are tied because they induce the same node ranking in this FE local formulation\. This supports the scope statement that local GEAF is a useful structural heuristic but does not provide a distinct correction signal beyond degree centrality in the FE agent\-graph setting\.
Figure 22:Agent\-level correction policy comparison\.Error\-flag reduction is reported for random, degree, GEAF\-proxy, and oracle correction under a 10% node\-correction budget\.
#### 7\.6\.4Long\-horizon agent task completion
We evaluate agent task completion on the agent calling\-tree and platform skill\-graph testbeds\. We pool 160 instances per testbed and report task success rate \(SR\) at horizonsH∈\{1,5,10,20\}H\\in\\\{1,5,10,20\\\}\. Results are averaged over 6 testbed\-seed cells for each baseline and horizon\.
Table 4:Agent task completion across rollout horizons\. All methods decline from approximately 0\.49 atH=1H=1to approximately 0\.42 forH≥5H\\geq 5, indicating that heterogeneous agent\-graph planning remains difficult at longer horizons even when the world model uses graph structure\.Figure 23:Agent task completion across planning horizons\.Error\-Aware GWM has a small short\-horizon advantage atH=1H=1, but all baselines converge to SR≈0\.42\\approx 0\.42forH≥5H\\geq 5\.Table[4](https://arxiv.org/html/2606.27780#S7.T4)and Figure[23](https://arxiv.org/html/2606.27780#S7.F23)show that architecture choice has limited impact on heterogeneous agent\-task completion\. AtH=1H=1, all methods cluster around SR≈0\.49\\approx 0\.49\. ForH≥5H\\geq 5, all methods decline toward SR≈0\.42\\approx 0\.42\. Thus, the long\-horizon planning challenge persists even with graph\-structured world models\.
These real\-world evaluations clarify the boundary of GWMs\. Specialized graph classifiers remain stronger on static node classification, GWM variants are competitive but not dominant on temporal link prediction, and all methods struggle on long\-horizon heterogeneous agent planning\. GWMs are therefore best viewed as a framework for dynamic graph rollouts and agent planning, not as a replacement for task\-specific graph learning methods\. Error\-Aware regularization provides modest real\-world gains, with its strongest benefit in controlled rollout regimes where long\-horizon error accumulation is the main failure mode\.
### 7\.7RQ6: What Is the Real\-World Scope of GWMs?
Finally, we test GWMs on standard real\-world graph benchmarks to clarify their practical boundary\. These experiments compare GWM variants with specialized graph baselines on static node classification and temporal link prediction\. The goal is not to claim state\-of\-the\-art performance, but to identify where GWMs remain competitive and where task\-specific graph models are preferable\.
#### 7\.7\.1Static Node Classification on Cora and Citeseer
We first evaluate static node classification on Cora and CiteseerYanget al\.\([2016](https://arxiv.org/html/2606.27780#bib.bib69)\)\. Cora contains 2,708 nodes, 5,278 edges, and 7 classes; Citeseer contains 3,327 nodes, 4,732 edges, and 6 classes\. We follow the standard public train/validation/test split\. The baselines are GCNKipf and Welling \([2017](https://arxiv.org/html/2606.27780#bib.bib13)\), GATVeličkovićet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib67)\), and GraphSAGEHamiltonet al\.\([2017](https://arxiv.org/html/2606.27780#bib.bib59)\)\. All models are trained for 200 epochs with Adam, learning rate10−310^\{\-3\}, weight decay5×10−45\\times 10^\{\-4\}, dropout0\.50\.5, and 3 random seeds\.
Table 5:Static node classification test accuracy on Cora and Citeseer\. Results are mean±\\pmstandard deviation over three seeds; specialized graph classifiers remain strongest on Cora, while Error\-Aware GWM narrows the gap to FE\-GWM on both datasets\.Table[5](https://arxiv.org/html/2606.27780#S7.T5)and Figure[24](https://arxiv.org/html/2606.27780#S7.F24)show that specialized static graph classifiers remain stronger on Cora\. GCN achieves the best Cora accuracy at 80\.6%, while FE\-GWM and Error\-Aware GWM achieve 73\.5% and 75\.0%, respectively\. On Citeseer, the gap is smaller: all methods fall within an approximately 3 pp range, with GAT best at 68\.1% and Error\-Aware GWM reaching 66\.9%\. Error\-Aware GWM consistently improves over FE\-GWM by 1–1\.5 pp, suggesting that spectral regularization is mildly beneficial even in a static regime\.
Figure 24:Static node classification on Cora and Citeseer\.GCN, GAT, and GraphSAGE are standard static graph baselines, while FE\-GWM and Error\-Aware GWM are GWM variants\. Standard graph classifiers dominate Cora by 4–7 pp; on Citeseer, all methods are within an approximately 3 pp range\.
#### 7\.7\.2Temporal Link Prediction on Bitcoin\-Alpha
We evaluate temporal link prediction on Bitcoin\-Alpha\(Kumaret al\.,[2019](https://arxiv.org/html/2606.27780#bib.bib68)\), a sparse temporal trust network with 3,783 nodes and 24,186 temporal edges\. We construct 60\-day windows and compare GWM variants with a JODIE\-style temporal interaction model\(Kumaret al\.,[2019](https://arxiv.org/html/2606.27780#bib.bib68)\)and a static GCN baseline\. We report AP, AUC, and MRR over 3 seeds\.
Table 6:Temporal link prediction on Bitcoin\-Alpha\. Results are mean±\\pmstandard deviation over three seeds\. No single method dominates all metrics: FE\-GWM has the best AP, Static GCN has the best AUC, and JODIE\-style has the best MRR\.Figure 25:Temporal link prediction on Bitcoin\-Alpha\.FE\-GWM obtains the best AP, Static GCN obtains the best AUC, and JODIE\-style obtains the best MRR\. DE\-GWM does not improve over FE\-GWM on this sparse real\-world temporal graph, indicating that dynamic\-edge modeling is not automatically beneficial outside controlled DE simulators\.Table[6](https://arxiv.org/html/2606.27780#S7.T6)and Figure[25](https://arxiv.org/html/2606.27780#S7.F25)show that GWM variants are competitive but do not dominate specialized baselines\. FE\-GWM achieves the best AP, while Static GCN has the best AUC and JODIE\-style has the best MRR\. DE\-GWM and Error\-Aware DE\-GWM do not consistently improve over FE\-GWM, suggesting that dynamic\-edge modeling is most useful when edge dynamics provide stable predictive signal over the rollout horizon\.
## 8Additional Experiments
This section provides concise result summaries for all other experiments\.
### 8\.1Seed Stability of Error Metrics
We first check whether the observed error patterns are artifacts of random seed choice\. Figure[26](https://arxiv.org/html/2606.27780#S8.F26)reports seed\-level distributions of NodeMSE@32 and GrowthSlope across the seven topology families\. The distributions are stable across seeds, and deterministic topologies show limited variation\. This supports the use of three outer seeds in the main experiments and suggests that the observed GEAF\-related trends are not driven by a small number of random graph instances\.
Figure 26:Seed stability across topologies\.Left: seed\-level distributions of NodeMSE@32\. Right: seed\-level distributions of GrowthSlope\.
### 8\.2Multi\-Hop Error Propagation Is Limited Under Contractive FE Dynamics
We test whether local perturbations propagate across multiple graph hops\. We inject perturbations at depths\{1,2,4,6,8,10\}\\\{1,2,4,6,8,10\\\}on chain and scale\-free graphs across six baselines and three seeds\. Figure[27](https://arxiv.org/html/2606.27780#S8.F27)shows that multi\-hop error propagation is limited for all non\-diverged models\. On chain graphs, NodeMSE@H=20H=20remains below10−1210^\{\-12\}across all injection depths, with failure propagation depth close to 1\. On scale\-free graphs, all non\-diverged models remain below10−1310^\{\-13\}, while the diverged GCN baseline reaches NodeMSE@H=20H=20around101010^\{10\}\.
This result supports a regime\-specific interpretation: FE dynamics are often contractive and absorb local perturbations, while severe multi\-hop amplification appears mainly when the underlying baseline has already entered a divergent rollout regime\.
Figure 27:Multi\-hop error propagation under perturbation injection\.NodeMSE@H=20H=20is plotted as a function of hop distance from the injection node\. Chain graphs remain contractive with near\-zero error at all hop distances\.
### 8\.3Zero\-Shot Topology Shift
We test zero\-shot generalization under topology shift by training on one topology and evaluating on another\. The evaluated train–test pairs are chain→\\togrid, grid→\\tosmall\-world, small\-world→\\toscale\-free, scale\-free→\\tostar, and tree→\\toscale\-free\. The experiment covers six baselines and three seeds\.
Figure[28](https://arxiv.org/html/2606.27780#S8.F28)shows that cross\-topology transfer causes a large increase in NodeMSE@H=20H=20\. Excluding diverged vanilla GCN cells, the median NodeMSE@H=20H=20across baselines is0\.1960\.196, compared with an in\-distribution level of approximately3×10−43\\times 10^\{\-4\}\. Error\-Aware GWM obtains a median of0\.1830\.183, which is slightly better than the overall median but still far above the in\-distribution floor\. These results indicate that zero\-shot topology shift is a structural difficulty for graph rollouts, not merely an architecture\-specific failure\.
Figure 28:Zero\-shot topology shift\.The heatmap reports NodeMSE@H=20H=20for train–test topology pairs\. Rows indicate training topologies and columns indicate test topologies\.
### 8\.4Sparse and Dense Message\-Passing Layers
We compare sparse and dense message\-passing behavior on the scale\-free topology\. The experiment evaluates the vanilla GCN world model, MPNN world model, and GPS\-style graph transformer across rollout horizonsH∈\{1,4,8,16,32\}H\\in\\\{1,4,8,16,32\\\}and three seeds\.
Figure[29](https://arxiv.org/html/2606.27780#S8.F29)shows that the vanilla GCN world model diverges as the horizon increases, with median NodeMSE@H=32H=32reaching6\.5×1096\.5\\times 10^\{9\}\. In contrast, MPNN and GPS remain near the low\-error floor, with median NodeMSE@H=32H=32equal to2\.2×10−42\.2\\times 10^\{\-4\}and4\.0×10−44\.0\\times 10^\{\-4\}, respectively\. This suggests that the observed divergence is not explained by message\-passing density alone\. Instead, it is tied to the interaction between the GCN transition operator and the high\-risk scale\-free topology\.
Figure 29:Sparse and dense message\-passing comparison\.NodeMSE@HHis reported across rollout horizons for the vanilla GCN world model, MPNN world model, and GPS\-style graph transformer on the scale\-free topology\. Shaded regions indicate variation across seeds\.
### 8\.5Directed and Undirected Graph Variants
We evaluate whether graph directionality affects rollout growth\. The experiment uses chain, tree, and scale\-free graphs with directed and undirected variants, covering three baselines and three seeds\. Figure[30](https://arxiv.org/html/2606.27780#S8.F30)reports GrowthSlope8→32\.
Directionality changes the error\-growth distribution, especially on scale\-free graphs, where the variance is larger\. Chain and tree graphs show smaller directionality effects\. These results indicate that directed edges can change the effective rollout operator, but the effect depends strongly on topology\.
Figure 30:Directed and undirected graph variants\.GrowthSlope8→32is reported for directed and undirected variants of chain, scale\-free, and tree graphs\. Bars show means and error bars show variation across runs\.
### 8\.6Temporal Memory Variants
We compare temporal\-memory variants for long\-horizon rollout prediction\. The main ablation evaluates Markov, last\-2, and last\-4 history variants on scale\-free and small\-world graphs\. The extended experiment compares recurrent, transformer, and retrieval\-augmented memory architectures\. Figure[31](https://arxiv.org/html/2606.27780#S8.F31)reports NodeMSE@HHacross rollout horizons\.
Last\-kkhistory reduces NodeMSE relative to the Markov variant, with most of the gain appearing from the last\-2 context\. In the extended comparison, retrieval\-augmented memory achieves lower long\-horizon NodeMSE than recurrent and transformer variants\. These results suggest that adding temporal context can improve long\-horizon rollout accuracy, especially when the memory mechanism provides stable retrieval rather than only recurrent propagation\.
Figure 31:Temporal memory variants\.NodeMSE@HHis reported across rollout horizons for recurrent, transformer, and retrieval\-augmented memory variants\. Shaded regions indicate variation across topologies and seeds\.
## 9Conclusion
We present a unified framework for analyzing and improving Graph World Models in graph\-structured planning environments\. The framework covers fixed\-edge and dynamic\-edge regimes, introduces topology\-aware rollout\-error analysis through GEAF and the joint node\-edge operatorBB, and connects rollout error to planning regret\. Empirically, long\-horizon graph planning is sensitive to topology and regime: regret increases with rollout horizon, dynamic\-edge training is needed when graph structure evolves, and Error\-Aware GWM improves stability while preserving prediction accuracy\. Real\-world experiments show that GWMs are best viewed as tools for dynamic graph rollouts and agent planning, not as replacements for specialized graph learning methods\. Future work includes applying this framework to LLM\-based multi\-agent systems, heterogeneous knowledge graphs, and real\-time network planning\.
## 10Discussion
Our results show that GWM rollout error is regime\-dependent\. In fixed\-edge settings, instability mainly comes from the interaction between topology and the learned node\-transition operator\. This explains why the vanilla GCN world model fails on high\-ρ\(A\)\\rho\(A\)graphs, while Error\-Aware GWM avoids divergence by combining spectral control, rollout consistency, and critical\-node weighting\. In dynamic\-edge settings, edge prediction activates node\-edge cross\-coupling through the joint operatorBB, and DE training substantially improves DE rollout accuracy\. GEAF is therefore best interpreted as a topology\-aware indicator of rollout\-error growth, not as a universal predictor of final error level\. Locally, the FE version of GEAF reduces to degree centrality, so future correction policies should use activation\-dependent, uncertainty\-aware, or DE\-regime coupling terms to provide a distinct node\-priority signal\.
## 11Limitations and Future Work
The strongest evidence in this work comes from controlled synthetic graph simulators\. Real\-world benchmarks show more limited transfer: GWMs are not replacements for specialized static graph classifiers, and DE\-GWM does not consistently dominate on sparse temporal link prediction\. The joint\-operator theory is also only partially tested: FE models collapse to theMX=0M\_\{X\}=0case, while larger DE datasets are needed to evaluate the GEAF proxy forρ\(B\)\\rho\(B\)\. Future work should study larger and denser temporal graphs, real agent\-system traces, activation\-aware local GEAF, uncertainty\-guided correction, and memory\-augmented GWMs for long\-horizon planning\.
## References
- Learning knowledge graph\-based world models of textual environments\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p2.3),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px3.p1.1)\.
- P\. Anokhin, N\. Semenov, A\. Sorokin, D\. Evseev, M\. Burtsev, and A\. Panov \(2024\)AriGraph: learning knowledge graph world models with episodic memory for llm agents\.InInternational Joint Conference on Artificial Intelligence \(IJCAI\),Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px4.p1.1)\.
- K\. Asadi, D\. K\. Misra, and M\. L\. Littman \(2018\)Lipschitz continuity in model\-based reinforcement learning\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1),[§1](https://arxiv.org/html/2606.27780#S1.p4.11),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px3.p1.1)\.
- A\. Barabási and R\. Albert \(1999\)Emergence of scaling in random networks\.Science286\(5439\),pp\. 509–512\.Cited by:[§5\.1](https://arxiv.org/html/2606.27780#S5.SS1.p2.9),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px2.p1.3)\.
- P\. W\. Battaglia, J\. B\. Hamrick, V\. Bapst, A\. Sanchez\-Gonzalez, V\. Zambaldi, M\. Malinowski, A\. Tacchetti, D\. Raposo, A\. Santoro, R\. Faulkner, C\. Gulcehre, F\. Song, A\. Ballard, J\. Gilmer, G\. Dahl, A\. Vaswani, K\. Allen, C\. Nash, V\. Langston, C\. Dyer, N\. Heess, D\. Wierstra, P\. Kohli, M\. Botvinick, O\. Vinyals, Y\. Li, and R\. Pascanu \(2018\)Relational inductive biases, deep learning, and graph networks\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p3.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px5.p1.1)\.
- P\. W\. Battaglia, R\. Pascanu, M\. Lai, D\. J\. Rezende, and K\. Kavukcuoglu \(2016\)Interaction networks for learning about objects, relations and physics\.InAdvances in Neural Information Processing Systems,Vol\.29\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p3.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Bruce, M\. D\. Dennis, A\. Edwards, J\. Parker\-Holder, Y\. Shi, E\. Hughes, M\. Lai, A\. Mavalankar, R\. Steigerwald, C\. Apps, Y\. Aytar, S\. Bechtle, F\. Behbahani, S\. Chan, N\. Heess, L\. Gonzalez, S\. Osindero, S\. Ozair, S\. Reed, J\. Zhang, K\. Zolna, J\. Clune, N\. de Freitas, S\. Singh, and T\. Rocktäschel \(2024\)Genie: generative interactive environments\.InProceedings of the 41st International Conference on Machine Learning,Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1)\.
- B\. Chamberlain, J\. Rowbottom, M\. I\. Gorinova, M\. Bronstein, S\. Webb, and E\. Rossi \(2021\)GRAND: graph neural diffusion\.InProceedings of the 38th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.139,pp\. 1407–1418\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p3.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1)\.
- R\. Chen, T\. Zhao, A\. Jaiswal, N\. Shah, and Z\. Wang \(2024\)LLaGA: large language and graph assistant\.InProceedings of the 41st International Conference on Machine Learning,Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p2.3)\.
- T\. Chen, X\. Hu, Z\. Ding, and C\. Jin \(2025\)Learning world models for interactive video generation\.arXiv preprint arXiv:2505\.21996\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- K\. Chua, R\. Calandra, R\. McAllister, and S\. Levine \(2018\)Deep reinforcement learning in a handful of trials using probabilistic dynamics models\.InAdvances in Neural Information Processing Systems,Vol\.31\.Cited by:[§B\.1](https://arxiv.org/html/2606.27780#A2.SS1.p1.4),[§1](https://arxiv.org/html/2606.27780#S1.p1.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- T\. Feng, Y\. Wu, G\. Lin, and J\. You \(2025\)Graph world model\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p2.3),[§1](https://arxiv.org/html/2606.27780#S1.p6.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px3.p1.1),[§3\.2](https://arxiv.org/html/2606.27780#S3.SS2.p1.1),[§3\.3](https://arxiv.org/html/2606.27780#S3.SS3.SSS0.Px1.p1.3),[§6](https://arxiv.org/html/2606.27780#S6.SS0.SSS0.Px1.p1.2),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px1.p1.1)\.
- J\. Gilmer, S\. S\. Schoenholz, P\. F\. Riley, O\. Vinyals, and G\. E\. Dahl \(2017\)Neural message passing for quantum chemistry\.InProceedings of the 34th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.70,pp\. 1263–1272\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p6.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px1.p1.1)\.
- N\. Gürtler and G\. Martius \(2025\)Long\-horizon planning with predictable skills\.Reinforcement Learning Journal6,pp\. 1247–1272\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- D\. Ha and J\. Schmidhuber \(2018\)World models\.arXiv preprint arXiv:1803\.10122\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1)\.
- D\. Hafner, T\. Lillicrap, J\. Ba, and M\. Norouzi \(2020\)Dream to control: learning behaviors by latent imagination\.InInternational Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1)\.
- D\. Hafner, T\. Lillicrap, I\. Fischer, R\. Villegas, D\. Ha, H\. Lee, and J\. Davidson \(2018\)Learning latent dynamics for planning from pixels\.arXiv preprint arXiv:1811\.04551\.Cited by:[§B\.1](https://arxiv.org/html/2606.27780#A2.SS1.p1.4),[§1](https://arxiv.org/html/2606.27780#S1.p1.1)\.
- D\. Hafner, T\. Lillicrap, M\. Norouzi, and J\. Ba \(2023\)Mastering diverse domains through world models\.arXiv preprint arXiv:2301\.04104\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1)\.
- W\. L\. Hamilton, R\. Ying, and J\. Leskovec \(2017\)Inductive representation learning on large graphs\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§7\.7\.1](https://arxiv.org/html/2606.27780#S7.SS7.SSS1.p1.3),[Table 5](https://arxiv.org/html/2606.27780#S7.T5.8.6.6.3)\.
- K\. Huang and J\. Huang \(2025\)Audited skill\-graph self\-improvement for agentic llms via verifiable rewards, experience synthesis, and continual memory\.arXiv preprint arXiv:2512\.23760\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px4.p1.1)\.
- X\. Huang, J\. Chen, Y\. Fei, Z\. Li, P\. Schwaller, and G\. Ceder \(2026\)CASCADE: cumulative agentic skill creation through autonomous development and evolution\.External Links:2512\.23880,[Link](https://arxiv.org/abs/2512.23880)Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px4.p1.1)\.
- M\. Janner, J\. Fu, M\. Zhang, and S\. Levine \(2019\)When to trust your model: model\-based policy optimization\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§B\.1](https://arxiv.org/html/2606.27780#A2.SS1.p1.4),[§1](https://arxiv.org/html/2606.27780#S1.p1.1),[§1](https://arxiv.org/html/2606.27780#S1.p4.11),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px3.p1.1)\.
- S\. Kakade and J\. Langford \(2002\)Approximately optimal approximate reinforcement learning\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- T\. N\. Kipf and M\. Welling \(2017\)Semi\-supervised classification with graph convolutional networks\.InInternational Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p6.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px1.p1.1),[§7\.7\.1](https://arxiv.org/html/2606.27780#S7.SS7.SSS1.p1.3),[Table 5](https://arxiv.org/html/2606.27780#S7.T5.4.2.2.3)\.
- T\. Kipf, E\. van der Pol, and M\. Welling \(2020\)Contrastive learning of structured world models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px3.p1.1)\.
- S\. Kumar, X\. Zhang, and J\. Leskovec \(2019\)Predicting dynamic embedding trajectory in temporal interaction networks\.InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,pp\. 1269–1278\.Cited by:[§7\.7\.2](https://arxiv.org/html/2606.27780#S7.SS7.SSS2.p1.1),[Table 6](https://arxiv.org/html/2606.27780#S7.T6.5.3.3.4)\.
- M\. Li, X\. Zhang, C\. Yang, Z\. Zheng, and Y\. Yang \(2025a\)Prismatic world model: learning compositional dynamics for planning in hybrid systems\.arXiv preprint arXiv:2512\.08411\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- X\. Li, Z\. Xia, W\. Lu, C\. Hao, and Y\. Chen \(2025b\)SmallWorlds: assessing dynamics understanding of world models in isolated environments\.arXiv preprint arXiv:2511\.23465\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- H\. Liu, J\. Feng, L\. Kong, N\. Liang, D\. Tao, Y\. Chen, and M\. Zhang \(2024\)One for all: towards training one graph model for all classification tasks\.InThe Twelfth International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p2.3)\.
- T\. Ni, E\. Derman, V\. Jain, V\. Taboga, S\. Ravanbakhsh, and P\. Bacon \(2025\)Long\-horizon model\-based offline reinforcement learning without conservatism\.arXiv preprint arXiv:2512\.04341\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- K\. Oono and T\. Suzuki \(2020\)Graph neural networks exponentially lose expressive power for node classification\.International Conference on Learning Representations \(ICLR\)\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p3.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1)\.
- L\. Rampášek, M\. Galkin, V\. P\. Dwivedi, A\. T\. Luu, G\. Wolf, and D\. Beaini \(2022\)Recipe for a general, powerful, scalable graph transformer\.InAdvances in Neural Information Processing Systems,Vol\.35,pp\. 14501–14515\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p6.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px1.p1.1)\.
- L\. Rivard, S\. Sun, H\. Guo, W\. Chen, and Y\. Deng \(2025\)NeuralOS: towards simulating operating systems via neural generative models\.arXiv preprint arXiv:2507\.08800\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- A\. Sanchez\-Gonzalez, N\. Heess, J\. T\. Springenberg, J\. Merel, M\. Riedmiller, R\. Hadsell, and P\. Battaglia \(2018\)Graph networks as learnable physics engines for inference and control\.InProceedings of the 35th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.80,pp\. 4470–4479\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p3.1),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Tori, V\. Holst, and V\. Ginis \(2024\)The effectiveness of curvature\-based rewiring and the role of hyperparameters in gnns revisited\.External Links:2407\.09381,[Link](https://arxiv.org/abs/2407.09381)Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1)\.
- P\. Veličković, G\. Cucurull, A\. Casanova, A\. Romero, P\. Liò, and Y\. Bengio \(2018\)Graph attention networks\.InInternational Conference on Learning Representations,Cited by:[§7\.7\.1](https://arxiv.org/html/2606.27780#S7.SS7.SSS1.p1.3),[Table 5](https://arxiv.org/html/2606.27780#S7.T5.6.4.4.3)\.
- J\. Wang, Q\. Yan, Y\. Wang, Y\. Tian, S\. S\. Mishra, Z\. Xu, M\. Gandhi, P\. Xu, and L\. L\. Cheong \(2026\)Reinforcement learning for self\-improving agent with skill library\.External Links:2512\.17102,[Link](https://arxiv.org/abs/2512.17102)Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px4.p1.1)\.
- Z\. Wang, K\. Wang, L\. Zhao, P\. Stone, and J\. Bian \(2025\)Dyn\-o: building structured world models with object\-centric representations\.External Links:2507\.03298,[Link](https://arxiv.org/abs/2507.03298)Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px3.p1.1)\.
- D\. J\. Watts and S\. H\. Strogatz \(1998\)Collective dynamics of ‘small\-world’ networks\.Nature393\(6684\),pp\. 440–442\.Cited by:[§5\.1](https://arxiv.org/html/2606.27780#S5.SS1.p2.9),[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px2.p1.3)\.
- J\. Wu, S\. Hu, N\. Guo, L\. Guo, J\. Du, J\. Yan, J\. Hao, and M\. Huang \(2024\)iVideoGPT: interactive VideoGPTs are scalable world models\.arXiv preprint arXiv:2405\.15223\.Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p1.1)\.
- T\. Xia, L\. Hu, Y\. Sun, M\. Xu, L\. Xu, S\. Wang, W\. Xu, and J\. Jiang \(2026\)GraSP: graph\-structured skill compositions for llm agents\.External Links:2604\.17870,[Link](https://arxiv.org/abs/2604.17870)Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px4.p1.1)\.
- Z\. Yang, W\. W\. Cohen, and R\. Salakhutdinov \(2016\)Revisiting semi\-supervised learning with graph embeddings\.InProceedings of the 33rd International Conference on Machine Learning,Cited by:[§7\.7\.1](https://arxiv.org/html/2606.27780#S7.SS7.SSS1.p1.3)\.
- J\. You, T\. Du, and J\. Leskovec \(2022\)ROLAND: graph learning framework for dynamic graphs\.InACM SIGKDD Conference on Knowledge Discovery and Data Mining,Cited by:[§7\.1](https://arxiv.org/html/2606.27780#S7.SS1.SSS0.Px4.p1.1)\.
- Y\. Yu, J\. Choi, J\. Park, K\. Lee, and N\. Park \(2025\)PIORF: physics\-informed ollivier\-ricci flow for long\-range interactions in mesh graph neural networks\.External Links:2504\.04052,[Link](https://arxiv.org/abs/2504.04052)Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px2.p1.1)\.
- L\. Zhang, G\. Yang, and B\. C\. Stadie \(2021\)World model as a graph: learning latent landmarks for planning\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§1](https://arxiv.org/html/2606.27780#S1.p2.3),[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px3.p1.1)\.
- W\. Zhang, B\. Terver, A\. Zholus, S\. Chitnis, H\. Sutaria, M\. Assran, R\. Balestriero, A\. Bar, A\. Bardes, Y\. LeCun, and N\. Ballas \(2026\)Hierarchical planning with latent world models\.arXiv preprint arXiv:2604\.03208\.Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px1.p1.1)\.
- D\. Zhu, E\. L\. Li, and M\. Elhoseiny \(2022\)Value memory graph: a graph\-structured world model for offline reinforcement learning\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.27780#S2.SS0.SSS0.Px3.p1.1)\.
## Appendix AProofs and mathematical background
### A\.1Proof of Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1)
###### Proof\.
Since the graph is fixed in the FE regime, we haveAt≡AA\_\{t\}\\equiv Afor all rollout steps\. LetFθX\(⋅,A,a\)F\_\{\\theta\}^\{X\}\(\\cdot,A,a\)denote the learned node\-transition map andFX⋆\(⋅,A,a\)F^\{X\\star\}\(\\cdot,A,a\)denote the true node\-transition map\. By Assumption[2](https://arxiv.org/html/2606.27780#Thmassumption2), for any input\(X,A,a\)\(X,A,a\)in the compact state space,
‖FθX\(X,A,a\)−FX⋆\(X,A,a\)‖F≤ϵX\.\\\|F\_\{\\theta\}^\{X\}\(X,A,a\)\-F^\{X\\star\}\(X,A,a\)\\\|\_\{F\}\\leq\\epsilon\_\{X\}\.\(A\.1\)
LetX^k\\hat\{X\}\_\{k\}andXkX\_\{k\}be the predicted and true node features at rollout stepkk\. Then
ek\+1X\\displaystyle e^\{X\}\_\{k\+1\}=‖X^k\+1−Xk\+1‖F\\displaystyle=\\\|\\hat\{X\}\_\{k\+1\}\-X\_\{k\+1\}\\\|\_\{F\}\(A\.2\)=‖FθX\(X^k,A,ak\)−FX⋆\(Xk,A,ak\)‖F\.\\displaystyle=\\\|F\_\{\\theta\}^\{X\}\(\\hat\{X\}\_\{k\},A,a\_\{k\}\)\-F^\{X\\star\}\(X\_\{k\},A,a\_\{k\}\)\\\|\_\{F\}\.Adding and subtractingFθX\(Xk,A,ak\)F\_\{\\theta\}^\{X\}\(X\_\{k\},A,a\_\{k\}\)and applying the triangle inequality gives
ek\+1X\\displaystyle e^\{X\}\_\{k\+1\}≤‖FθX\(X^k,A,ak\)−FθX\(Xk,A,ak\)‖F\\displaystyle\\leq\\\|F\_\{\\theta\}^\{X\}\(\\hat\{X\}\_\{k\},A,a\_\{k\}\)\-F\_\{\\theta\}^\{X\}\(X\_\{k\},A,a\_\{k\}\)\\\|\_\{F\}\(A\.3\)\+‖FθX\(Xk,A,ak\)−FX⋆\(Xk,A,ak\)‖F\.\\displaystyle\+\\\|F\_\{\\theta\}^\{X\}\(X\_\{k\},A,a\_\{k\}\)\-F^\{X\\star\}\(X\_\{k\},A,a\_\{k\}\)\\\|\_\{F\}\.The second term is bounded byϵX\\epsilon\_\{X\}from Equation \([A\.1](https://arxiv.org/html/2606.27780#A1.E1)\)\. For the first term, Assumption[3](https://arxiv.org/html/2606.27780#Thmassumption3)gives a message\-passing node transition with coordinate\-wise Lipschitz activationσ\\sigma:
FθX\(X,A,a\)=σ\(A^XW1\+Ua𝟏⊤\)W2\.F\_\{\\theta\}^\{X\}\(X,A,a\)=\\sigma\(\\hat\{A\}XW\_\{1\}\+Ua\\mathbf\{1\}^\{\\top\}\)W\_\{2\}\.\(A\.4\)For a multi\-layer message\-passing model, repeated use of the Lipschitz property ofσ\\sigmaand submultiplicativity of operator norms gives
‖FθX\(X^k,A,ak\)−FθX\(Xk,A,ak\)‖F\\displaystyle\\\|F\_\{\\theta\}^\{X\}\(\\hat\{X\}\_\{k\},A,a\_\{k\}\)\-F\_\{\\theta\}^\{X\}\(X\_\{k\},A,a\_\{k\}\)\\\|\_\{F\}\(A\.5\)≤Lσ‖A‖2∏ℓ‖Wℓ‖2⋅‖X^k−Xk‖F\.\\displaystyle\\leq L\_\{\\sigma\}\\\|A\\\|\_\{2\}\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\\cdot\\\|\\hat\{X\}\_\{k\}\-X\_\{k\}\\\|\_\{F\}\.For symmetric adjacency matrices,‖A‖2=ρ\(A\)\\\|A\\\|\_\{2\}=\\rho\(A\)\. Hence,
‖FθX\(X^k,A,ak\)−FθX\(Xk,A,ak\)‖F≤LXekX,\\displaystyle\\\|F\_\{\\theta\}^\{X\}\(\\hat\{X\}\_\{k\},A,a\_\{k\}\)\-F\_\{\\theta\}^\{X\}\(X\_\{k\},A,a\_\{k\}\)\\\|\_\{F\}\\leq L\_\{X\}e^\{X\}\_\{k\},\(A\.6\)LX=Lσρ\(A\)∏ℓ‖Wℓ‖2\.\\displaystyle L\_\{X\}=L\_\{\\sigma\}\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\.Combining Equations \([A\.3](https://arxiv.org/html/2606.27780#A1.E3)\) and \([A\.6](https://arxiv.org/html/2606.27780#A1.E6)\), we obtain the affine recursion
ek\+1X≤LXekX\+ϵX\.e^\{X\}\_\{k\+1\}\\leq L\_\{X\}e^\{X\}\_\{k\}\+\\epsilon\_\{X\}\.\(A\.7\)Unrolling Equation \([A\.7](https://arxiv.org/html/2606.27780#A1.E7)\) forkksteps yields
ekX≤LXke0X\+ϵX∑i=0k−1LXi\.e^\{X\}\_\{k\}\\leq L\_\{X\}^\{k\}e^\{X\}\_\{0\}\+\\epsilon\_\{X\}\\sum\_\{i=0\}^\{k\-1\}L\_\{X\}^\{i\}\.\(A\.8\)IfLX≠1L\_\{X\}\\neq 1, the finite geometric sum gives
∑i=0k−1LXi=LXk−1LX−1\.\\sum\_\{i=0\}^\{k\-1\}L\_\{X\}^\{i\}=\\frac\{L\_\{X\}^\{k\}\-1\}\{L\_\{X\}\-1\}\.\(A\.9\)Therefore,
ekX≤LXke0X\+ϵXLXk−1LX−1\.e^\{X\}\_\{k\}\\leq L\_\{X\}^\{k\}e^\{X\}\_\{0\}\+\\epsilon\_\{X\}\\frac\{L\_\{X\}^\{k\}\-1\}\{L\_\{X\}\-1\}\.\(A\.10\)This proves Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1)\. WhenLX=1L\_\{X\}=1, the same recursion gives the limiting formekX≤e0X\+kϵXe^\{X\}\_\{k\}\\leq e^\{X\}\_\{0\}\+k\\epsilon\_\{X\}\. ∎
### A\.2Proof of Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2)
###### Proof\.
By assumption, the node and edge rollout errors satisfy
ek\+1X\\displaystyle e^\{X\}\_\{k\+1\}≤LXekX\+LAekA\+ϵX,\\displaystyle\\leq L\_\{X\}e^\{X\}\_\{k\}\+L\_\{A\}e^\{A\}\_\{k\}\+\\epsilon\_\{X\},\(A\.11\)ek\+1A\\displaystyle e^\{A\}\_\{k\+1\}≤MXekX\+MAekA\+ϵA\.\\displaystyle\\leq M\_\{X\}e^\{X\}\_\{k\}\+M\_\{A\}e^\{A\}\_\{k\}\+\\epsilon\_\{A\}\.Define
uk\\displaystyle u\_\{k\}=\(ekX,ekA\)⊤,\\displaystyle=\(e^\{X\}\_\{k\},e^\{A\}\_\{k\}\)^\{\\top\},\(A\.12\)B\\displaystyle B=\(LXLAMXMA\),\\displaystyle=\\begin\{pmatrix\}L\_\{X\}&L\_\{A\}\\\\ M\_\{X\}&M\_\{A\}\\end\{pmatrix\},ϵ\\displaystyle\\epsilon=\(ϵX,ϵA\)⊤\.\\displaystyle=\(\\epsilon\_\{X\},\\epsilon\_\{A\}\)^\{\\top\}\.Since all constants are nonnegative, Equation \([A\.11](https://arxiv.org/html/2606.27780#A1.E11)\) is equivalent to the componentwise vector inequality
uk\+1⪯Buk\+ϵ\.u\_\{k\+1\}\\preceq Bu\_\{k\}\+\\epsilon\.\(A\.13\)This proves Equation \([4\.14](https://arxiv.org/html/2606.27780#S4.E14)\)\.
It remains to computeρ\(B\)\\rho\(B\)\. The characteristic polynomial ofBBis
det\(λI−B\)\\displaystyle\\det\(\\lambda I\-B\)=λ2−\(LX\+MA\)λ\\displaystyle=\\lambda^\{2\}\-\(L\_\{X\}\+M\_\{A\}\)\\lambda\(A\.14\)\+\(LXMA−LAMX\)\.\\displaystyle\\quad\+\(L\_\{X\}M\_\{A\}\-L\_\{A\}M\_\{X\}\)\.Thus, the two eigenvalues are
λ±\\displaystyle\\lambda\_\{\\pm\}=12\[\(LX\+MA\)\\displaystyle=\\frac\{1\}\{2\}\\Big\[\(L\_\{X\}\+M\_\{A\}\)\(A\.15\)±\(LX−MA\)2\+4LAMX\]\.\\displaystyle\\quad\\pm\\sqrt\{\(L\_\{X\}\-M\_\{A\}\)^\{2\}\+4L\_\{A\}M\_\{X\}\}\\Big\]\.BecauseBBis nonnegative, Perron–Frobenius implies that its spectral radius is the larger eigenvalue\. Hence,
ρ\(B\)\\displaystyle\\rho\(B\)=λ\+\\displaystyle=\\lambda\_\{\+\}\(A\.16\)=12\[\(LX\+MA\)\\displaystyle=\\frac\{1\}\{2\}\\Big\[\(L\_\{X\}\+M\_\{A\}\)\+\(LX−MA\)2\+4LAMX\]\.\\displaystyle\\quad\+\\sqrt\{\(L\_\{X\}\-M\_\{A\}\)^\{2\}\+4L\_\{A\}M\_\{X\}\}\\Big\]\.This proves Equation \([4\.15](https://arxiv.org/html/2606.27780#S4.E15)\)\.
Finally, supposeLAMX\>0L\_\{A\}M\_\{X\}\>0\. Then
\(LX−MA\)2\+4LAMX\\displaystyle\(L\_\{X\}\-M\_\{A\}\)^\{2\}\+4L\_\{A\}M\_\{X\}\>\(LX−MA\)2,\\displaystyle\>\(L\_\{X\}\-M\_\{A\}\)^\{2\},\(A\.17\)\(LX−MA\)2\+4LAMX\\displaystyle\\sqrt\{\(L\_\{X\}\-M\_\{A\}\)^\{2\}\+4L\_\{A\}M\_\{X\}\}\>\|LX−MA\|\.\\displaystyle\>\|L\_\{X\}\-M\_\{A\}\|\.Substituting this inequality into Equation \([A\.16](https://arxiv.org/html/2606.27780#A1.E16)\) gives
ρ\(B\)\\displaystyle\\rho\(B\)\>12\[\(LX\+MA\)\+\|LX−MA\|\]\\displaystyle\>\\frac\{1\}\{2\}\\left\[\(L\_\{X\}\+M\_\{A\}\)\+\|L\_\{X\}\-M\_\{A\}\|\\right\]\(A\.18\)=max\(LX,MA\)\.\\displaystyle=\\max\(L\_\{X\},M\_\{A\}\)\.This proves Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2)\. ∎
### A\.3Proof of Lemma[3](https://arxiv.org/html/2606.27780#Thmtheorem3)
###### Proof\.
SinceBBis a nonnegative matrix, the Gershgorin circle theorem gives
ρ\(B\)≤max\(LX\+LA,MX\+MA\)\.\\rho\(B\)\\leq\\max\(L\_\{X\}\+L\_\{A\},M\_\{X\}\+M\_\{A\}\)\.\(A\.19\)In the signal\-dominant regime, the first\-row error term is controlled by the topology\-dependent message\-passing contribution plus a bounded non\-topological residual\. Thus,
LX\+LA\\displaystyle L\_\{X\}\+L\_\{A\}≤GEAF^\(G\)\+GEAF^\(G\)RX‖A‖2\\displaystyle\\leq\\widehat\{\\mathrm\{GEAF\}\}\(G\)\+\\widehat\{\\mathrm\{GEAF\}\}\(G\)\\frac\{R\_\{X\}\}\{\\\|A\\\|\_\{2\}\}\(A\.20\)=GEAF^\(G\)\(1\+RX‖A‖2\)\.\\displaystyle=\\widehat\{\\mathrm\{GEAF\}\}\(G\)\\left\(1\+\\frac\{R\_\{X\}\}\{\\\|A\\\|\_\{2\}\}\\right\)\.Under the same signal\-dominant condition, the second row is no larger than this bound, so combining Equations \([A\.19](https://arxiv.org/html/2606.27780#A1.E19)\) and \([A\.20](https://arxiv.org/html/2606.27780#A1.E20)\) yields
ρ\(B\)≤GEAF^\(G\)\(1\+RX‖A‖2\)\.\\rho\(B\)\\leq\\widehat\{\\mathrm\{GEAF\}\}\(G\)\\left\(1\+\\frac\{R\_\{X\}\}\{\\\|A\\\|\_\{2\}\}\\right\)\.\(A\.21\)Since‖A‖2≥RX\\\|A\\\|\_\{2\}\\geq R\_\{X\}, we have
1\+RX‖A‖2≤2\.1\+\\frac\{R\_\{X\}\}\{\\\|A\\\|\_\{2\}\}\\leq 2\.\(A\.22\)Therefore,
ρ\(B\)\\displaystyle\\rho\(B\)≤GEAF^\(G\)\(1\+RX‖A‖2\)\\displaystyle\\leq\\widehat\{\\mathrm\{GEAF\}\}\(G\)\\left\(1\+\\frac\{R\_\{X\}\}\{\\\|A\\\|\_\{2\}\}\\right\)\(A\.23\)≤2GEAF^\(G\)\.\\displaystyle\\leq 2\\widehat\{\\mathrm\{GEAF\}\}\(G\)\.This proves the bound in Equation \([4\.17](https://arxiv.org/html/2606.27780#S4.E17)\)\.
Finally, if model weights are fixed across topologies, then
GEAF^\(G\)=ρ\(A\)∏ℓ‖Wℓ‖2\.\\widehat\{\\mathrm\{GEAF\}\}\(G\)=\\rho\(A\)\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}\.\(A\.24\)The factor∏ℓ‖Wℓ‖2\\prod\_\{\\ell\}\\\|W\_\{\\ell\}\\\|\_\{2\}is constant across topologies\. Hence,GEAF^\(G\)\\widehat\{\\mathrm\{GEAF\}\}\(G\)is monotone inρ\(A\)\\rho\(A\), and the rank order ofGEAF^\(G\)\\widehat\{\\mathrm\{GEAF\}\}\(G\)matches the rank order ofρ\(A\)\\rho\(A\)\. This proves Lemma[3](https://arxiv.org/html/2606.27780#Thmtheorem3)\. ∎
### A\.4Proof of Lemma[4](https://arxiv.org/html/2606.27780#Thmtheorem4)
###### Proof\.
Letπ∗\\pi^\{\*\}be the policy selected by planning with the true dynamics, and letπ^θ\\hat\{\\pi\}\_\{\\theta\}be the policy selected by planning with the learned GWM\. A standard simulation\-lemma argument bounds the regret by the accumulated reward error along the planning horizon:
J∗\(π∗\)−J∗\(π^θ\)\\displaystyle J^\{\*\}\(\\pi^\{\*\}\)\-J^\{\*\}\(\\hat\{\\pi\}\_\{\\theta\}\)≤2∑k=0H−1γk\\displaystyle\\leq 2\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\(A\.25\)×\|R⋆\(Gk,ak\)\\displaystyle\\quad\\times\\Big\|R^\{\\star\}\(G\_\{k\},a\_\{k\}\)−Rθ\(G^k,ak\)\|\.\\displaystyle\\qquad\-R\_\{\\theta\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)\\Big\|\.Adding and subtractingR⋆\(G^k,ak\)R^\{\\star\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)gives
\|R⋆\(Gk,ak\)−Rθ\(G^k,ak\)\|\\displaystyle\\left\|R^\{\\star\}\(G\_\{k\},a\_\{k\}\)\-R\_\{\\theta\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)\\right\|\(A\.26\)≤\|R⋆\(Gk,ak\)−R⋆\(G^k,ak\)\|\\displaystyle\\leq\\left\|R^\{\\star\}\(G\_\{k\},a\_\{k\}\)\-R^\{\\star\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)\\right\|\+\|R⋆\(G^k,ak\)−Rθ\(G^k,ak\)\|\.\\displaystyle\+\\left\|R^\{\\star\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)\-R\_\{\\theta\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)\\right\|\.By Assumption[5](https://arxiv.org/html/2606.27780#Thmassumption5), the first term is bounded by the Lipschitz reward condition, and the second term is bounded byϵR\\epsilon\_\{R\}:
\|R⋆\(Gk,ak\)−Rθ\(G^k,ak\)\|\\displaystyle\\left\|R^\{\\star\}\(G\_\{k\},a\_\{k\}\)\-R\_\{\\theta\}\(\\hat\{G\}\_\{k\},a\_\{k\}\)\\right\|\(A\.27\)≤LRdG\(Gk,G^k\)\+ϵR\.\\displaystyle\\leq L\_\{R\}d\_\{G\}\(G\_\{k\},\\hat\{G\}\_\{k\}\)\+\\epsilon\_\{R\}\.Using Equation \([4\.18](https://arxiv.org/html/2606.27780#S4.E18)\),
dG\(Gk,G^k\)=c⊤uk≤‖c‖2‖uk‖2\.d\_\{G\}\(G\_\{k\},\\hat\{G\}\_\{k\}\)=c^\{\\top\}u\_\{k\}\\leq\\\|c\\\|\_\{2\}\\\|u\_\{k\}\\\|\_\{2\}\.\(A\.28\)From Lemma[2](https://arxiv.org/html/2606.27780#Thmtheorem2), the joint error recursion satisfies
uk\+1⪯Buk\+ϵ\.u\_\{k\+1\}\\preceq Bu\_\{k\}\+\\epsilon\.\(A\.29\)Letϵ~\\tilde\{\\epsilon\}denote a uniform scale for the one\-step error vector\. Standard matrix\-recursion bounds give
‖uk‖2≤κ\(B\)ϵ~∑i=0k−1ρ\(B\)i\.\\\|u\_\{k\}\\\|\_\{2\}\\leq\\kappa\(B\)\\tilde\{\\epsilon\}\\sum\_\{i=0\}^\{k\-1\}\\rho\(B\)^\{i\}\.\(A\.30\)Substituting Equations \([A\.27](https://arxiv.org/html/2606.27780#A1.E27)\)–\([A\.30](https://arxiv.org/html/2606.27780#A1.E30)\) into Equation \([A\.25](https://arxiv.org/html/2606.27780#A1.E25)\) yields
J∗\(π∗\)−J∗\(π^θ\)\\displaystyle J^\{\*\}\(\\pi^\{\*\}\)\-J^\{\*\}\(\\hat\{\\pi\}\_\{\\theta\}\)≤2LR‖c‖2κ\(B\)ϵ~\\displaystyle\\leq 2L\_\{R\}\\\|c\\\|\_\{2\}\\kappa\(B\)\\tilde\{\\epsilon\}\(A\.31\)×∑k=0H−1γk∑i=0k−1ρ\(B\)i\\displaystyle\\quad\\times\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\sum\_\{i=0\}^\{k\-1\}\\rho\(B\)^\{i\}\+2ϵR∑k=0H−1γk\.\\displaystyle\\quad\+2\\epsilon\_\{R\}\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\.The reward\-model error term is
∑k=0H−1γk=1−γH1−γ\.\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}=\\frac\{1\-\\gamma^\{H\}\}\{1\-\\gamma\}\.\(A\.32\)For the rollout\-error term,
∑k=0H−1γk∑i=0k−1ρi\\displaystyle\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\sum\_\{i=0\}^\{k\-1\}\\rho^\{i\}=∑k=0H−1γkρk−1ρ−1\\displaystyle=\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\frac\{\\rho^\{k\}\-1\}\{\\rho\-1\}\(A\.33\)=1ρ−1\[∑k=0H−1\(γρ\)k−∑k=0H−1γk\]\\displaystyle=\\frac\{1\}\{\\rho\-1\}\\left\[\\sum\_\{k=0\}^\{H\-1\}\(\\gamma\\rho\)^\{k\}\-\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\right\]=1ρ−1\[1−\(γρ\)H1−γρ−1−γH1−γ\]\.\\displaystyle=\\frac\{1\}\{\\rho\-1\}\\left\[\\frac\{1\-\(\\gamma\\rho\)^\{H\}\}\{1\-\\gamma\\rho\}\-\\frac\{1\-\\gamma^\{H\}\}\{1\-\\gamma\}\\right\]\.Thus,
ΦH\(γ,ρ\)\\displaystyle\\Phi\_\{H\}\(\\gamma,\\rho\)=1ρ−1\[1−\(γρ\)H1−γρ−1−γH1−γ\]\.\\displaystyle=\\frac\{1\}\{\\rho\-1\}\\left\[\\frac\{1\-\(\\gamma\\rho\)^\{H\}\}\{1\-\\gamma\\rho\}\-\\frac\{1\-\\gamma^\{H\}\}\{1\-\\gamma\}\\right\]\.\(A\.34\)Settingρ=ρ\(B\)\\rho=\\rho\(B\)in Equation \([A\.31](https://arxiv.org/html/2606.27780#A1.E31)\) gives
J∗\(π∗\)−J∗\(π^θ\)\\displaystyle J^\{\*\}\(\\pi^\{\*\}\)\-J^\{\*\}\(\\hat\{\\pi\}\_\{\\theta\}\)≤2LR‖c‖2κ\(B\)ϵ~\\displaystyle\\leq 2L\_\{R\}\\\|c\\\|\_\{2\}\\kappa\(B\)\\tilde\{\\epsilon\}\(A\.35\)×ΦH\(γ,ρ\(B\)\)\\displaystyle\\quad\\times\\Phi\_\{H\}\(\\gamma,\\rho\(B\)\)\+2ϵR1−γH1−γ\.\\displaystyle\\quad\+2\\epsilon\_\{R\}\\frac\{1\-\\gamma^\{H\}\}\{1\-\\gamma\}\.Ifγρ\(B\)\>1\\gamma\\rho\(B\)\>1, the term\(γρ\(B\)\)H\(\\gamma\\rho\(B\)\)^\{H\}appears inΦH\(γ,ρ\(B\)\)\\Phi\_\{H\}\(\\gamma,\\rho\(B\)\), so the rollout\-error contribution grows faster than linearly inHH\. This proves Lemma[4](https://arxiv.org/html/2606.27780#Thmtheorem4)\. ∎
### A\.5Proof of Corollary[5](https://arxiv.org/html/2606.27780#Thmtheorem5)
###### Proof\.
From Lemma[1](https://arxiv.org/html/2606.27780#Thmtheorem1), the FE rollout error satisfies
ekX≤LXke0X\+ϵXLXk−1LX−1\.e^\{X\}\_\{k\}\\leq L\_\{X\}^\{k\}e^\{X\}\_\{0\}\+\\epsilon\_\{X\}\\frac\{L\_\{X\}^\{k\}\-1\}\{L\_\{X\}\-1\}\.\(A\.36\)When the rollout is not dominated by the constant error floor, the leading dependence onkkis proportional toLXkL\_\{X\}^\{k\}\. Thus, for some constantC\>0C\>0independent ofkk,
ekX≈CLXk\.e^\{X\}\_\{k\}\\approx CL\_\{X\}^\{k\}\.\(A\.37\)Taking logarithms gives
logekX≈logC\+klogLX\.\\log e^\{X\}\_\{k\}\\approx\\log C\+k\\log L\_\{X\}\.\(A\.38\)Therefore,
∂logekX∂k→logLX\.\\frac\{\\partial\\log e^\{X\}\_\{k\}\}\{\\partial k\}\\to\\log L\_\{X\}\.\(A\.39\)This proves Corollary[5](https://arxiv.org/html/2606.27780#Thmtheorem5)\. ∎
## Appendix BAuxiliary latent world\-model bounds
### B\.1Latent world\-model setup
This appendix records the scalar latent\-world\-model bounds that underlie the graph\-valued results in the main text\. They are not a replacement for Lemmas[1](https://arxiv.org/html/2606.27780#Thmtheorem1)–[4](https://arxiv.org/html/2606.27780#Thmtheorem4); rather, they show how the familiar compounding\-error argument for latent dynamicsHafneret al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib20)\); Chuaet al\.\([2018](https://arxiv.org/html/2606.27780#bib.bib70)\); Janneret al\.\([2019](https://arxiv.org/html/2606.27780#bib.bib4)\)becomes the graph\-specific operator analysis used in Section[4](https://arxiv.org/html/2606.27780#S4)\. We consider a sequential decision\-making setting in which an agent interacts with a partially observed environment over discrete time steps\. At each time steptt, the agent receives an observationot∈𝒪o\_\{t\}\\in\\mathcal\{O\}, selects an actionat∈𝒜a\_\{t\}\\in\\mathcal\{A\}, and receives a scalar rewardrt∈ℝr\_\{t\}\\in\\mathbb\{R\}\. Since the true environment state is generally not directly observable, the agent must infer a compact latent state from its interaction history\. Let
ht=\(o1,a1,o2,a2,…,ot\)h\_\{t\}=\(o\_\{1\},a\_\{1\},o\_\{2\},a\_\{2\},\\ldots,o\_\{t\}\)\(B\.1\)denote the observable history up to timett\. A latent world model learns an internal state representation
zt=eθ\(ht\),z\_\{t\}=e\_\{\\theta\}\(h\_\{t\}\),\(B\.2\)wherezt∈𝒵z\_\{t\}\\in\\mathcal\{Z\}is a latent state andeθe\_\{\\theta\}is an encoder parameterized byθ\\theta\.
The purpose of the latent world model is not only to encode the current observation, but also to predict how the latent state evolves under actions\. We define a latent world model as
𝒲θ=\(\\displaystyle\\mathcal\{W\}\_\{\\theta\}=\(eθ,pθ\(zt\+1∣zt,at\),\\displaystyle e\_\{\\theta\},\\,p\_\{\\theta\}\(z\_\{t\+1\}\\mid z\_\{t\},a\_\{t\}\),\(B\.3\)pθ\(ot∣zt\),pθ\(rt∣zt,at\)\),\\displaystyle p\_\{\\theta\}\(o\_\{t\}\\mid z\_\{t\}\),\\,p\_\{\\theta\}\(r\_\{t\}\\mid z\_\{t\},a\_\{t\}\)\),whereeθe\_\{\\theta\}maps observation histories to latent states,pθ\(zt\+1∣zt,at\)p\_\{\\theta\}\(z\_\{t\+1\}\\mid z\_\{t\},a\_\{t\}\)is the latent transition model,pθ\(ot∣zt\)p\_\{\\theta\}\(o\_\{t\}\\mid z\_\{t\}\)is the observation reconstruction or prediction model, andpθ\(rt∣zt,at\)p\_\{\\theta\}\(r\_\{t\}\\mid z\_\{t\},a\_\{t\}\)is the reward prediction model\.
Given a trajectory
τ=\(o1,a1,r1,o2,a2,r2,…,oT\),\\tau=\(o\_\{1\},a\_\{1\},r\_\{1\},o\_\{2\},a\_\{2\},r\_\{2\},\\ldots,o\_\{T\}\),\(B\.4\)the latent world model induces the following factorization:
pθ\(o1:T,r1:T,z1:T∣a1:T\)=pθ\(z1∣o1\)\\displaystyle p\_\{\\theta\}\(o\_\{1:T\},r\_\{1:T\},z\_\{1:T\}\\mid a\_\{1:T\}\)=p\_\{\\theta\}\(z\_\{1\}\\mid o\_\{1\}\)\(B\.5\)×∏t=1Tpθ\(ot∣zt\)pθ\(rt∣zt,at\)pθ\(zt\+1∣zt,at\)\.\\displaystyle\\times\\prod\_\{t=1\}^\{T\}p\_\{\\theta\}\(o\_\{t\}\\mid z\_\{t\}\)p\_\{\\theta\}\(r\_\{t\}\\mid z\_\{t\},a\_\{t\}\)p\_\{\\theta\}\(z\_\{t\+1\}\\mid z\_\{t\},a\_\{t\}\)\.This factorization shows that the model compresses the observable history into latent states while preserving the information needed to predict future observations and rewards\.
The learning objective is to fit the latent world model to trajectories collected from the environment\. A common training objective is the negative log\-likelihood
ℒWM\(θ\)=−∑t=1T\[logpθ\(ot∣zt\)\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{WM\}\}\(\\theta\)=\-\\sum\_\{t=1\}^\{T\}\\Big\[\\log p\_\{\\theta\}\(o\_\{t\}\\mid z\_\{t\}\)\(B\.6\)\+logpθ\(rt∣zt,at\)\+logpθ\(zt\+1∣zt,at\)\],\\displaystyle\+\\log p\_\{\\theta\}\(r\_\{t\}\\mid z\_\{t\},a\_\{t\}\)\+\\log p\_\{\\theta\}\(z\_\{t\+1\}\\mid z\_\{t\},a\_\{t\}\)\\Big\],wherezt=eθ\(ht\)z\_\{t\}=e\_\{\\theta\}\(h\_\{t\}\)\. In variational latent world models, the latent transition term can be replaced by an evidence lower bound that includes an inference modelqθ\(zt∣ht\)q\_\{\\theta\}\(z\_\{t\}\\mid h\_\{t\}\)and a prior transition modelpθ\(zt∣zt−1,at−1\)p\_\{\\theta\}\(z\_\{t\}\\mid z\_\{t\-1\},a\_\{t\-1\}\)\.
After training, the latent world model can be used for multi\-step imagination or planning\. Starting from an inferred latent stateztz\_\{t\}, the model recursively predicts future latent states under a candidate action sequence\(at,at\+1,…,at\+H−1\)\(a\_\{t\},a\_\{t\+1\},\\ldots,a\_\{t\+H\-1\}\):
z^t\+k\+1∼pθ\(z^t\+k\+1∣z^t\+k,at\+k\),\\hat\{z\}\_\{t\+k\+1\}\\sim p\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\+1\}\\mid\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\),\(B\.7\)k=0,1,…,H−1k=0,1,\\ldots,H\-1, withz^t=zt\\hat\{z\}\_\{t\}=z\_\{t\}\. The predicted return of this imagined trajectory is
J^θ\(at:t\+H−1∣zt\)=𝔼𝒲θ\[∑k=0H−1γkr^t\+k\],\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)=\\mathbb\{E\}\_\{\\mathcal\{W\}\_\{\\theta\}\}\\left\[\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\hat\{r\}\_\{t\+k\}\\right\],\(B\.8\)where
r^t\+k∼pθ\(r^t\+k∣z^t\+k,at\+k\)\.\\hat\{r\}\_\{t\+k\}\\sim p\_\{\\theta\}\(\\hat\{r\}\_\{t\+k\}\\mid\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\.\(B\.9\)The planning problem can then be written as
at:t\+H−1⋆=argmaxat:t\+H−1J^θ\(at:t\+H−1∣zt\)\.a\_\{t:t\+H\-1\}^\{\\star\}=\\arg\\max\_\{a\_\{t:t\+H\-1\}\}\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\.\(B\.10\)
Thus, a latent world model is an action\-conditioned predictive representation of the environment\. It learns a latent state space in which the agent can simulate future trajectories, estimate their rewards, and select actions according to predicted long\-term outcomes\.
### B\.2Error Propagation in Latent World Models
We first study how a one\-step latent transition error propagates through multi\-step imagination\. We consider the deterministic mean dynamics induced by the latent world model\. Let
z^t\+1=Fθ\(z^t,at\)\\hat\{z\}\_\{t\+1\}=F\_\{\\theta\}\(\\hat\{z\}\_\{t\},a\_\{t\}\)\(B\.11\)denote the learned latent transition, and let
zt\+1=F⋆\(zt,at\)z\_\{t\+1\}=F^\{\\star\}\(z\_\{t\},a\_\{t\}\)\(B\.12\)denote the latent dynamics induced by the true environment under the same latent representation\. Here,F⋆F^\{\\star\}is not required to be a physical\-state transition; it is the true transition map in the latent space\.
###### Assumption 8\(Uniform one\-step latent transition error\)\.
There exists a constantϵF≥0\\epsilon\_\{F\}\\geq 0such that
supz∈𝒵,a∈𝒜‖Fθ\(z,a\)−F⋆\(z,a\)‖≤ϵF\.\\sup\_\{z\\in\\mathcal\{Z\},a\\in\\mathcal\{A\}\}\\left\\\|F\_\{\\theta\}\(z,a\)\-F^\{\\star\}\(z,a\)\\right\\\|\\leq\\epsilon\_\{F\}\.\(B\.13\)
###### Assumption 9\(Lipschitz continuity of the learned latent transition\)\.
There exists a constantLF≥0L\_\{F\}\\geq 0such that, for anyz,z′∈𝒵z,z^\{\\prime\}\\in\\mathcal\{Z\}anda∈𝒜a\\in\\mathcal\{A\},
‖Fθ\(z,a\)−Fθ\(z′,a\)‖≤LF‖z−z′‖\.\\left\\\|F\_\{\\theta\}\(z,a\)\-F\_\{\\theta\}\(z^\{\\prime\},a\)\\right\\\|\\leq L\_\{F\}\\left\\\|z\-z^\{\\prime\}\\right\\\|\.\(B\.14\)
Given an action sequenceat:t\+H−1a\_\{t:t\+H\-1\}, define the true latent rollout and the imagined latent rollout as
zt\+k\+1\\displaystyle z\_\{t\+k\+1\}=F⋆\(zt\+k,at\+k\),\\displaystyle=F^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\),\(B\.15\)z^t\+k\+1\\displaystyle\\hat\{z\}\_\{t\+k\+1\}=Fθ\(z^t\+k,at\+k\),\\displaystyle=F\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\),fork=0,1,…,H−1k=0,1,\\ldots,H\-1\. Define thekk\-step rollout error by
ek=‖z^t\+k−zt\+k‖\.e\_\{k\}=\\left\\\|\\hat\{z\}\_\{t\+k\}\-z\_\{t\+k\}\\right\\\|\.\(B\.16\)
###### Theorem 6\(Multi\-step latent transition error propagation\)\.
Suppose Assumptions[8](https://arxiv.org/html/2606.27780#Thmassumption8)and[9](https://arxiv.org/html/2606.27780#Thmassumption9)hold\. Then, for any fixed action sequenceat:t\+H−1a\_\{t:t\+H\-1\}and anyk=0,1,…,Hk=0,1,\\ldots,H,
ek\\displaystyle e\_\{k\}≤LFke0\+ϵF∑i=0k−1LFi\.\\displaystyle\\leq L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\.\(B\.17\)Equivalently, ifLF≠1L\_\{F\}\\neq 1,
ek\\displaystyle e\_\{k\}≤LFke0\+ϵFLFk−1LF−1\\displaystyle\\leq L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\frac\{L\_\{F\}^\{k\}\-1\}\{L\_\{F\}\-1\}\(B\.18\)=O\(LFke0\+ϵFLFk\)\.\\displaystyle=O\(L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}L\_\{F\}^\{k\}\)\.IfLF=1L\_\{F\}=1, thenek≤e0\+kϵF=O\(e0\+kϵF\)e\_\{k\}\\leq e\_\{0\}\+k\\epsilon\_\{F\}=O\(e\_\{0\}\+k\\epsilon\_\{F\}\)\. In particular, if the initial latent state is exact, namelye0=0e\_\{0\}=0, then
ek≤ϵF∑i=0k−1LFi=\{O\(ϵF\),0≤LF<1,O\(kϵF\),LF=1,O\(ϵFLFk\),LF\>1\.e\_\{k\}\\leq\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}=\\begin\{cases\}O\(\\epsilon\_\{F\}\),&0\\leq L\_\{F\}<1,\\\\ O\(k\\epsilon\_\{F\}\),&L\_\{F\}=1,\\\\ O\(\\epsilon\_\{F\}L\_\{F\}^\{k\}\),&L\_\{F\}\>1\.\\end\{cases\}\(B\.19\)
###### Proof\.
For eachk=0,1,…,H−1k=0,1,\\ldots,H\-1, we have
ek\+1=‖Fθ\(z^t\+k,at\+k\)−F⋆\(zt\+k,at\+k\)‖\.e\_\{k\+1\}=\\left\\\|F\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-F^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\\\|\.\(B\.20\)Adding and subtractingFθ\(zt\+k,at\+k\)F\_\{\\theta\}\(z\_\{t\+k\},a\_\{t\+k\}\)gives
ek\+1≤‖Fθ\(z^t\+k,at\+k\)−Fθ\(zt\+k,at\+k\)‖\\displaystyle e\_\{k\+1\}\\leq\\left\\\|F\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-F\_\{\\theta\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\\\|\(B\.21\)\+‖Fθ\(zt\+k,at\+k\)−F⋆\(zt\+k,at\+k\)‖\.\\displaystyle\+\\left\\\|F\_\{\\theta\}\(z\_\{t\+k\},a\_\{t\+k\}\)\-F^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\\\|\.\(B\.22\)By Assumption[9](https://arxiv.org/html/2606.27780#Thmassumption9), the first term is bounded by
LF‖z^t\+k−zt\+k‖=LFek\.L\_\{F\}\\left\\\|\\hat\{z\}\_\{t\+k\}\-z\_\{t\+k\}\\right\\\|=L\_\{F\}e\_\{k\}\.\(B\.23\)By Assumption[8](https://arxiv.org/html/2606.27780#Thmassumption8), the second term is bounded byϵF\\epsilon\_\{F\}\. Therefore,
ek\+1≤LFek\+ϵF\.e\_\{k\+1\}\\leq L\_\{F\}e\_\{k\}\+\\epsilon\_\{F\}\.\(B\.24\)Unrolling this recursion of Equation \([B\.24](https://arxiv.org/html/2606.27780#A2.E24)\) yields
ek≤LFke0\+ϵF∑i=0k−1LFi\.e\_\{k\}\\leq L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\.\(B\.25\)WhenLF≠1L\_\{F\}\\neq 1, the geometric sum satisfies
∑i=0k−1LFi=LFk−1LF−1\.\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}=\\frac\{L\_\{F\}^\{k\}\-1\}\{L\_\{F\}\-1\}\.\(B\.26\)WhenLF=1L\_\{F\}=1, the recursion becomesek\+1≤ek\+ϵFe\_\{k\+1\}\\leq e\_\{k\}\+\\epsilon\_\{F\}, and henceek≤e0\+kϵFe\_\{k\}\\leq e\_\{0\}\+k\\epsilon\_\{F\}\. ∎
Theorem[6](https://arxiv.org/html/2606.27780#Thmtheorem6)shows that even a uniformly bounded one\-step transition error can increase across a multi\-step imagined trajectory\. WhenLF<1L\_\{F\}<1, the transition is contractive and the accumulated error remains controlled\. WhenLF=1L\_\{F\}=1, the error grows at most linearly in the planning horizon\. WhenLF\>1L\_\{F\}\>1, the error bound grows geometrically, which gives a formal explanation of compounding error in latent world models\.
### B\.3Reward Prediction and Return Error Bound
We next show that latent transition error induces reward prediction error and therefore biases the planning objective\. Let the learned reward model be
r^t\+k=Rθ\(z^t\+k,at\+k\),\\hat\{r\}\_\{t\+k\}=R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\),\(B\.27\)and let the true latent reward function be
rt\+k=R⋆\(zt\+k,at\+k\)\.r\_\{t\+k\}=R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\.\(B\.28\)For a fixed horizonHH, define the learned imagined return and the true return under the same action sequence as
J^θ\(at:t\+H−1∣z^t\)=∑k=0H−1γkRθ\(z^t\+k,at\+k\)\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid\\hat\{z\}\_\{t\}\)=\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\(B\.29\)and
J⋆\(at:t\+H−1∣zt\)=∑k=0H−1γkR⋆\(zt\+k,at\+k\)\.J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)=\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\.\(B\.30\)
###### Assumption 10\(Uniform one\-step reward model error\)\.
There exists a constantϵR≥0\\epsilon\_\{R\}\\geq 0such that
supz∈𝒵,a∈𝒜\|Rθ\(z,a\)−R⋆\(z,a\)\|≤ϵR\.\\sup\_\{z\\in\\mathcal\{Z\},a\\in\\mathcal\{A\}\}\\left\|R\_\{\\theta\}\(z,a\)\-R^\{\\star\}\(z,a\)\\right\|\\leq\\epsilon\_\{R\}\.\(B\.31\)
The previous theorem bounds the deviation of the imagined latent trajectory from the true latent trajectory\. We now show that this latent rollout error directly induces reward prediction error, because the reward model is evaluated on the imagined latent states rather than on the true latent states\.
###### Assumption 11\(Lipschitz continuity of the learned reward model\)\.
There exists a constantLR≥0L\_\{R\}\\geq 0such that, for anyz,z′∈𝒵z,z^\{\\prime\}\\in\\mathcal\{Z\}anda∈𝒜a\\in\\mathcal\{A\},
\|Rθ\(z,a\)−Rθ\(z′,a\)\|≤LR‖z−z′‖\.\\left\|R\_\{\\theta\}\(z,a\)\-R\_\{\\theta\}\(z^\{\\prime\},a\)\\right\|\\leq L\_\{R\}\\left\\\|z\-z^\{\\prime\}\\right\\\|\.\(B\.32\)
###### Proposition 7\(Reward prediction error bound\)\.
Suppose Assumptions[8](https://arxiv.org/html/2606.27780#Thmassumption8),[9](https://arxiv.org/html/2606.27780#Thmassumption9),[10](https://arxiv.org/html/2606.27780#Thmassumption10), and[11](https://arxiv.org/html/2606.27780#Thmassumption11)hold\. For any fixed action sequenceat:t\+H−1a\_\{t:t\+H\-1\}and anyk=0,1,…,H−1k=0,1,\\ldots,H\-1, define
ΔkR=\|Rθ\(z^t\+k,at\+k\)−R⋆\(zt\+k,at\+k\)\|\.\\Delta^\{R\}\_\{k\}=\\left\|R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\|\.\(B\.33\)Then
ΔkR≤LR\(LFke0\+ϵF∑i=0k−1LFi\)\+ϵR\.\\displaystyle\\Delta^\{R\}\_\{k\}\\leq L\_\{R\}\\left\(L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\\right\)\+\\epsilon\_\{R\}\.\(B\.34\)In particular, ife0=0e\_\{0\}=0, then
ΔkR≤LRϵF∑i=0k−1LFi\+ϵR\.\\Delta^\{R\}\_\{k\}\\leq L\_\{R\}\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\+\\epsilon\_\{R\}\.\(B\.35\)
Proposition[7](https://arxiv.org/html/2606.27780#Thmtheorem7)separates the reward prediction error into two sources\. The first term comes from latent state mismatch and scales with the transition error accumulated through the rollout\. The second term,ϵR\\epsilon\_\{R\}, is the intrinsic one\-step approximation error of the reward model\. Therefore, even when the reward model is uniformly accurate at one step, the reward estimate can still degrade over a long imagined trajectory because the latent inputs themselves become increasingly biased\.
###### Proof\.
For anyk=0,1,…,H−1k=0,1,\\ldots,H\-1, adding and subtractingRθ\(zt\+k,at\+k\)R\_\{\\theta\}\(z\_\{t\+k\},a\_\{t\+k\}\)gives
\|Rθ\(z^t\+k,at\+k\)−R⋆\(zt\+k,at\+k\)\|\\displaystyle\\left\|R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\|≤\|Rθ\(z^t\+k,at\+k\)−Rθ\(zt\+k,at\+k\)\|\\displaystyle\\leq\\left\|R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-R\_\{\\theta\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\|\+\|Rθ\(zt\+k,at\+k\)−R⋆\(zt\+k,at\+k\)\|\.\\displaystyle\+\\left\|R\_\{\\theta\}\(z\_\{t\+k\},a\_\{t\+k\}\)\-R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\|\.\(B\.36\)By Assumption[11](https://arxiv.org/html/2606.27780#Thmassumption11), the first term is bounded by
LR‖z^t\+k−zt\+k‖=LRek\.L\_\{R\}\\left\\\|\\hat\{z\}\_\{t\+k\}\-z\_\{t\+k\}\\right\\\|=L\_\{R\}e\_\{k\}\.\(B\.37\)By Assumption[10](https://arxiv.org/html/2606.27780#Thmassumption10), the second term is bounded byϵR\\epsilon\_\{R\}\. Thus,
\|Rθ\(z^t\+k,at\+k\)−R⋆\(zt\+k,at\+k\)\|≤LRek\+ϵR\.\\left\|R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\|\\leq L\_\{R\}e\_\{k\}\+\\epsilon\_\{R\}\.\(B\.38\)Applying Theorem[6](https://arxiv.org/html/2606.27780#Thmtheorem6)gives
ek≤LFke0\+ϵF∑i=0k−1LFi\.e\_\{k\}\\leq L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\.\(B\.39\)Substituting this bound into the reward error inequality proves the result\. ∎
###### Proposition 8\(Multi\-step return error bound\)\.
Suppose Assumptions[8](https://arxiv.org/html/2606.27780#Thmassumption8),[9](https://arxiv.org/html/2606.27780#Thmassumption9),[10](https://arxiv.org/html/2606.27780#Thmassumption10), and[11](https://arxiv.org/html/2606.27780#Thmassumption11)hold\. Then, for any fixed action sequenceat:t\+H−1a\_\{t:t\+H\-1\},
\|J^θ\(at:t\+H−1∣z^t\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid\\hat\{z\}\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.40\)≤∑k=0H−1γk\[LR\(LFke0\+ϵF∑i=0k−1LFi\)\+ϵR\]\.\\displaystyle\\leq\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\left\[L\_\{R\}\\left\(L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\\right\)\+\\epsilon\_\{R\}\\right\]\.\(B\.41\)If the initial latent state is exact, namelye0=0e\_\{0\}=0, then
\|J^θ\(at:t\+H−1∣zt\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.42\)≤∑k=0H−1γk\(LRϵF∑i=0k−1LFi\+ϵR\)\.\\displaystyle\\leq\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\left\(L\_\{R\}\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\+\\epsilon\_\{R\}\\right\)\.Furthermore, whene0=0e\_\{0\}=0,LF=1L\_\{F\}=1, andγ=1\\gamma=1, the return error satisfies
\|J^θ\(at:t\+H−1∣zt\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.43\)≤LRϵFH\(H−1\)2\+ϵRH\.\\displaystyle\\leq\\frac\{L\_\{R\}\\epsilon\_\{F\}H\(H\-1\)\}\{2\}\+\\epsilon\_\{R\}H\.\(B\.44\)Consequently,
\|J^θ−J⋆\|=O\(LRϵFH2\+ϵRH\)\.\\left\|\\hat\{J\}\_\{\\theta\}\-J^\{\\star\}\\right\|=O\(L\_\{R\}\\epsilon\_\{F\}H^\{2\}\+\\epsilon\_\{R\}H\)\.\(B\.45\)
Proposition[8](https://arxiv.org/html/2606.27780#Thmtheorem8)shows that transition error affects planning through the cumulative return objective\. Whene0=0e\_\{0\}=0,LF=1L\_\{F\}=1, andγ=1\\gamma=1, the return error scales asO\(LRϵFH2\+ϵRH\)O\(L\_\{R\}\\epsilon\_\{F\}H^\{2\}\+\\epsilon\_\{R\}H\)\. Thus, a one\-step latent transition error of orderϵF\\epsilon\_\{F\}can lead to a quadratic dependence on the planning horizon, while the reward approximation error accumulates linearly\. This explains why a latent world model with small one\-step prediction error may still produce biased long\-horizon planning estimates\.
###### Proof\.
By definition,
\|J^θ\(at:t\+H−1∣z^t\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid\\hat\{z\}\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.46\)=\|∑k=0H−1γk\[Rθ\(z^t\+k,at\+k\)−R⋆\(zt\+k,at\+k\)\]\|\.\\displaystyle=\\left\|\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\left\[R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\]\\right\|\.\(B\.47\)Using the triangle inequality,
\|J^θ\(at:t\+H−1∣z^t\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid\\hat\{z\}\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|≤∑k=0H−1γk\|Rθ\(z^t\+k,at\+k\)−R⋆\(zt\+k,at\+k\)\|\.\\displaystyle\\leq\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\left\|R\_\{\\theta\}\(\\hat\{z\}\_\{t\+k\},a\_\{t\+k\}\)\-R^\{\\star\}\(z\_\{t\+k\},a\_\{t\+k\}\)\\right\|\.\(B\.48\)Applying Proposition[7](https://arxiv.org/html/2606.27780#Thmtheorem7)to each summand yields
\|J^θ\(at:t\+H−1∣z^t\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid\\hat\{z\}\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.49\)≤∑k=0H−1γk\[LR\(LFke0\+ϵF∑i=0k−1LFi\)\+ϵR\]\.\\displaystyle\\leq\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\left\[L\_\{R\}\\left\(L\_\{F\}^\{k\}e\_\{0\}\+\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\\right\)\+\\epsilon\_\{R\}\\right\]\.\(B\.50\)Ife0=0e\_\{0\}=0, this reduces to
\|J^θ\(at:t\+H−1∣zt\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.51\)≤∑k=0H−1γk\(LRϵF∑i=0k−1LFi\+ϵR\)\.\\displaystyle\\leq\\sum\_\{k=0\}^\{H\-1\}\\gamma^\{k\}\\left\(L\_\{R\}\\epsilon\_\{F\}\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}\+\\epsilon\_\{R\}\\right\)\.\(B\.52\)WhenLF=1L\_\{F\}=1, we have∑i=0k−1LFi=k\\sum\_\{i=0\}^\{k\-1\}L\_\{F\}^\{i\}=k\. When alsoγ=1\\gamma=1, the bound becomes
∑k=0H−1\(LRϵFk\+ϵR\)=LRϵF∑k=0H−1k\+ϵRH\.\\sum\_\{k=0\}^\{H\-1\}\\left\(L\_\{R\}\\epsilon\_\{F\}k\+\\epsilon\_\{R\}\\right\)=L\_\{R\}\\epsilon\_\{F\}\\sum\_\{k=0\}^\{H\-1\}k\+\\epsilon\_\{R\}H\.\(B\.53\)Since∑k=0H−1k=H\(H−1\)2\\sum\_\{k=0\}^\{H\-1\}k=\\frac\{H\(H\-1\)\}\{2\}, we obtain
\|J^θ\(at:t\+H−1∣zt\)−J⋆\(at:t\+H−1∣zt\)\|\\displaystyle\\left\|\\hat\{J\}\_\{\\theta\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\-J^\{\\star\}\(a\_\{t:t\+H\-1\}\\mid z\_\{t\}\)\\right\|\(B\.54\)≤LRϵFH\(H−1\)2\+ϵRH\.\\displaystyle\\leq\\frac\{L\_\{R\}\\epsilon\_\{F\}H\(H\-1\)\}\{2\}\+\\epsilon\_\{R\}H\.\(B\.55\)This proves the stated asymptotic rate\. ∎Similar Articles
Imagined Rollouts are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure
This paper diagnoses long-horizon failures in world models, attributing them to kinematic rather than dynamic imagination. The authors introduce a metric (iKCE) and show that imagined rollouts fail to capture dynamic regime changes even as policy rewards collapse.
GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation
This paper systematically studies world models for robotic policy evaluation, introduces the WMBench benchmark and GigaWorld-1 model, and shows that long-horizon rollout consistency is more critical than short-term visual realism.
GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning
This paper introduces GAVEL, a framework that uses graph world models to verify and repair long-horizon LLM planning for robotic tasks, significantly improving success rates and efficiency in simulations.
Beyond Static Graph World Models: Learning Stochastic Latent Dynamics over Evolving Topologies
The paper proposes the Graph Dynamics Model (GDM) for learning in graph-structured environments with evolving topologies and stochasticity, and introduces the Graph Distribution Distance (GDD) metric for evaluation.
Repair the Amplifier, Not the Symptom: Stable World-Model Correction for Agent Rollouts
This paper introduces WM-SAR, a world-model correction method for agent planning that repairs causal subgraphs rather than visible symptoms, achieving better stabilization under token budgets compared to standard LLM correctors.