Beyond Static Forecasting: Unleashing the Power of World Models for Mobile Traffic Extrapolation Xiaoqian Qi
arXiv:2604.08199v1 [cs.NI] 9 Apr 2026
Department of Electronic Engineering, BNRist, Tsinghua University Beijing, China [email protected]
Yue Wang
Yong Li
Department of Electronic Engineering, BNRist, Tsinghua University Beijing, China [email protected]
Department of Electronic Engineering, BNRist, Tsinghua University Beijing, China [email protected]
ABSTRACT Mobile traffic prediction is a fundamental yet challenging problem for wireless network planning and optimization. Existing models focus on learning static long-term temporal patterns in mobile traffic series, which limits their ability to capture the dynamics between mobile traffic and network parameter adjustments. In this paper, we propose MobiWM, a world model for mobile networks. Taking mobile traffic as the system state, MobiWM models the dynamics between the states and network parameter actions, including power, azimuth, mechanical tilt, and electrical tilt through a predictive backbone. It fuses multimodal environmental contexts, comprising both image and sequential data, with encoded actions, leveraging shared spatial semantics to enhance spatial understanding. Leveraging the capacity of world models to capture real-world operational dynamics, MobiWM supports unlimited-horizon rollout over continuous network-adjustment action trajectories, providing operators with an explorable counterfactual simulation environment for network planning and optimization. Extensive experiments on variable-parameter mobile traffic data covering 31,900 cells across 9 districts demonstrate that MobiWM achieves the best distributional fidelity across all evaluation scenarios, significantly outperforming existing traffic prediction baselines and representative world models. A downstream RL-based case study further validates MobiWM as a simulation environment for network optimization, establishing a new paradigm for digital twin-driven wireless network management.
KEYWORDS World model, mobile traffic prediction, multi-modal fusion.
1
Haoye Chai State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications Beijing, China [email protected]
INTRODUCTION
Mobile traffic prediction serves as a cornerstone for wireless network planning and optimization. Accurate forecasting of mobile traffic enables operators to proactively allocate radio resources, trigger load-balancing mechanisms, and schedule energy-saving strategies before congestion or service degradation occurs [29, 35]. However, mobile traffic exhibits intricate spatio-temporal dynamics shaped by heterogeneous urban activities, irregular event patterns,
and complex engineering configurations and topological correlations among densely deployed base stations (BS), making reliable prediction a persistently challenging problem. As mobile networks evolve toward 5G-Advanced and 6G, the expanding scale of network nodes imposes rigorous demands on resource allocation and optimization. Directly tuning parameters on live networks entails prohibitive costs and non-negligible risks of service disruption, rendering extensive trial-and-error infeasible in production environments. A more desirable paradigm is to construct high-fidelity digital twin environments that simulate network states in a virtual space, enabling operators to conduct counterfactual inference and exploratory what-if analyses before any configuration change is actually deployed. Existing approaches have made substantial progress in modeling the spatio-temporal patterns of mobile traffic [4, 5, 8, 19, 21, 31, 33, 34]. These methods attempt to capture diverse spatio-temporal dynamics through tailored architectural designs [4, 19, 31, 34], as well as to incorporate complex environmental correlations by fusing external context [6, 7, 32]. As network topology has a strong influence on the network operation, graph-based methods such as FedGTP [33] exploit spatial dependencies across distributed BSs under privacy constraints. State-space architectures like HiSTM [5] leverage hierarchical Mamba modules for efficient long-horizon cellular traffic forecasting. Meanwhile, emerging foundation-model paradigms, exemplified by MobiFM [8], attempt to unify heterogeneous mobile data types within a single pre-trained backbone, advancing scalability and generalization. Despite these advances, all existing methods share a fundamental limitation: they essentially learn static mappings from historical observations to future values, treating mobile traffic prediction as a pattern-fitting problem. In real-world network operations, however, traffic patterns are not determined solely by historical trends; they are continuously shaped by parameter tuning actions that operators perform to optimize coverage, capacity, and quality of service. In other words, mobile traffic is inherently coupled with network configuration parameters, and any change in these parameters can trigger substantial traffic redistribution across neighboring cells. Therefore, to enable counterfactual inference in digital twin environments, a critical capability is modeling the dynamic interplay
Conference’17, July 2017, Washington, DC, USA
between network parameters and traffic states, transforming traditional static modeling into dynamic extrapolation. World models, originally proposed as learned simulators that capture environment dynamics for model-based reinforcement learning [11], offer a principled framework to bridge this gap. By jointly modeling how actions transform states over time, world models internalize the causal structure of the underlying system and enable forward rollout, counterfactual inference, and policy optimization within a learned latent space. Recent breakthroughs have demonstrated the power of this paradigm across diverse domains. STORM [38] introduces Transformer-based stochastic world models that achieve state-of-the-art sample efficiency in Atari benchmarks. DreamerV3 [14] masters over 150 control tasks with a single set of hyperparameters by imagining future trajectories in a learned world model. These achievements offer innovative insights into the modeling requirements for mobile network state dynamics. The construction of a mobile network world model can enable an explorable, counterfactual environment for digital twin network optimization. In this paper, we propose MobiWM, adopting the world model paradigm for mobile traffic prediction. MobiWM formulates celllevel traffic as the system state and network parameter adjustments as actions, learning the dynamics between the time-varying parameter and network states through a predictive architecture. Figure 1 shows the difference between traditional static mobile traffic prediction models and MobiWM. Firstly, MobiWM adopts the world model paradigm to achieve cell-level mobile traffic prediction by modeling the transition distribution from historical states and parameter actions to future states. Multi-modal environmental context information, including Points of Interest (POI), Origin-Destination (OD) flows, and an image-modality facility map encompassing building distributions and BS layouts, constitutes the conditional space for this transition distribution. Secondly, MobiWM adopts an encoderdecoder architecture that jointly encodes and fuses the state space, action space, and multimodal urban context, while sharing spatial semantics across modalities to strengthen spatial understanding over the network graph. The designed Factorized Spatio-Temporal Blocks (FSTBlocks) can capture spatial topology features and temporal dependencies of mobile networks in a decoupled manner, which achieves spatio-temporally factorized dynamics modeling. Thirdly, we adopt a graph-batch prediction strategy with cell masking to enable efficient map-level global forecasting output and support unlimited-horizon rollout over continuous action trajectories. By contrast, traditional traffic prediction models only fit the distribution of long-term steady-state patterns for individual cell traffic conditioned on external factors, failing to achieve action-aware dynamic modeling. In summary, the main contributions of this paper are as follows: • We propose MobiWM, the first world model for mobile networks that learns the dynamics between network parameter adjustments and traffic variations through predictive modeling, providing operators with an explorable counterfactual simulation environment for network planning and optimization. • We design Factorized Spatio-Temporal Blocks (FSTBlocks) that decouple spatial topology and temporal dependency
Xiaoqian Qi, Haoye Chai, Yue Wang, and Yong Li
modeling via factorized attention. Encoding networks tailored for multi-modal environmental contexts, integrated with a learnable conditional gating mechanism, are employed to achieve multi-modal information fusion. • We conduct extensive evaluations on variable-parameter mobile traffic data spanning 9 districts from Nanchang City, China, and covering 31,900 cells. Experimental results demonstrate that MobiWM significantly outperforms existing traffic prediction and world model baselines in both accuracy and action-awareness, while exhibiting strong generalization to out-of-distribution actions and bursty events, validating its advantage in constructing digital twin-driven counterfactual environments for wireless network management.
2
RELATED WORK
Mobile Traffic Prediction. Mobile traffic prediction has been extensively studied, evolving from classical statistical models to modern deep learning architectures. Early approaches relied on statistical time-series methods [22, 26], which capture temporal autocorrelations but assume stationary linear dynamics, and traditional machine learning methods, such as Random Forests and XGBoost [10], which improved predictive capacity by incorporating hand-crafted features. The advent of deep learning brought substantial advances [5, 21, 24, 33, 34, 36]. Yang et al. [33] introduce FedGTP, which exploits inter-client spatial dependencies in a federated graph learning framework for privacy-preserving cellular traffic prediction. Bettouche et al. [5] propose HiSTM, which combines hierarchical Mamba-based state-space modules with dual spatial encoders for efficient long-horizon cellular traffic forecasting with significantly reduced parameter counts. Wang et al. [21] present MobiMixer, a lightweight multi-scale spatiotemporal mixing model that achieves competitive accuracy while substantially reducing computational cost. More recently, generative paradigms have been introduced for mobile traffic synthesis and prediction [6– 8, 37]. Chai et al. [7] develop STK-Diff, a spatio-temporal knowledgedriven diffusion model that constructs urban knowledge graphs to enable controllable mobile traffic generation. MobiFM [8] is constructed as a foundation model for mobile data, unifying heterogeneous mobile data types within a single diffusion-Transformer backbone to advance scalable prediction. However, none of these models addresses the dynamic coupling between network parameter adjustments and the resulting traffic variations, rendering them unable to answer counterfactual "what-if" questions that are essential for network planning and optimization. World Models. World models learn environment dynamics by modeling transitions between actions and states. Ha and Schmidhuber [11] first formalize this concept, combining a variational autoencoder with a recurrent network to train policies entirely within imagined rollouts. Hafner et al. introduce DreamerV1–V3 [12–14], proposing the Recurrent State-Space Model (RSSM) to learn latent dynamics and train actor-critic policies from imagination alone. More recently, Transformer-based world models have shown advantages in capturing long-range dependencies: TransDreamer [9] replaces recurrent dynamics with a Transformer State-Space Model,
Beyond Static Forecasting: Unleashing the Power of World Models for Mobile Traffic Extrapolation
Weekly/Monthly mobile traffic Static traffic patterns
𝜞𝒕 𝜞𝒕
𝜞𝒕 𝒕 𝜞𝒕 𝒕
𝜞𝒕+𝟏
𝒕
𝒕
…
𝜞𝒕+𝑷
𝒔𝒕
System state (cell traffic)
𝒂𝒕
Actions (BS parameters)
𝒄𝒏
Environment contexts
𝓔𝑺
𝓔𝑨
Static Mobile Traffic Prediction 𝜞𝒕−𝑯+𝟏 𝜞𝒕
…
𝒄𝟏
𝜞𝒕
𝒄𝟐
…
𝒄𝑵
𝜞𝒕 azimuth rotation
𝒕
Encoders
𝓔𝑪
rollout power surge
𝒕
…
𝜞𝒕
𝜞𝒕
Conference’17, July 2017, Washington, DC, USA
𝒔𝒕+𝟏
…
𝒔𝒕+𝑷
Prediction head
Mobile Network World Model (MobiWM) 𝒔𝒕−𝑯+𝟏
𝒔𝒕
…
𝓔𝑺
𝒂𝒕−𝑯+𝟏
𝒂𝒕+𝑷
…
𝒄𝟏
𝒄𝟐
POI
OD flow
𝓔𝑨
…
𝒄𝑵
𝓔𝑪
𝜞𝒕 𝜞𝒕
mask
𝜞𝒕 mask 𝒕 𝜞𝒕 mask mask 𝒕 mask 𝒕 𝒕
mobile traffic curves
𝒕 𝜞𝒕 mask
𝒑𝒕
POI
OD flow
facility map
cell-level traffic map
(a) Traditional static mobile traffic prediction models
𝜽𝒕 𝒕
𝒕
𝜸𝒕 𝒕
𝒕
[power, azimuth, mtilt, etilt]
facility map
(b) Proposed mobile network world model (MobiWM)
Figure 1: Comparison of the traditional static mobile traffic prediction models and the proposed mobile network world model (MobiWM). STORM [38] achieves state-of-the-art sample efficiency via stochastic Transformer architectures, and TD-MPC2 [16] predicts future states directly in latent space to avoid high-dimensional complexity. The key advantage of world models is providing virtual counterfactual environments without real-world interaction. However, their application in mobile communications remains largely unexplored. Zhao et al. [39] propose a conceptual architecture for edge intelligence in wireless networks, but it stays at the vision level without addressing large-scale multi-cell dynamics modeling. Existing world models lack customized modeling of network topology and spatio-temporal traffic dynamics, making the construction of a mobile network world model capable of learning network state-parameter dynamics an urgent open problem.
3
PRELIMINARY AND PROBLEM FORMULATION 3.1 Mobile Traffic-Network Parameter Dynamics The traffic load of a cell is governed by its radio coverage footprint, which is jointly shaped by the four antenna parameters. We briefly formalize this coupling to motivate the world model design. According to the 3GPP channel model [1], the received power at location r from cell 𝑣𝑖 can be denoted as 𝑖 𝑃rx (r)(dB) = 𝑃𝑡𝑖 + 𝐺 𝑖 (𝜑, 𝜗) − PL(𝑑 𝑖 (r)) + 𝑋𝜎 ,
(1)
where 𝐺 𝑖 is the directional antenna gain, PL(·) is the path loss, and 𝑋𝜎 is shadow fading. The gain 𝐺 𝑖 is determined by horizontal and vertical radiation patterns, in which the azimuth 𝜃 𝑡𝑖 steers the horizontal main lobe while the mechanical and electrical downtilts 𝑖 , 𝛾 𝑖 jointly control the vertical beam direction and thereby 𝛾 m,𝑡 e,𝑡 the effective cell radius. A user equipment is associated with the strongest-signal cell, so the aggregate traffic of cell 𝑣𝑖 is: ∫ h i 𝑗 𝑠𝑡𝑖 = 1 𝑖 = arg max 𝑃rx (r) 𝜌𝑡 (r) dr, (2) r∈ R
𝑗 ∈V
where 𝜌𝑡 (r) is the spatio-temporal traffic demand density. This formulation reveals that any parameter adjustment at one cell alters the received power landscape, triggers user re-association, and redistributes traffic across neighboring cells. The resulting dynamics are inherently non-local and nonlinear, due to the arg max cell selection and the squared angular attenuation in the antenna gain pattern. These properties, compounded by time-varying urban demand 𝜌𝑡 (r), render closed-form prediction intractable and motivate a data-driven world model to learn the action-state transition distribution directly from operational network data.
3.2
Problem Formulation
System Representation. We consider a cellular network with 𝑁 cells V = {𝑣 1, . . . , 𝑣 𝑁 } organized as a directed graph G = (V, E), where edges in E encode spatial adjacency. At each time step 𝑡, the system state s𝑡 ∈ R𝑁 records the traffic load Γ𝑡 of all cells, and the action a𝑡 ∈ R𝑁 ×4 collects the four tunable antenna parameters of 𝑖 , each cell: transmit power 𝑃𝑡𝑖 , azimuth 𝜃 𝑡𝑖 , mechanical downtilt 𝛾 m,𝑡 𝑖 and electrical downtilt 𝛾 e,𝑡 . The multimodal environmental context c = {cpoi, cod, cfac } comprises POI features characterizing land use, an OD flow matrix capturing mobility patterns, and image-modality facility maps encoding building distributions and BS layouts. Mobile Network World Modeling. The objective is to learn a parameterized dynamics model 𝑓Ω that captures the transition distribution from historical states and actions to future states, conditioned on environmental context: s𝑡 +1:𝑡 +𝑃 ∼ 𝑓Ω (s𝑡 +1:𝑡 +𝑃 | s𝑡 −𝐻 +1:𝑡 , a𝑡 −𝐻 +1:𝑡 , c),
(3)
where 𝐻 is the historical window length, 𝑃 is the prediction horizon, and Ω denotes the learnable parameters. The training objective is to find Ω∗ that minimizes the discrepancy between the predicted and true transition distributions: h i Ω ∗ = arg min E L s𝑡 +1:𝑡 +𝑃 , 𝑓Ω (s𝑡 +1:𝑡 +𝑃 | s𝑡 −𝐻 +1:𝑡 , a𝑡 −𝐻 +1:𝑡 , c) , Ω
(4)
Conference’17, July 2017, Washington, DC, USA
Feed-Forward +
Temporal Transformers
Topology Representation Network Mask FFN
spatial bias 𝑬𝒕𝒐𝒑𝒐
×
𝓟𝒓 cell coordinates
𝜽𝜽 ∆𝒚 ∆𝒙 ∆𝒙∆𝒚
𝓜
map batch
cell mask
+
map 𝓰𝟐 (5 cells)
…
map 𝓰𝑩 (4 cells)
distance Emb
cell b-1
bearing Emb
cell b-2
mask Emb
…
Spatial Transformers
map 𝓰𝟏 (8 cells)
Feature FFN
Shared FFN
+
Xiaoqian Qi, Haoye Chai, Yue Wang, and Yong Li
cell b-N State Embedding
𝜞𝒕
𝜽𝒕 𝒕 𝒕
𝒕
𝜸𝒆𝒕 𝜸𝒆𝒕 𝒕 𝜸𝒎𝒕 𝒕 𝒕
𝒕
Action Encoder
𝒑𝒕 𝒕 𝜸𝒎𝒕
[𝒂𝒕−𝑯+𝟏 , … , 𝒂𝒕 , … 𝒂𝒕+𝑷 ]
+
latent state
Visual Tokenizer
Facility Encoder
OD flow
Flow Tokenizer
OD Encoder POI Encoder
𝓖𝑭
+
𝓖𝑶
+
𝓖𝑷
+
𝓖𝑨
+
Multi-modal Context Fusion
Figure 2: Overview of the mobile network world model, MobiWM. It is composed of an encoder-decoder backbone for the state with factorized spatio-temporal blocks (FSTBlocks), conditioned on actions and multimodal environmental context. where L (·) is the prediction loss. Critically, the learned model supports unlimited-horizon rollout: the predicted state is fed back as input along with a new action sequence to autoregressively extend the trajectory. For the 𝑘-th rollout step: ŝ𝑡 +𝑘𝑃+1:𝑡 +(𝑘+1)𝑃 ∼ 𝑓Ω · | ŝ𝑡 +(𝑘 −1)𝑃 +1:𝑡 +𝑘𝑃 , a𝑡 +(𝑘 −1)𝑃 +1:𝑡 +𝑘𝑃 , c , (5) enabling trajectories of arbitrary length over continuous action sequences for counterfactual what-if inference.
4
MOBIWM: THE MOBILE NETWORK WORLD MODEL
We propose MobiWM, a world model for mobile networks that learns the dynamics between network parameter adjustments and traffic variations through an encoder-decoder architecture. The model is designed to capture the complex spatio-temporal dependencies and topological features of mobile networks while fusing multimodal environmental context. Figure 2 illustrates the overall architecture of MobiWM.
4.1
graph batch 𝓑 = {𝓰𝟏 , 𝓰𝟐 , … , 𝓰𝑩 }
cell mask 𝓜 = {𝒎𝟏 , 𝒎𝟐 , … , 𝒎𝑩 }
Figure 3: Diagram of the graph batch and cell mask for irregular network topology.
State Decoder
facility map
POI
Positional Encoding
FSTBlocks
State Encoder
𝜽𝒕
𝒑𝒕
Layer Norm
FSTBlocks
𝒕
[𝒔𝒕−𝑯+𝟏 , … , 𝒔𝒕−𝟏 , 𝒔𝒕 ]
FSTBlocks
mask mask
𝒕 𝒕
FSTBlocks
𝜞mask 𝒕 mask mask 𝒕 𝒕 𝒕 𝜞𝒕 𝜞 𝒕 𝜞mask 𝒕
State Encoder
𝜞𝒕
Factorized Spatial-temporal Block
Base Model
The base model of MobiWM is an encoder-decoder architecture for state representation. To enable efficient map-level global forecasting output, we adopt a graph-batch strategy with cell masking. 4.1.1 Graph Batch for Irregular Network Topology. Cells in a mobile network are deployed at irregular locations dictated by terrain, population density, and infrastructure constraints. To handle this irregular topology within a batch-parallel framework, we adopt
a graph-batch strategy inspired by mini-batch graph processing in scalable graph neural networks [15, 18]. Specifically, all 𝑁 cells on a district-level map are organized into a single graph ℊ = (𝓋, ℯ) and processed as one batch sample, where each cell serves as a node carrying its own state and action series. Accordingly, we introduce a binary cell mask M ∈ {0, 1}𝑁 ×𝑇 to indicate active cells within each map, allowing maps of different sizes to be batched together with zero-padding and masked attention. Figure 3 shows how the graph batch and cell mask are defined. 4.1.2 Encoder-Decoder State Representation. As illustrated in Figure 2, MobiWM utilizes a state encoder and a state decoder to predict the latent representations of future states. The historical state [s𝑡 −𝐻 +1, . . . , s𝑡 ] is first projected into a 𝑑-dimensional embedding space via a linear layer, yielding the state token sequence S0 ∈ R𝑁 ×𝐻 ×𝑑 . The state encoder E𝑆 , composed of 𝐿𝑒 stacked FSTBlocks, followed by Layer Normalization, compresses the historical observations into a latent space: Z𝐻 = E𝑆 S0 ∈ R𝑁 ×𝐻 ×𝑑 . (6) The state decoder D𝑆 , consisting of 𝐿𝑑 FSTBlocks, takes Z as input and generates the predicted future latent states: Z𝑃 = D𝑆 Z𝐻 ∈ R𝑁 ×𝑃 ×𝑑 . (7) By learning the transition dynamics in latent space rather than directly predicting in the high-dimensional raw space, we can significantly reduce modeling complexity and enhance generalization. 4.1.3 Action Encoding. We define four key parameters of the cell antenna as actions: transmission power, azimuth, mechanical tilt (mtilt), and electrical tilt (etilt). The action sequence spans both the historical and future horizons: [a𝑡 −𝐻 +1, . . . , a𝑡 , . . . , a𝑡 +𝑃 ]. Each percell action vector a𝑖𝑡 ∈ R4 is projected to the same 𝑑-dimensional space through the action encoder E𝐴 : h𝐴 = E𝐴 [a𝑡 −𝐻 +1, . . . , a𝑡 +𝑃 ] ∈ R𝑁 × (𝐻 +𝑃 ) ×𝑑 . (8) Including future actions in the encoding is essential for the world model to answer counterfactual queries: it allows the decoder to condition its predictions on hypothetical parameter adjustments that have not yet occurred. During rollout (Eq. (5)), operators can specify arbitrary future action trajectories to explore how different parameter configurations would reshape traffic distributions.
Beyond Static Forecasting: Unleashing the Power of World Models for Mobile Traffic Extrapolation
4.2
Factorized Spatio-Temporal Block
where Q, K, V are the query, key, and value projections of X.
Mobile traffic exhibits two distinct types of regularity: spatial patterns governed by network topology, where geographically adjacent or functionally related cells share correlated traffic profiles, and temporal patterns driven by periodic human activity, where traffic follows diurnal and weekly cycles. A naive joint spatio-temporal attention over 𝑁 cells and 𝑇 time steps incurs O (𝑁 2𝑇 2 ) complexity, which is prohibitive for large-scale networks. We therefore adopt a factorized design that decouples spatial and temporal modeling into separate attention stages, reducing the complexity to O (𝑁 2𝑇 +𝑁𝑇 2 ) while allowing each stage to specialize in capturing its respective dependency structure. 4.2.1 Factorized Attentions. Each FSTBlock consists of a Spatial Transformer, a Temporal Transformer, and a Feed-Forward Network (FFN), connected via residual additions. Given an input tensor X ∈ R𝑁 ×𝑇 ×𝑑 , the block first applies self-attention via Transformer across the spatial dimension: for each time step, the 𝑁 cell tokens are gathered and attended over, producing spatially contextualized representations. A topology-based spatial bias Etopo is injected into the attention scores to encode the network topological structure. The spatially enriched tokens are then passed to a temporal selfattention: for each cell independently, its 𝑇 temporal tokens attend to one another to capture periodic patterns and temporal trends. A position-wise FFN with residual connection produces the final block output. Formally: X′ = X + SpatialAttn(X), ′′
′
out
′′
(9) ′
X = X + TemporalAttn(X ), X
′′
= X + FFN(X ).
(10) (11)
The spatial-first ordering allows temporal attention to operate on spatially contextualized representations, enabling the model to distinguish, for example, whether a traffic surge is a local event at one cell or a network-wide pattern. 4.2.2 Topology-based Spatial Bias. Standard self-attention treats all cell pairs symmetrically and lacks awareness of the underlying network topology. We inject structural inductive bias into SpatialAttn through a learned Topology Representation Network (TRN). 𝑁 of all cells, Given the geographic coordinates P𝑟 = {(𝑥𝑖 , 𝑦𝑖 )}𝑖=1 we compute for each cell pair the relative displacement (Δ𝑥𝑖 𝑗 , Δ𝑦𝑖 𝑗 ), from which the Euclidean distance 𝑑𝑖 𝑗 and bearing angle 𝛼𝑖 𝑗 are derived. These geometric features are projected through a shared FFN to obtain a distance embedding and a bearing embedding, which are summed to form a pairwise feature representation e𝑖 𝑗 . Simultaneously, the cell mask M is similarly processed into a mask embedding e𝑖mask for each pair. The topology-aware spatial bias is 𝑗 computed via an outer product: mask Etopo [𝑖, 𝑗] = e𝑖feat ∈ R𝑛ℎ , 𝑗 ⊗ e𝑖 𝑗
Conference’17, July 2017, Washington, DC, USA
(12)
where 𝑛ℎ is the number of attention heads. This bias is then added to the attention logits of SpatialAttn, yielding the topology-aware spatial attention: QK⊤ SpatialAttn(X) = softmax √ + Etopo V, (13) 𝑑𝑘
4.3
Multi-modal Context Fusion
MobiWM fuses multimodal environmental context with the stateaction dynamics through three complementary mechanisms: modalityspecific encoding that extracts compact representations from heterogeneous data formats, shared positional encoding that aligns spatial semantics across modalities, and learnable gating that adaptively controls each modality’s contribution to the decoder. 4.3.1 Multi-modal Context Encoding. MobiWM encodes each environmental context: POI and OD flow in meshing-vector-modal, and facility map in image-modal, through a dedicated tokenizerencoder pipeline. For POI, the categorical distribution over 𝐾 POI types at each grid location forms a tensor cpoi ∈ R𝑆 ×𝑆 ×𝐾 , which is processed by a convolutional POI Encoder with multi-scale filters followed by global average pooling. 𝑆 denotes the number of pixels within the map area, corresponding to the adjustable spatial resolution. For OD flow, the time-varying origin-destination matrix cod ∈ R𝑆 ×𝑆 ×𝑇 captures population mobility patterns across the service area. A Flow Tokenizer applies spatial convolutions to extract structural features from each temporal slice, and an OD Encoder aggregates them via cross-attention with a learnable query token to produce a compact representation. For the facility map, the building distribution image cfac ∈ R𝐻 𝑓 ×𝑊𝑓 ×1 is encoded by a Visual Tokenizer into multi-scale convolutional feature maps. The Facility Encoder extracts representations at three granularities: a fine-grained map, a coarse-grained map aligned with the POI/OD grid, and a global feature vector: h𝑃 = POIEnc(cpoi ), h𝑂 = ODEnc(cod ), (h𝐹 𝑓 , h𝐹𝑐 , h𝐹 ) = FacEnc(cfac ); h𝐶 = {h𝑃 , h𝑂 , h𝐹 }; h𝑃 , h𝑂 , h𝐹 𝑓 , h𝐹𝑐 , h𝐹 ∈ R𝑑 . (14) 4.3.2 Positional Encoding. For temporal encoding, we extract the time-slot index 𝑞𝑡 ∈ {0, . . . , 𝑄−1} (e.g., 𝑄=96 for 15-minute intervals) and day-of-week index 𝑤𝑡 ∈ {0, . . . , 6} at each step 𝑡, and map them through learnable embedding tables: temp
p𝑡
= Eslot (𝑞𝑡 ) + Edow (𝑤𝑡 ) ∈ R𝑑 ,
∈ R𝑄 ×𝑑 and E
(15)
∈ R7×𝑑 . This encoding is added to
where Eslot dow both state and action tokens, providing an explicit reference for diurnal and weekly cycles. For spatial encoding, a key challenge is that cell-level states, grid-level POI/OD, and pixel-level facility maps share the same geographic space but differ in resolution and format. We design a shared Fourier-based positional encoding: for any coordinate r = (𝑥, 𝑦), we normalize it and project through 𝐿 log-spaced frequency bands to obtain Fourier features, then map them via a single sharedparameter MLP: 𝐿 pspat (r) = MLPshared sin(𝜔𝑙 r̄), cos(𝜔𝑙 r̄) 𝑙=1 ∈ R𝑑 , (16) where r̄ is the coordinate normalized by the maximum spatial ex𝐿 are 𝐿 logarithmically spaced frequency bands. tent and {𝜔𝑙 }𝑙=1 This MLPshared is applied to cell coordinates, 𝑆×𝑆 grid centers, and 𝐻 𝑓 ×𝑊 𝑓 pixel grids alike. The resulting encodings are injected into
Conference’17, July 2017, Washington, DC, USA
Xiaoqian Qi, Haoye Chai, Yue Wang, and Yong Li
intermediate feature maps of each context encoder before aggregation. spat spat h𝐹 𝑓 ← h𝐹 𝑓 + pfine , h𝐹𝑐 ← h𝐹𝑐 + pcoarse, (17) spat
and the POI/OD encoders similarly receive pcoarse . Since all coordinates pass through the same Fourier basis and MLP, tokens from different modalities at the same location receive identical positional signals, establishing cross-modal spatial alignment without explicit cross-attention. 4.3.3 Learnable Gating. The encoded context representations and action embeddings are injected into the state decoder through learnable gating modules. As shown in Figure 2, MobiWM employs four parallel gating modules G𝐹 , G𝑂 , G𝑃 , and G𝐴 for the facility map, OD flow, POI, and the action, respectively. Each condition h𝑖 ∈ {h𝑃 , h𝑂 , h𝐹 , h𝐴 } is first transformed by a conditioning projection 𝜙𝑖 (·) = Linear(LN(h𝑖 )) to align with the decoder’s hidden space. A learnable scalar gate 𝑔𝑖 ∈ R, initialized to a small negative value and passed through a sigmoid, controls the contribution of each condition: G𝑖 (h𝑖 ) = 𝜎 (𝑔𝑖 ) · 𝜙𝑖 (h𝑖 ). (18) The gated representations are then added to the decoder’s hidden states: ∑︁ Hdec ← Hdec + G𝑖 (h𝑖 ). (19) 𝑖 ∈ {𝐹,𝑂,𝑃,𝐴}
The scalar gate design provides a lightweight yet effective mechanism for the model to learn the relative importance of each modality during training. Finally, the hidden state is input into a linear prediction head Oout to transfer the latent future state in the traffic space: ŝ𝑡 +1:𝑡 +𝑃 = Oout Hdec ).
(20)
5 EXPERIMENTS 5.1 Experimental Setup 5.1.1 Dataset. We construct a simulation-augmented dataset from real network measurements of Nanchang City, China, covering 9 districts and 31,900 cells with 15-minute granularity over one week (𝑇 =672). The raw data provide per-cell traffic, coordinates, and static engineering parameters. We build two variable-parameter subsets, Para and Topo, via the ray-tracing pipeline, to generate action-aware datasets. For each map, we reconstruct a 3D urban scene from OSM building footprints, sample UE locations from WorldPop [30] population grids, and disaggregate cell traffic to UEs. Sionna ray-tracing [17] computes parameter-dependent RSRP maps. To enhance realism, UE-cell association follows a 3GPP A3 eventtriggered handover procedure [2] with access threshold 𝑄 rxlevmin [3] and hysteresis. Cell traffic at each step is the sum of associated UE traffic under the modified RSRP landscape. Para: Parameter Variation Dataset. The cell set V is fixed; 40% of cells undergo parameter changes. Figure 4 illustrates a sample from the Para dataset. It can be observed that, compared to static traffic patterns, our constructed action-aware traffic is capable of generating distinct fluctuation patterns in response to varying actions. Topo: Topology & Parameter Variation Dataset. The cell set V is additionally modified: existing cells are deactivated, or new cells
facility map
RSRP map
action-aware traffic
actions (power & azimuth)
static traffic pattern
Figure 4: View of the variable-parameter mobile traffic dataset. The green arrows indicate the significant correlation between action variations and traffic fluctuations. are inserted at random locations with sampled parameters. 40% of maps also include parameter-variation actions. 5.1.2 Baselines. We compare MobiWM against three categories of baselines. (1) Mobile traffic prediction models: FedGTP [33], HiSTM [5], and MobiFM [8]. For each, we evaluate both the original model (which predicts without action conditioning) and a worldmodel variant (suffixed with -WM) that augments the original architecture with our action-state formulation. (2) Spatio-temporal prediction models: iTransformer [20], Informer [28], TimeMoE [25], and CSDI [27], all adapted to the world-model formulation to accept action inputs. (3) Representative world models: TD-MPC2 [16], STORM [38], and DreamerV3 [14], which natively support actionconditioned state prediction. 5.1.3 Evaluation Metrics. We evaluate the rollout performance of MobiWM and baselines using three metrics: Jensen-Shannon Divergence (JSD) to measure distributional similarity between predicted and true traffic distributions, Mean Absolute Error (MAE) to quantify the average magnitude of prediction errors, and Normalized Root Mean Square Error (NRMSE) to assess the overall prediction accuracy normalized by the range of true values. These metrics together provide a comprehensive evaluation of both the fidelity and accuracy of the predicted traffic patterns under variable parameter scenarios.
5.2
Overall Rollout Performance
Table 1 summarizes the long-horizon rollout results across all four scenarios: Urban-Para, Urban-Topo, Suburb-Para, and SuburbTopo. 9 districts were categorized into two groups, Urban and Suburb, based on their geographical locations and building density. It can be noticed that MobiWM achieves the best JSD in all four scenarios and the best MAE in three out of four, demonstrating consistently superior distributional fidelity and prediction accuracy. Notably, it attains the lowest JSD of 0.311 on Suburb-Para and 0.324 on Urban-Topo, outperforming the strongest baseline by 9.4% and 23.1%, respectively. In the challenging Topo scenarios, MobiWM maintains stable rollout quality, whereas most baselines degrade substantially, confirming its robustness to action-induced structural changes. Besides, comparing original mobile traffic predictors (FedGTP, HiSTM, MobiFM) with their WM-augmented variants (FedGTP-WM, HiSTM-WM, MobiFM-WM), all three pairs show consistent improvements after introducing action conditioning, validating that the action-state dynamics paradigm is broadly effective and not architecture-specific. TD-MPC2, STORM, and DreamerV3,
Beyond Static Forecasting: Unleashing the Power of World Models for Mobile Traffic Extrapolation
Conference’17, July 2017, Washington, DC, USA
Table 1: Performance comparison across different scenarios. ( Bold indicates best, underlined indicates second best.) Urban-Para
Model
JSD
Urban-Topo
MAE(×1e5) NRMSE
JSD
Suburb-Para
MAE(×1e5) NRMSE
JSD
Suburb-Topo
MAE(×1e5) NRMSE
JSD
MAE(×1e5) NRMSE
FedGTP FedGTP-WM HiSTM HiSTM-WM MobiFM MobiFM-WM
0.7833 0.4741 0.4865 0.5058 0.4228 0.344
3.993 3.377 4.106 3.931 4.28 2.974
0.7824 0.7523 0.8495 0.7751 0.8528 0.6634
0.7644 0.5557 0.5365 0.4909 0.4846 0.4216
3.32 2.956 3.923 3.398 6.413 2.902
0.8758 0.8982 1.146 0.9415 2.119 0.8724
0.7625 0.4554 0.5154 0.4393 0.4228 0.3775
3.788 3.668 5.197 3.804 4.28 3.776
0.729 0.7233 1.127 0.7722 0.8528 0.8132
0.7898 0.5149 0.5558 0.5661 0.51 0.4401
2.976 2.758 5.17 2.916 2.935 4.48
0.8966 0.8425 1.888 0.8861 0.8629 1.534
iTransformer-WM Informer-WM TimeMoE-WM CSDI-WM
0.478 0.4674 0.4301 0.5195
3.33 3.736 3.346 6.035
0.7394 0.8139 0.6962 2.109
0.6224 0.6088 0.5233 0.4323
3.475 3.716 4.794 12.13
1.938 0.9728 1.311 4.66
0.5392 0.4709 0.4881 0.4979
4.717 3.808 3.491 9.349
0.8771 0.7143 0.7069 2.597
0.641 0.6506 0.4081 0.5151
4.678 5.102 2.958 15.37
1.512 1.635 1.114 6.962
TD-MPC2 STORM DreamerV3
0.6392 0.6337 0.6171
4.089 3.944 3.792
0.966 0.7581 0.765
0.6671 0.6291 0.5451
3.074 3.09 3.084
0.7963 0.7994 0.8247
0.7111 0.7032 0.5407
3.822 4.107 3.623
0.935 0.7839 0.722
0.6132 0.6656 0.5378
5.304 2.776 2.301
16.12 0.8058 0.7273
MobiWM (Ours)
0.3341
3.063
0.6731
0.3242
2.893
0.793
0.311
3.14
0.6882
0.3693
2.152
0.7987
w/o FA/OD w/o TRN/SPE/LG
w/o FA w/o TRN
w/o OD w/o SPE
3.6
w/o LG
3.2
3.0 2.8
Urban-Topo
(a) Performance on Urban-Para and Urban-Topo MobiWM 2.6
MAE (×1e5)
4.0
Urban-Para
MAE (×1e5)
3.0
MAE (×1e5)
MobiWM 3.4 3.2
MAE (×1e5)
3.4
2.4
3.5 3.0
2.2
Suburb-Para
Suburb-Topo
(b) Performance on Suburb-Para and Suburb-Topo Figure 5: Ablation study on environment context modalities (w/o FA, w/o OD) and fusion mechanism components (w/o TRN, w/o SPE, w/o LG). The dashed line marks the full MobiWM. designed for low-dimensional continuous control, rely on compact latent representations and lack explicit spatial modeling, making them unable to encode the high-dimensional, graph-structured state space of mobile networks. This gap underscores the necessity of domain-specific designs that MobiWM introduces.
5.3
Ablation Studies
We ablate the environmental context modalities and the multimodal fusion mechanism. Results are reported in Figure 5. 5.3.1 Environment context ablation. Removing the facility map (w/o FA) causes consistent MAE increases across all four scenarios, indicating that static infrastructure layout provides essential spatial priors for traffic dynamics modeling. Excluding OD flow (w/o OD) leads to comparable degradation, as it captures time-varying mobility demand that directly drives traffic redistribution across
cells. When both are removed simultaneously (w/o FA/OD), the degradation is most pronounced, particularly on Urban-Para and Suburb-Topo, confirming that the two modalities provide complementary information and their joint presence is critical for accurate rollout. 5.3.2 Multimodal fusion ablation. Removing all fusion components (w/o TRN/SPE/LG) yields the largest MAE increase among all ablated variants, even exceeding the context-removal settings on several scenarios, demonstrating that how modalities are fused matters as much as which modalities are included. Among individual components, removing shared positional encoding (w/o SPE) or learnable gating (w/o LG) each causes notable performance drops, as the former disrupts cross-modal spatial alignment and the latter disables adaptive modality weighting based on local context. Removing the Topology Representation Network (w/o TRN) leads to moderate degradation, indicating that the learned spatial bias enriches the model’s understanding of network topology but is partially compensated by the remaining spatial encoding. The full fusion design consistently achieves the lowest MAE (dashed line), validating that all three components synergistically contribute to effective multimodal integration.
5.4
Action sensitivity
To evaluate the generalization of MobiWM to out-of-distribution actions and emergency events, we test two extreme scenarios (Figure 6): (a) Action exceeds threshold: the transmit power surpasses the training-set maximum (red dashed line) during several intervals, accompanied by large azimuth swings; (b) Sudden power outage: the base station is completely powered off mid-week, dropping all parameters to zero. In both cases, MobiWM (top-right panels) closely tracks the real traffic variations, promptly responding to the abrupt parameter changes with accurate rollout predictions. In contrast, MobiFM (bottom-right panels), which lacks action conditioning, continues to predict traffic based solely on historical temporal patterns and fails to reflect the drastic state changes caused by out-of-distribution actions or power-off events. This comparison highlights a fundamental advantage of the world-model paradigm:
Conference’17, July 2017, Washington, DC, USA
Xiaoqian Qi, Haoye Chai, Yue Wang, and Yong Li
Table 2: Steady-state RMSE (↓) of RL-based network optimization using different world models as the simulation environment. Results over 18 runs (95% CI).
threshold
Model
RMSE (Std)
Model
RMSE (Std)
MobiWM
22369 (4848)
Static-MobiWM
23838 (4952)
TimeMoE-WM
27703 (4586)
MobiFM-WM
27892 (8102)
(a) Emergency situation: Action exceeds threshold.
power outage
(b) Emergency: Sudden power outage.
Figure 6: Performance comparison during emergency events.
STORM-S MobiWM-S
MobiWM-XL
DreamerV3-L TD-MPC2-S TD-MPC2-M
DreamerV3-S DreamerV3-M
TD-MPC2-L STORM-L
MobiWM-L MobiWM-M
(a) Model efficiency analysis on Urban-Para dataset. STORM-L DreamerV3-L DreamerV3-S MobiWM-S MobiWM-L DreamerV3-M MobiWM-XL MobiWM-M
TD-MPC2-S
TD-MPC2-L
STORM-S
TD-MPC2-M
than generic world model designs. Notably, a clear scaling law does not emerge for any model family: increasing parameters does not monotonically improve performance. For instance, DreamerV3-L is the largest model, yet it performs worse than DreamerV3-M on both datasets, and TD-MPC2-L shows no gain over TD-MPC2-M. Within MobiWM, the MobiWM-M achieves the best overall accuracy, while MobiWM-XL offers only marginal or no improvement. This suggests that, for the mobile traffic dynamics task, the bottleneck lies in architectural inductive biases rather than raw model capacity, further justifying MobiWM’s design choices of factorized spatio-temporal blocks and shared spatial semantics over simply scaling up parameters.
5.6
Case Study
A core motivation of MobiWM is to serve as a counterfactual simulation environment for network optimization. To validate this, we couple the frozen MobiWM with a PPO [23] agent to solve a practical task: traffic load balancing via parameter adjustment. In real-world operations, uneven traffic distribution across cells leads to localized congestion and resource waste. Operators typically tune antenna parameters (e.g., tilt, power, azimuth) to redistribute traffic toward a balanced target profile. However, evaluating the effect of each adjustment on the live network is costly and risky. We formulate this as a sequential optimization problem. Given a desired traffic distribution s★1:𝑇 derived from operational planning targets, the RL agent learns to find parameter adjustments that steer the network toward this target: 𝑇
(b) Model efficiency analysis on Suburb-Para dataset.
Figure 7: Model efficiency comparison. Bubble position encodes rollout accuracy (JSD, MAE); bubble size encodes parameter count. by explicitly modeling the action-state dynamics, MobiWM generalizes to unseen parameter regimes and emergency scenarios that static predictors cannot handle.
5.5
Model Efficiency
Figure 7 compares the rollout accuracy (JSD, MAE) against parameter count for MobiWM (S/M/L/XL) and three general-purpose world models (TD-MPC2, STORM, DreamerV3), each also scaled to S/M/L variants. All MobiWM variants cluster achieve the lowest JSD and MAE, while using substantially fewer parameters than the baselines. This confirms that MobiWM’s domain-specific architecture is inherently more parameter-efficient for mobile network dynamics
1 ∑︁ ŝ𝑡 − s★ 𝑡 1 + 𝜆 Δ |Δa𝑡 |, a1:𝑇 𝑠¯★ 𝑡 =1
min
(21)
s.t. ŝ𝑡 +1 = 𝑓Ω (ŝ𝑡 , a𝑡 , c), where the first term measures the deviation from the target distribution, 𝜆Δ penalizes excessive parameter oscillation to ensure operational stability, and the dynamics constraint is enforced by the frozen world model 𝑓Ω . The agent interacts exclusively with MobiWM with no real network access is needed during training. As shown in Table 2, MobiWM achieves the lowest steady-state RMSE after policy convergence, outperforming TimeMoE-WM, MobiFM-WM, and Static-MobiWM (which removes action conditioning). The gap between MobiWM and Static-MobiWM confirms that faithful action-conditioned dynamics are essential for the RL agent to discover effective optimization strategies. The larger gaps against TimeMoE-WM and MobiFM-WM further indicate that inaccurate environment dynamics mislead the RL policy, resulting in suboptimal parameter adjustments.
Beyond Static Forecasting: Unleashing the Power of World Models for Mobile Traffic Extrapolation
6
CONCLUSION
In this paper, we present MobiWM, the first world model for mobile networks that learns the dynamics between network parameter adjustments and traffic variations. By formulating cell-level traffic as the system state and engineering parameters as actions, MobiWM transforms mobile traffic prediction from a static forecasting task into an action-conditioned dynamics modeling problem. The proposed Transformer-based architecture, equipped with Factorized Spatio-Temporal Blocks and a multimodal fusion mechanism with shared spatial semantics, effectively captures the complex spatio-temporal dependencies of mobile networks while integrating heterogeneous urban context. Extensive experiments on variableparameter traffic data spanning 31,900 cells across 9 districts demonstrate that MobiWM consistently achieves the best distributional fidelity across all evaluation scenarios, outperforming both domainspecific traffic predictors and general-purpose world models. A downstream RL-based case study further validates MobiWM as a reliable counterfactual simulation environment for network optimization. We believe MobiWM opens a promising direction toward digital twin-driven wireless network management, where operators can explore and evaluate optimization strategies in an imagined environment before deploying them to the live network.
REFERENCES [1] 3GPP. 2022. Study on Channel Model for Frequencies from 0.5 to 100 GHz. Technical Report TR 38.901 V17.0.0. 3rd Generation Partnership Project. [2] 3GPP. 2023. NR; Radio Resource Control (RRC); Protocol specification. Technical Specification TS 38.331. 3rd Generation Partnership Project. [3] 3GPP. 2023. NR; User Equipment (UE) procedures in idle mode and in RRC inactive state. Technical Specification TS 38.304. 3rd Generation Partnership Project. [4] Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. 2020. Adaptive graph convolutional recurrent network for traffic forecasting. In Proceedings of the 34th International Conference on Neural Information Processing Systems (Vancouver, BC, Canada) (NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, Article 1494, 12 pages. [5] Zineddine Bettouche, Khalid Ali, Andreas Fischer, and Andreas Kassler. 2025. HiSTM: Hierarchical Spatiotemporal Mamba for Cellular Traffic Forecasting. arXiv e-prints, Article arXiv:2508.09184 (Aug. 2025), arXiv:2508.09184 pages. arXiv:2508.09184 [cs.NI] doi:10.48550/arXiv.2508.09184 [6] Haoye Chai, Tao Jiang, and Li Yu. 2024. Diffusion Model-based Mobile Traffic Generation with Open Data for Network Planning and Optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, New York, NY, USA, 4828–4838. doi:10.1145/3637528.3671544 [7] Haoye Chai, Xiaoqian Qi, and Yong Li. 2025. Spatio-Temporal Knowledge Driven Diffusion Model for Mobile Traffic Generation. IEEE Transactions on Mobile Computing 24, 6 (June 2025), 4939–4956. doi:10.1109/TMC.2025.3527966 [8] Haoye Chai, Xiaoqian Qi, Yibo Ma, Zhaocheng Wang, Lei Yue, and Yong Li. 2026. MobiFM: A Foundation Model for Mobile Data Forecasting. IEEE Journal on Selected Areas in Communications 44 (2026), 2494–2509. doi:10.1109/JSAC.2025. 3642851 [9] Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. 2022. TransDreamer: Reinforcement Learning with Transformer World Models. arXiv preprint arXiv:2202.09481 (2022). [10] Qingbo Du, Faming Yin, and Zongchen Li. 2020. Base station traffic prediction using XGBoost-LSTM with feature enhancement. IET Networks 9, 1 (2020), 29–37. arXiv:https://ietresearch.onlinelibrary.wiley.com/doi/pdf/10.1049/ietnet.2019.0103 doi:10.1049/iet-net.2019.0103 [11] David Ha and Jürgen Schmidhuber. 2018. Recurrent world models facilitate policy evolution. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (Montréal, Canada) (NIPS’18). Curran Associates Inc., Red Hook, NY, USA, 2455–2467. [12] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. 2019. Dream to Control: Learning Behaviors by Latent Imagination. arXiv e-prints, Article arXiv:1912.01603 (Dec. 2019), arXiv:1912.01603 pages. arXiv:1912.01603 [cs.LG] doi:10.48550/arXiv.1912.01603 [13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. 2020. Mastering Atari with Discrete World Models. arXiv e-prints, Article arXiv:2010.02193
Conference’17, July 2017, Washington, DC, USA
(Oct. 2020), arXiv:2010.02193 pages. arXiv:2010.02193 [cs.LG] doi:10.48550/arXiv. 2010.02193 [14] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. 2023. Mastering Diverse Domains through World Models. arXiv e-prints, Article arXiv:2301.04104 (Jan. 2023), arXiv:2301.04104 pages. arXiv:2301.04104 [cs.AI] doi:10.48550/arXiv.2301.04104 [15] William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, USA) (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 1025–1035. [16] Nicklas Hansen, Hao Su, and Xiaolong Wang. 2023. TD-MPC2: Scalable, Robust World Models for Continuous Control. arXiv e-prints, Article arXiv:2310.16828 (Oct. 2023), arXiv:2310.16828 pages. arXiv:2310.16828 [cs.LG] doi:10.48550/arXiv. 2310.16828 [17] Jakob Hoydis, Faycal Ait Aoudia, Sebastian Cammerer, Merlin Nimier-David, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. 2023. Sionna RT: Differentiable Ray Tracing for Radio Propagation Modeling. In 2023 IEEE Globecom Workshops (GC Wkshps). 317–321. doi:10.1109/GCWkshps58843.2023.10465179 [18] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open Graph Benchmark: Datasets for Machine Learning on Graphs. In Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 22118–22133. https://proceedings.neurips.cc/ paper_files/paper/2020/file/fb60d411a5c5b72b2e7d3527cfc84fd0-Paper.pdf [19] Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2017. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. arXiv e-prints, Article arXiv:1707.01926 (July 2017), arXiv:1707.01926 pages. arXiv:1707.01926 [cs.LG] doi:10.48550/arXiv.1707.01926 [20] Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2023. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. arXiv e-prints, Article arXiv:2310.06625 (Oct. 2023), arXiv:2310.06625 pages. arXiv:2310.06625 [cs.LG] doi:10.48550/arXiv.2310.06625 [21] Jiaming Ma, Binwu Wang, Pengkun Wang, Zhengyang Zhou, Yudong Zhang, Xu Wang, and Yang Wang. 2025. MobiMixer: A Multi-Scale Spatiotemporal Mixing Model for Mobile Traffic Prediction. IEEE Transactions on Mobile Computing 24, 11 (2025), 11972–11986. doi:10.1109/TMC.2025.3585007 [22] Ali Yadavar Nikravesh, Samuel A. Ajila, Chung-Horng Lung, and Wayne Ding. 2016. Mobile Network Traffic Prediction Using MLP, MLPWD, and SVM. In 2016 IEEE International Congress on Big Data (BigData Congress). 402–409. doi:10.1109/ BigDataCongress.2016.63 [23] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal Policy Optimization Algorithms. arXiv e-prints, Article arXiv:1707.06347 (July 2017), arXiv:1707.06347 pages. arXiv:1707.06347 [cs.LG] doi:10.48550/arXiv.1707.06347 [24] Xingjian SHI, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun WOO. 2015. Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. In Advances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett (Eds.), Vol. 28. Curran Associates, Inc. https://proceedings.neurips.cc/paper_ files/paper/2015/file/07563a3fe3bbe7e3ba84431ad9d055af-Paper.pdf [25] Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2024. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. arXiv e-prints, Article arXiv:2409.16040 (Sept. 2024), arXiv:2409.16040 pages. arXiv:2409.16040 [cs.LG] doi:10.48550/arXiv.2409.16040 [26] Yantai Shu, Minfang Yu, Jiakun Liu, and O.W.W. Yang. 2003. Wireless traffic modeling and prediction using seasonal ARIMA models. In IEEE International Conference on Communications, 2003. ICC ’03., Vol. 3. 1675–1679 vol.3. doi:10. 1109/ICC.2003.1203886 [27] Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. 2021. CSDI: conditional score-based diffusion models for probabilistic time series imputation. In Proceedings of the 35th International Conference on Neural Information Processing Systems (NIPS ’21). Curran Associates Inc., Red Hook, NY, USA, Article 1900, 13 pages. [28] Hasita Veluri and Dilip Vasudevan. 2025. InFormer: A High-throughput, Ultraefficient In-memory Compute-based Floating-point Arithmetic Accelerator for Transformers. In Proceedings of the Great Lakes Symposium on VLSI 2025 (GLSVLSI ’25). Association for Computing Machinery, New York, NY, USA, 718–725. doi:10. 1145/3716368.3735246 [29] Xing Wang, Zhendong Wang, Kexin Yang, Zhiyan Song, Chong Bian, Junlan Feng, and Chao Deng. 2024. A Survey on Deep Learning for Cellular Traffic Prediction. Intelligent Computing 3 (01 2024). doi:10.34133/icomputing.0054 [30] WorldPop. 2018. WorldPop Open Population Data. https://www.worldpop.org/. School of Geography and Environmental Science, University of Southampton. [31] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. 2019. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19. International Joint Conferences on Artificial Intelligence Organization, 1907– 1913. doi:10.24963/ijcai.2019/264
Conference’17, July 2017, Washington, DC, USA
[32] Kai Xu, Rajkarn Singh, Hakan Bilen, Marco Fiore, Mahesh K. Marina, and Yue Wang. 2022. CartaGenie: Context-Driven Synthesis of City-Scale Mobile Network Traffic Snapshots. In 2022 IEEE International Conference on Pervasive Computing and Communications (PerCom). 119–129. doi:10.1109/PerCom53586.2022.9762395 [33] Linghua Yang, Wantong Chen, Xiaoxi He, Shuyue Wei, Yi Xu, Zimu Zhou, and Yongxin Tong. 2024. FedGTP: Exploiting Inter-Client Spatial Dependency in Federated Graph-based Traffic Prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, New York, NY, USA, 6105–6116. doi:10.1145/3637528.3671613 [34] Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2018. Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18. International Joint Conferences on Artificial Intelligence Organization, 3634–3640. doi:10.24963/ijcai.2018/505 [35] Chaoyun Zhang, Paul Patras, and Hamed Haddadi. 2019. Deep Learning in Mobile and Wireless Networking: A Survey. IEEE Communications Surveys & Tutorials 21, 3 (2019), 2224–2287. doi:10.1109/COMST.2019.2904897 [36] Junbo Zhang, Yu Zheng, and Dekang Qi. 2017. Deep spatio-temporal residual networks for citywide crowd flows prediction. In Proceedings of the Thirty-First AAAI
Xiaoqian Qi, Haoye Chai, Yue Wang, and Yong Li
Conference on Artificial Intelligence (San Francisco, California, USA) (AAAI’17). AAAI Press, 1655–1661. [37] Shiyuan Zhang, Yilai Liu, Yuwei Du, Ruoxuan Yang, Dong In Kim, and Hongyang Du. 2026. U-MASK: User-adaptive Spatio-Temporal Masking for Personalized Mobile AI Applications. arXiv e-prints, Article arXiv:2601.06867 (Jan. 2026), arXiv:2601.06867 pages. arXiv:2601.06867 [stat.ML] doi:10.48550/arXiv.2601. 06867 [38] Weipu Zhang, Gang Wang, Jian Sun, Yetian Yuan, and Gao Huang. 2023. STORM: efficient stochastic transformer based world models for reinforcement learning. In Proceedings of the 37th International Conference on Neural Information Processing Systems (New Orleans, LA, USA) (NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, Article 1182, 20 pages. [39] Changyuan Zhao, Guangyuan Liu, Ruichen Zhang, Yinqiu Liu, Jiacheng Wang, Jiawen Kang, Dusit Niyato, Zan Li, Xuemin Shen, Zhu Han, Sumei Sun, Chau Yuen, and Dong In Kim. 2026. Edge General Intelligence Through World Models, Large Language Models, and Agentic AI: Fundamentals, Solutions, and Challenges. IEEE Transactions on Cognitive Communications and Networking 12 (2026), 5649–5675. doi:10.1109/TCCN.2026.3658762