Mobile Edge General Intelligence

• Mobile Edge General Intelligence (MEGI) is an edge-centric paradigm enabling autonomous multi-task decision-making across mobile, UAV, vehicular, and IoT devices.
• MEGI architectures integrate device, edge, and cloud resources using hybrid LLM/SLM orchestration, world models, and digital twin integration to overcome wireless constraints.
• MEGI addresses challenges in scalability, reliability, and standardization by optimizing latency, energy, and privacy, ensuring adaptive and decentralized intelligence at the network edge.

Mobile Edge General Intelligence (MEGI) denotes the mobile and wireless instantiation of Edge General Intelligence (EGI): an edge-centric intelligence paradigm in which autonomous agents operate across mobile devices, UAVs, vehicles, roadside units, base stations, and edge servers under stringent constraints of latency, energy, bandwidth, reliability, and privacy. Across the literature, MEGI is described as bringing long-term, multi-task autonomy close to the physical world in 6G-and-beyond environments, while integrating mobile endpoints, device–edge–cloud orchestration, and wireless-specific constraints such as URLLC, mobility, intermittent connectivity, and spectrum dynamics (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025, Chen et al., 2024). The concept is broader than narrow edge inference: it encompasses perception, reasoning, planning, memory, tool use, communication-aware execution, and continual adaptation across heterogeneous edge nodes (Luo et al., 27 Sep 2025, Wu et al., 25 Nov 2025).

1. Definition, scope, and relation to edge intelligence

MEGI is consistently positioned as the realization of EGI in mobile contexts. In this usage, EGI extends traditional Edge Intelligence from task-specific inference toward general-purpose cognitive capabilities—comprehension, reasoning, generation, adaptability, multimodality, and contextualization—executed at the network edge rather than exclusively in centralized cloud infrastructure (Chen et al., 2024). MEGI narrows that broader paradigm to mobile, wireless, and heterogeneous edge environments, including smartphones, wearables, vehicular units, IoT gateways, UAVs, low-altitude platforms, RAN nodes, and space–air–ground systems (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025).

The goals attributed to MEGI recur across the surveyed work. First, autonomy is defined as closed-loop sensing–modeling–planning–actuation without continuous cloud dependence. Second, adaptability refers to continual operation in dynamic, partially observable, and non-stationary edge environments. Third, scalability refers to decentralized intelligence across many devices, vehicles, UAVs, and base stations (Zheng et al., 18 Mar 2026). Complementary surveys emphasize that MEGI must also be communication-aware, exchanging compact and semantically meaningful representations rather than raw data when possible, and adapting execution to current channel states, fading, interference, and mobility (Wu et al., 25 Nov 2025).

A recurring distinction is that MEGI is not identical to conventional MEC nor to narrow AI offloading. MEC provides compute, storage, and caching near users; MEGI adds general-purpose reasoning, planning, memory, and adaptive orchestration on top of that infrastructure (Wang et al., 2022, Lin et al., 2020). Similarly, traditional edge AI usually deploys specialized models for specific tasks, whereas MEGI seeks cross-task generalization and complex instruction following through LLMs, SLMs, world models, or distilled student models, depending on deployment constraints (Chen et al., 2024, Wu et al., 25 Nov 2025).

A common misconception is that MEGI is simply “running an LLM on a phone.” The surveyed literature does not support that reduction. Some strands frame MEGI through LLM/SLM orchestration across device–edge–cloud (Chen et al., 2024, Luo et al., 27 Sep 2025), others through world-model-driven agentic AI (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025), others through knowledge distillation for edge-native student models (Wu et al., 25 Nov 2025), and others through distributed generative pipelines, caching, or RL-enabled orchestration (Zhong et al., 2023, Wei et al., 2022). This suggests that MEGI is best understood as a systems-level convergence of mobile edge computing, AI-native orchestration, and general-purpose decision-making rather than a single model family.

2. From digital twins and narrow inference to agent-centric world models

One major line of work treats MEGI as the outcome of a transition from Digital Twins (DTs) to World Models (WMs) at the edge. In this formulation, DTs are physics-based, high-fidelity, often centralized virtual replicas of physical systems, useful for monitoring, analysis, and offline optimization; WMs are learned internal models, built from interaction data, that encode compact latent dynamics relevant to decision-making (Zheng et al., 18 Mar 2026). The distinction is also framed as centralized versus decentralized and system-centric versus agent-centric: DTs manage global system state and lifecycle, whereas WMs represent an agent’s local observations, actions, and reward-relevant dynamics (Zheng et al., 18 Mar 2026).

This distinction matters because highly dynamic mobile edge environments expose the limitations of DT-only approaches. The literature attributes greater autonomy, adaptability, and scalability to WMs for four stated reasons: latent abstraction reduces compute, storage, and communication costs; action-conditioned imagination enables foresight without risky real-world trial-and-error; continual learning adapts to unmodeled dynamics and distribution shifts; and agent-centric local modeling decouples complexity from the full environment, supporting multi-agent scaling (Zheng et al., 18 Mar 2026). A related survey describes world models as the “cognitive backbone” of edge agentic AI because they can learn compact latent representations, predict and imagine future trajectories, and support multi-step planning under uncertainty and partial observability (Zhao et al., 13 Aug 2025).

The formal control perspective typically uses MDPs or POMDPs. The canonical MDP is given as

$$\text{MDP} = (\mathcal{S}, \mathcal{A}, T, R, \gamma),$$

with transitions $T(s_{t+1}\mid s_t,a_t)$ and reward $R(s_t,a_t)$, while mobile edge settings are more naturally modeled as POMDPs with observations $o_t$ and latent beliefs over hidden state (Zheng et al., 18 Mar 2026). In WM-based formulations, the model supplies

$$p(z_{t+1}\mid z_t,a_t),\quad p(o_t\mid z_t),\quad r(z_t,a_t),$$

and planning seeks to maximize

$$\max_{\pi}\ \mathbb{E}\left[\sum_{t=0}^{T}\gamma^t r_t\right].$$

Representation learning is commonly expressed via a variational objective such as

$$\mathcal{L}=\mathbb{E}_{q(z\mid o)}[\log p(o\mid z)]-\mathrm{KL}\big(q(z\mid o)\,\|\,p(z)\big),$$

with predictive or reward-consistency losses layered on top (Zheng et al., 18 Mar 2026).

The literature also argues for hybrid DT–WM integration rather than replacement. In this hybrid view, DTs provide physics priors, radio propagation constraints, mobility envelopes, safety rules, global synchronization, and cross-agent consistency, while WMs learn residuals, latent dynamics, and task-relevant abstractions for local decision-making (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025). A plausible implication is that MEGI architectures may increasingly separate global governance and validation from local cognition: DTs handle calibration and system boundaries, while WMs handle edge-native anticipation and control.

3. Architectural patterns and algorithmic foundations

MEGI architectures are repeatedly described as spanning device, edge, RAN, and cloud. In WM-centric formulations, on-device or RAN sensing from cameras, radar, CSI, or IMUs feeds an encoder, such as a CNN, Transformer, or VAE, that compresses observations $o_t$ into latent states $z_t$ (Zheng et al., 18 Mar 2026). Dynamics learning then models action-conditioned temporal evolution in latent space, for example with RSSM, Mamba-based variants, or TD-MPC2-style transitions (Zheng et al., 18 Mar 2026). Planning is performed through latent rollouts via MPC, CEM, or MCTS, and policy/value learning may proceed from imagined trajectories in Dreamer-style actor–critic form (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025).

Representative algorithmic choices named across the surveys include Dreamer-style RSSM, MuZero-style MCTS over learned dynamics, TD-MPC2 with MPC/CEM for continuous control, Bayesian filters for latent-state belief updates, and episodic plus semantic memory modules for retrieval-augmented planning (Zheng et al., 18 Mar 2026). Edge-oriented WM adaptations also include Sparse Imagination, which reduced PushT episode time from 173 s to 82 s at comparable success, and SGF, which reduced Atari-100k total training time from 12 h to 3 h versus DreamerV3 (Zhao et al., 13 Aug 2025). These are presented as examples of reducing planning or training cost under edge constraints rather than as universal defaults.

A parallel architectural lineage frames MEGI through LLM/SLM deployment patterns. The literature groups these into centralized, hybrid, and decentralized EGI systems, all applicable to mobile edge settings (Chen et al., 2024).

Architecture	Core arrangement	MEGI characteristics
Centralized	GI capabilities reside in cloud or powerful edge servers	Simple thin-client model; suited to bandwidth-sufficient, latency-tolerant tasks (Chen et al., 2024)
Hybrid	Cloud hosts an LLM, devices run SLMs, routing decides local vs offloaded execution	Identified as the dominant pattern for mobile devices (Chen et al., 2024)
Decentralized	Each device hosts an SLM and collaborates peer-to-peer	Useful under intermittent or constrained networks (Chen et al., 2024)

In hybrid architectures, the device typically runs an SLM for low-latency or privacy-sensitive steps, while the edge or cloud augments complex reasoning, long-context processing, or multimodal fusion (Chen et al., 2024). The control plane includes a quality/cost-aware router, offloading policy, trust and identity management, update distribution, and fallback rules; the data plane includes token streams, compressed multimodal payloads, semantic messages, KV-cache segments, and function-calling interfaces (Chen et al., 2024). Additional work extends this with adaptive CoT depth selection and distributed MoE routing, jointly optimizing reasoning depth, expert activation, and transmission power under latency, memory, and energy constraints (Luo et al., 27 Sep 2025).

Knowledge distillation is presented as a complementary enabler for making MEGI practical on mobile and IoT endpoints. The canonical soft-target objective is given as

$$L_{\mathrm{KD}}=\alpha\,\mathrm{CE}(y,p_s)+\beta\,T^2\,\mathrm{KL}\!\left(\mathrm{softmax}\!\left(\tfrac{z_t}{T}\right)\,\big\Vert\,\mathrm{softmax}\!\left(\tfrac{z_s}{T}\right)\right),$$

with cross-architecture distillation, feature-based distillation, and communication-aware weighting $T(s_{t+1}\mid s_t,a_t)$0 proposed for wireless conditions (Wu et al., 25 Nov 2025). Architectures singled out as edge-suited include Mamba and RWKV because of linear-time or constant-state characteristics, and the survey reports examples such as SenseMamba at 0.021M parameters and RWKV-Lite with 3.4–5× footprint reduction (Wu et al., 25 Nov 2025).

4. Communication, computation, and deployment at the mobile edge

MEGI is usually modeled over a device–edge–cloud continuum with explicit latency, energy, bandwidth, and memory budgets. A standard end-to-end latency decomposition used in LLM-oriented EGI work is

$T(s_{t+1}\mid s_t,a_t)$1

with

$T(s_{t+1}\mid s_t,a_t)$2

and total energy

$T(s_{t+1}\mid s_t,a_t)$3

Offloading is then chosen to minimize a weighted cost such as

$T(s_{t+1}\mid s_t,a_t)$4

subject to memory, battery, thermal, and bandwidth constraints (Chen et al., 2024).

World-model and wireless-edge work expresses the same constraints in radio-compute form. Representative formulas include

$T(s_{t+1}\mid s_t,a_t)$5

for communication, and

$T(s_{t+1}\mid s_t,a_t)$6

for on-device compute, leading to

$T(s_{t+1}\mid s_t,a_t)$7

as an end-to-end latency model (Zhao et al., 13 Aug 2025). RL-enabled MEC surveys use similar decompositions for local compute, offloading, queueing, and drift-plus-penalty scheduling in dynamic environments (Wei et al., 2022).

Deployment guidance in the literature is consistently modular. Devices host lightweight encoders or SLMs, policy inference, safety monitors, tokenizers, runtimes, local memory stores, and tool APIs; MEC servers host dynamics learning, imagination or planning services, vector databases or RAG stores, model caches, adapter registries, privacy gateways, and multi-agent coordination; RAN nodes contribute telemetry, slicing-aware scheduling, and model-serving endpoints; cloud backends handle pretraining, fine-tuning, distillation, model registries, and cross-site federation (Zheng et al., 18 Mar 2026, Chen et al., 2024). Interoperability appears as a recurrent requirement, usually described as standardized APIs, telemetry schemas, and model-serving interfaces across heterogeneous vendors and nodes (Zheng et al., 18 Mar 2026).

Distributed service placement and collaborative execution are also treated as core MEGI mechanisms. AMCoEdge, for example, formulates multi-edge workload allocation via fractions $T(s_{t+1}\mid s_t,a_t)$8 summing to one across candidate edge servers and minimizes average makespan with an online DRL scheduler plus closed-form workload allocation (Xu et al., 19 Dec 2025). The reported gains—at least 11.04% reduction in task offloading make-span, 44.86% decrease in failure rate, and 9.23%–31.98% lower service delays in a distributed prototype—are presented specifically for AIGC workloads, but they indicate that adaptive multi-edge collaboration is technically viable for broader MEGI service fabrics (Xu et al., 19 Dec 2025).

5. Application domains and empirical patterns

The application space attributed to MEGI is broad but internally consistent: it centers on settings where communication, sensing, and control are coupled, and where centralized cloud intelligence is either too slow, too bandwidth-intensive, or too privacy-sensitive. Integrated Sensing and Communications (ISAC/ISCC) is a recurrent example. There, MEGI supports joint waveform or resource design under power, latency, and URLLC constraints, with objectives such as

$T(s_{t+1}\mid s_t,a_t)$9

The survey reports approximately 33% end-to-end latency reduction and approximately 30% tail-latency reduction at approximately 2–7% energy cost for DT-integrated frameworks, and approximately 15–20% lower execution latency for WM-driven online adaptation via dynamic parameter tuning (Zheng et al., 18 Mar 2026).

Vehicular, UAV, and low-altitude wireless networks form another major cluster. In V2X, WMs are described as anticipating mobility-driven channel changes, interference, and blockages for proactive spectrum allocation, beam alignment, or power control (Zhao et al., 13 Aug 2025). Specific figures reported in the survey include predictive power allocation with 85.71% accuracy, packet-completeness-aware AoI reduced by 26% versus model-based RL and 16% versus model-free RL in vehicular mmWave scheduling, and radio success ratio improved by more than 50% using a neural POMDP world model with only 1/20 of the training data (Zhao et al., 13 Aug 2025). In low-altitude wireless networks, Wireless Dreamer is reported to accelerate convergence by 46.15% compared with DQN under weather uncertainty (Zhao et al., 13 Aug 2025).

Semantic communication is another recurrent target. Here, task-oriented criteria are given in terms of mutual information $R(s_t,a_t)$0 and rate–distortion $R(s_t,a_t)$1, and DT–WM combinations are described as calibrating offline policies while adapting encoding, transmission, and decoding to time-varying channels and semantic demands (Zheng et al., 18 Mar 2026). Wireless context engineering extends this logic by treating the air interface as a context pipeline that selects, compresses, and injects task-relevant context under budget constraints. In an ISAC-enabled beam prediction case study on DeepSense 6G Scenario 9, GPS-only top-3 accuracy is reported at approximately 60–63%, single missing-modality cases at approximately 74–76%, full observation at approximately 83–85%, and RL-driven context selection within approximately 2–3% of full-observation accuracy while improving reward by approximately 10–25% over fixed single-modality baselines (Zhao et al., 7 Feb 2026).

Generative services and Metaverse-oriented workloads provide a different empirical window on MEGI. Mobile Edge Generation (MEG) distributes generative models across edge servers and UEs, transmitting compact seeds or sketches rather than raw media (Zhong et al., 2023, Zhong et al., 2024). In the 2023 MEG formulation, the CIAG protocol exchanges approximately 57 kb total in latent seeds rather than approximately 2.6 Mb in full image round-trip transfers, corresponding to an approximately 47.2× communication reduction under the case study assumptions (Zhong et al., 2023). The 2024 MEG paper further reports transmitted data reductions from 1,048,576 floats in centralized generation to 8,192 floats for MEG with $R(s_t,a_t)$2, while showing that compression coding improves image quality under low-SNR conditions and that DRL-based power allocation improves FID by approximately 20 points over tie allocation at 1 mW (Zhong et al., 2024).

MEGI is also applied to task markets, trustless coordination, and AI service placement. BeTrustMEC relocates incentive settlement and verification to MEC edge servers with blockchain smart contracts, replacing remote platform relays and defining a reward–penalty mechanism for crowd-intelligence workflows (Xu et al., 2019). AI service placement work formulates a MINLP for deciding which users should receive updated AI programs for local execution versus offloading to the edge, and reports greedy and ADMM-based algorithms that perform extremely close to optimum while significantly outperforming representative benchmarks (Lin et al., 2020). These strands suggest that MEGI is not limited to inference or control loops; it also includes economic coordination and lifecycle management at the edge.

6. Challenges, misconceptions, and research directions

The dominant challenges are highly consistent across the surveys. Scalability remains difficult because MEGI must operate across multi-agent, multi-domain, multi-scale environments with heterogeneous devices, dynamic channels, and distributed services (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025). Reliability and safety remain open because imagination errors, distribution shifts, and policy failures can be consequential in beam alignment, driving, or industrial control; the literature therefore emphasizes uncertainty quantification, constrained planning, CMDP formulations, and safe fallback strategies (Zhao et al., 13 Aug 2025). Interoperability and standardization recur as unresolved issues, especially around DT–WM schemas, model exchange interfaces, O-RAN integration, edge-native APIs, and context-token metadata (Zheng et al., 18 Mar 2026, Zhao et al., 7 Feb 2026).

A second cluster of open problems concerns constrained deployment. LLM reasoning surveys emphasize that context windows, attention, and KV-cache memory scale poorly at the edge, and that long-context inference inflates latency and energy (Zhao et al., 7 Feb 2026). Quantization helps, but aggressive compression may reduce reasoning quality; one survey explicitly notes that 3-bit quantization can degrade accuracy by up to 40% compared to higher precision (Luo et al., 27 Sep 2025). World-model surveys likewise identify sub-100-ms decisions for beam alignment or driving, efficient training, low-FLOP architectures, quantization-aware training, and asynchronous rollouts as continuing problems (Zhao et al., 13 Aug 2025). A plausible implication is that MEGI research will keep converging on decision-oriented abstraction—modeling only what affects rewards or QoS—rather than general-purpose simulation of the entire environment.

A third issue is conceptual: the literature repeatedly pushes back against simple equivalences. DTs and WMs are not interchangeable (Zheng et al., 18 Mar 2026). Edge LLM deployment is not sufficient to constitute MEGI (Chen et al., 2024, Luo et al., 27 Sep 2025). Generative AI at the edge is important, but it covers only one subset of MEGI capabilities (Zhong et al., 2023, Lai et al., 2023). Likewise, knowledge distillation is not merely model compression; in the MEGI framing it is treated as capability transfer across perception, planning, action, and memory loops under wireless constraints (Wu et al., 25 Nov 2025).

The practical roadmaps proposed across the surveys are convergent. They recommend lightweight latent or student models at the device, stronger planning or reasoning services at MEC, cloud-based pretraining and distillation, periodic recalibration against DT ground truths or telemetry, OOD detection, rollback-safe policies, privacy-aware federated updates, and evaluation suites that jointly measure latency, energy, reward or mission success, safety events, outage probability, bandwidth usage, and model drift (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025). Several papers explicitly call for benchmarks blending sensing, communication, and control rather than isolated ML leaderboards (Zheng et al., 18 Mar 2026, Zhao et al., 7 Feb 2026). This suggests that the maturation of MEGI will depend as much on systems methodology and evaluation infrastructure as on any single model class.

In that sense, MEGI designates a research program rather than a finished architecture. The surveyed literature already supplies its principal ingredients—world models, LLM/SLM orchestration, distillation, distributed generative pipelines, RL-enabled scheduling, context engineering, and hybrid DT integration—but also makes clear that scalable, reliable, interoperable, and privacy-preserving edge-native general intelligence remains an active systems problem (Zheng et al., 18 Mar 2026, Chen et al., 2024, Zhao et al., 13 Aug 2025).