Edge General Intelligence

• Edge General Intelligence is a paradigm that equips edge devices with autonomous, multi-modal cognitive capabilities for perception, reasoning, planning, and continual adaptation.
• It leverages world models, agentification, and semantic-context optimization to achieve decentralized multi-task learning under resource-constrained conditions.
• EGI is applied in vehicular networks, IoT systems, UAVs, and wireless communications to enhance latency, energy efficiency, and reliability through coordinated agentic inference.

Searching arXiv for papers on Edge General Intelligence and closely related surveys to ground the article in current literature. Edge General Intelligence (EGI) denotes a paradigm in which edge nodes evolve from platforms for fixed inference or cloud offloading into autonomous agents that can perceive, reason, plan, and act across diverse tasks and time scales under stringent latency, energy, memory, and privacy constraints. Recent surveys describe this shift through several complementary lenses: world-model-driven edge cognition, agentic artificial intelligence, large- and small-language-model deployment, digital-twin integration, personalized federated intelligence, and context-aware orchestration at inference time (Zhao et al., 13 Aug 2025, Zhang et al., 26 Aug 2025, Zheng et al., 18 Mar 2026, Chen et al., 2024). In this formulation, EGI is not merely “AI at the edge,” but an edge-native synthesis of multimodal perception, internal simulation, sequential decision-making, continual adaptation, and decentralized coordination.

1. Definition, scope, and distinction from earlier edge intelligence

EGI extends traditional edge computing and traditional edge intelligence from task-specific inference toward decentralized systems with multi-task generalization, continual learning, and autonomous reasoning. One formalization treats an edge agent as operating in a Markov decision process $(\mathcal S,\mathcal A,P,R,\gamma)$, while another models EGI under partial observability as a POMDP $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$, with belief $b_t(s)=P(s_t\mid h_t)$ over histories $h_t=(o_{1:t},a_{1:t-1})$. In world-model formulations, the latent state $z_t$ plays the role of an approximate belief state, so that planning can proceed in latent belief space rather than directly over raw observations (Zhang et al., 26 Aug 2025, Zhao et al., 13 Aug 2025).

An important precursor is Edge Semantic Cognitive Intelligence, which framed edge intelligence around semantic entropy, semantic cognitive information, semantic compression, and semantic communication for 6G systems. A later step was the explicit taxonomy of LLM-empowered EGI into centralized, hybrid, and decentralized systems, which made the distinction between conventional edge AI and edge general intelligence operational at the systems level (Dong et al., 2022, Chen et al., 2024).

The distinction from traditional edge intelligence can be summarized as follows.

Feature	Edge General Intelligence	Traditional Edge Intelligence
Generalization	Multi-task capability across vision, NLP, planning	Single-task models
Adaptability	Dynamic adaptation via continual/federated learning	Static, requires manual updates
Model Architecture	Compact LLMs, MoEs, foundation models	Small CNNs, RNNs per task
Multi-Modality	Unified processing of text, images, audio	One modality at a time
Autonomy & Reasoning	Goal-driven planning and tool use	Pure inference, limited autonomy
Cloud Dependency	Low, most reasoning local	High, frequent cloud offloads
Cognitive Collaboration	Peer-to-peer knowledge sharing	Independent, cloud-orchestrated

This comparative frame is significant because it separates EGI from mere model miniaturization. EGI is defined not only by deployment location, but by cognitive properties: foresight, contextual adaptation, long-horizon planning, multimodal fusion, and collaborative autonomy.

2. Cognitive substrates: world models, agentification, semantics, and context

A central line of work presents world models as the cognitive backbone of EGI. In that view, an agent maintains an internal model $\hat{T}_\theta$ of the real-world transition $T$, enabling it to imagine future outcomes before acting. The canonical architecture consists of latent representation learning, dynamics modeling, and imagination-based planning. Raw observations $o_t$ are encoded into a compact latent state $z_t$ through an encoder-decoder pair such as

$q_\phi(z_t\mid o_t),\quad p_\theta(o_t\mid z_t),$

with representation learning driven by the evidence lower bound

$(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$0

Given $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$1 and action $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$2, the dynamics model predicts the next latent state through either a probabilistic transition $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$3 or a deterministic approximation $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$4. Planning then searches over action sequences to maximize

$(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$5

Because the rollouts occur in low-dimensional latent space, the agent can evaluate thousands of trajectories in milliseconds, which is presented as critical for real-time edge operation (Zhao et al., 13 Aug 2025).

Agentification generalizes this substrate into a full perception-reasoning-action loop. In the agentic formulation, passive inference pipelines are turned into autonomous agents that perceive multi-modal inputs, reason explicitly through chain-of-thought and tool invocation, plan and decompose tasks, act through control signals or API calls, and coordinate with peers. A representative workflow is: perception $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$6, retrieval-augmented memory $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$7, LLM-driven planning $(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$8, and action through an actor or tool interface. This gives EGI a closed perception-memory-reasoning-action loop rather than an inference-only loop (Zhang et al., 26 Aug 2025).

Semantic and context abstractions form a parallel substrate. Edge Semantic Cognitive Intelligence introduced semantic entropy

$(\mathcal{S},\mathcal{A},\mathcal{O},T,R,O,\gamma)$9

and semantic cognitive information

$b_t(s)=P(s_t\mid h_t)$0

to formalize the distinction between transmitting raw bits and transmitting meaning. Wireless context engineering develops this idea for agentic edge systems by selecting only high-impact context elements under inference-time constraints:

$b_t(s)=P(s_t\mid h_t)$1

Its design principles are selectivity, structuring, compression and prioritization, persistence and aging, and on-demand delivery (Dong et al., 2022, Zhao et al., 7 Feb 2026).

A closely related conceptual transition is the movement from digital twins to world models. Digital twins are characterized as physics-based, centralized, and system-centric replicas; world models as data-driven, decentralized, and agent-centric internal simulators. In hybrid systems, an offline digital twin supplies calibrated models, boundary conditions, safety certificates, or high-fidelity physics, while an online world model embedded within each agent adapts latent dynamics to local conditions. The same literature describes foundation models as providers of declarative knowledge and world models as providers of predictive and prescriptive reasoning, yielding a stack in which declarative, predictive, and prescriptive reasoning co-exist on-device (Zheng et al., 18 Mar 2026, Zhao et al., 13 Aug 2025).

3. System architectures and coordination patterns

EGI architectures are commonly divided into centralized, hybrid, and decentralized systems. In centralized EGI, edge devices collect data and expose APIs, but LLM inference, planning, and tool orchestration occur at a cloud or base-station server. In hybrid EGI, a cloud LLM is paired with an on-device SLM, and a router decides whether to answer locally or offload. In decentralized EGI, each device hosts a local SLM and communicates peer-to-peer, sharing compressed semantic representations rather than raw data. These three archetypes differ primarily in where reasoning resides, how much context is retained locally, and how coordination overhead is distributed (Chen et al., 2024).

Multi-LLM systems refine this taxonomy by decomposing reasoning across specialized agents. A representative edge framework comprises a pool of heterogeneous LLM agents deployed across one or more edge servers or devices, a light-weight director or router, a collaboration bus over which agents exchange intermediate reasoning tokens or confidence scores, a consensus and trust layer optionally backed by blockchain or weighted-Byzantine consensus, and an edge orchestrator that monitors compute, memory, and energy budgets. Dynamic orchestration mechanisms include cascaded inference with early exit, learned router policies that classify requests into “edge-only,” “edge+cloud,” or “ensemble,” distributed layer sharding following EdgeShard by Zhang et al., and confidence-weighted answer aggregation. The same survey specifies quantized or adapter-augmented LLM agents in the 0.5–7 B parameter range (Luo et al., 1 Jul 2025).

Personalized Federated Intelligence extends this systems view by combining foundation models with federated learning and client-specific adaptation. A server distributes a pre-trained foundation model $b_t(s)=P(s_t\mid h_t)$2, clients optimize local objectives over private datasets $b_t(s)=P(s_t\mid h_t)$3, return encrypted or compressed model deltas, and the server aggregates without seeing raw data. Each client then computes a personalized model $b_t(s)=P(s_t\mid h_t)$4 through a regularized local objective

$b_t(s)=P(s_t\mid h_t)$5

The same framework incorporates quantization, pruning, knowledge distillation, parameter-efficient fine-tuning, differential privacy, secure aggregation, and local retrieval-augmented generation, and is explicitly presented as a path from AGI-like foundation models toward artificial personalized intelligence at the edge (Qiao et al., 11 May 2025).

Mobile Edge General Intelligence introduces a BS-centric distributed Mixture-of-Experts architecture for reasoning. In that system, the base-station control unit tokenizes user queries, computes MoE gating scores, batches tokens by expert, schedules downlink transmissions, and re-aggregates outputs returned by distributed edge devices hosting specialized expert networks. The system jointly optimizes token-to-expert assignments, transmission powers, and adaptive Chain-of-Thought depth to minimize energy while satisfying latency, reasoning-quality, compute, and memory constraints. This architecture operationalizes EGI as a coordinated inference-and-reasoning fabric across wireless devices rather than as a monolithic model instance (Luo et al., 27 Sep 2025).

4. Applications and reported performance

World-model surveys emphasize four canonical edge scenarios: vehicular networks, UAV networks, large-scale IoT systems, and network functions virtualization. In vehicular networks, a world-model-augmented scheduler anticipates packet-completeness-aware age (CAoI) under intermittent blockage and rapid channel variation, reducing CAoI by up to 26% over purely model-free baselines. In UAV networks, imagined trajectory planning can halve real-world interactions and improve long-term throughput in weather-aware LAWNs. In large-scale IoT, world models simulate demand trajectories for proactive caching and routing, reducing backhaul usage and delivery latency. In URLLC/eMBB coexistence scheduling, a digital-twin-assisted world model forecasts URLLC bursts and learns slicing strategies robust to demand uncertainty (Zhao et al., 13 Aug 2025).

Agentic AI case studies make these abstractions concrete. In a UAV-assisted IoT network, “Agentic TD3” achieves up to 6.4 % reduction in total energy consumption by using a quantized LLM to shape reinforcement-learning rewards through chain-of-thought prompts and causal reasoning. In intent-driven networking, ACR improves intent-fulfillment accuracy by up to 14.8 % while reducing communication by 23.4 % against keyword-matching and semantic-only retrieval baselines. In vehicular edge computing, an agentic framework using onboard LLAVA and GAE-PPO converges about 61 % faster in cumulative return than pure PPO. In human-centric service provisioning, a preference-guided DRL policy driven by lightweight E-LAMs yields up to 27.3 % improvement in subjective QoE versus uniform-preference baselines (Zhang et al., 26 Aug 2025).

Wireless context engineering supplies a complementary class of results in which model capacity is held fixed and context is optimized instead. In the ISAC-enabled beam prediction case study on the DeepSense 6G V2I Dataset (Scenario 9), accuracy is approximately 62% with only GPS, approximately 74% with GPS+Image, approximately 76% with GPS+LiDAR, and approximately 84% with full observation. The RL-adaptive WCCF policy reaches approximately 82% accuracy with 50–60% of the average cost of full observation, operates within 2–3% of full-observation accuracy while cutting context cost by about 50%, and improves reward by 10–25% compared to single-modality baselines (Zhao et al., 7 Feb 2026).

Digital-twin/world-model integration reports similar gains in mobile wireless systems. In integrated sensing and communications, an adaptive DT+WM system in a UAV swarm achieves a 33% reduction in end-to-end latency and 30% tail-latency reduction at 5% extra energy cost; the world model alone reduces execution latency by 15–20%. In semantic communications, a WM-driven semantic link maintains more than 95% intent recognition accuracy under 30 dB SNR fluctuations and improves semantic fidelity by 20% over static DT-tuned schemes. In air-ground networks, a hierarchical DT+WM system reduces UAV mission failure rate by 40% under sudden interference events compared to DT-only control. In low-altitude wireless networks, WM-enabled obstacle avoidance maintains link stability 85% of the time, versus 60% for centralized replanning (Zheng et al., 18 Mar 2026).

These application studies collectively indicate that EGI research is not confined to abstract reasoning benchmarks. It is being framed as a methodology for long-horizon, uncertainty-aware optimization in wireless systems, robotics-adjacent control, service orchestration, and semantic communication.

5. Resource-efficient deployment and hardware substrates

A defining constraint of EGI is that the target platforms are resource-limited. Recent work emphasizes that edge devices may have no GPUs, stringent latency budgets below 100 ms, and, in 5G/6G settings, real-time constraints in the 1–10 ms range. Edge platforms such as base stations, vehicles, and drones are described as having only a few TOPS of compute and a few GB of RAM, while long-context attention causes super-linear latency growth because of KV-cache and quadratic attention costs. Large models in the 7B–100B parameter range are reported to require 10’s of watts merely to hold weights in memory, which exceeds typical edge budgets (Zhao et al., 13 Aug 2025, Zhao et al., 7 Feb 2026).

Model compression, adaptive inference, and efficient architectures therefore play a structural role in EGI rather than a merely supportive one. Reported methods include LoRA, prompt tuning, and adapters, which update only $b_t(s)=P(s_t\mid h_t)$6 parameters per client; 4-bit weight quantization with “lossless” performance on vision-language tasks; structured pruning that removes 20 % of attention heads with less than 6 % performance loss; early-exit inference that reduces energy by up to 46 %; and invocation-based DRL with DVFS scheduling that cuts invocation cost by 55 % and total energy by 23 %. World-model literature adds Dreamer-style methods with up to 200× data efficiency and lightweight planning variants such as Sparse Imagination and SGF, which achieve sub-10 ms planning at competitive performance (Qiao et al., 11 May 2025, Zhang et al., 26 Aug 2025, Zhao et al., 13 Aug 2025).

Knowledge distillation has been elevated from a generic compression technique to a specific enabler for mobile agentic AI. In wireless settings, cross-model CSI feedback distillation reduces NMSE by up to 7.6 dB while cutting encoder complexity to 25–43% of the teacher’s, and adversarially robust modulation-classification distillation improves PGD-attack accuracy from about 30% to about 78% at 30 dB SNR. The same survey highlights Mamba and RWKV as architectures suited to edge deployment: Mamba is reported to reduce MAC operations by up to 98.9% while matching accuracy on some wireless tasks, and RWKV has been compressed by 3.4×–5× for sensor-network intrusion detection with negligible loss (Wu et al., 25 Nov 2025).

MEGI adds a deployment-level result for reasoning workloads: adaptive MoE+CoT reduces total energy by approximately 30 % versus MoE without CoT and by approximately 45 % versus dense centralized inference, while sustaining accuracy satisfaction above 92 % and latency satisfaction above 90 % under dynamic load. The same work reports local validation on Gemma-2B-base and Gemma-2B-IT, where supervised fine-tuning improves raw answer accuracy by about 15–20 percentage points, and structured Chain-of-Thought adds about 10–20 ms per step, making selective use of CoT preferable for hard queries (Luo et al., 27 Sep 2025).

At the hardware level, AGI-chip research identifies reconfigurability, controlled randomness, rich feedback-feedforward topology, and multi-modal on-chip learning as design requirements for generalizable intelligence on edge hardware. The hardware catalog includes neuromorphic spiking-array chips such as IBM TrueNorth, Intel Loihi, SpiNNaker, and Neurogrid; memristive and in-memory crossbar accelerators; digital neural engines such as Google Edge TPU, NVIDIA Jetson TX2, Movidius Myriad, and Xilinx Versal AI; and emerging photonic and quantum neuromorphic chips. This literature predates current EGI surveys, but it maps directly onto the hardware-software co-design problem now posed by agentic edge systems (James, 2020).

6. Safety, trust, governance, and open research problems

Safety and robustness are recurrent concerns because EGI agents act under uncertainty and may control physical or networked systems. World-model surveys explicitly formulate safety constraints as

$b_t(s)=P(s_t\mid h_t)$7

and describe SafeDreamer as augmenting the world model with a safety critic $b_t(s)=P(s_t\mid h_t)$8 and constrained CEM so that planning searches only among no-violation rollouts. More broadly, open problems include safety guarantees, efficient training from limited edge observations, on-device continual adaptation, communication-efficient federated world-model learning, and constrained deployment on hardware lacking accelerator support (Zhao et al., 13 Aug 2025).

Multi-LLM EGI introduces a distinct threat surface. Reported intra-LLM threats include jailbreaks, prompt injection, emergent abilities, data leakage, and toxic or misaligned output. Inter-LLM threats include expanded attack surfaces, over-permissive integration, insecure communication, consensus manipulation, and cross-context data leakage. In response, zero-trust surveys propose a framework organized around strong identity and authentication, context-aware authorization, inter-LLM communication gateways with end-to-end encryption, prompt sanitization pipelines, multi-layer output verification, and behavioral auditing with anomaly detection. Model-level mechanisms include MFA, reputation-based credentials, ephemeral tokens, Attribute-Based Encryption, hierarchical data caches, PagedAttention or vAttention for cache isolation, BlockLLM micro-segmentation, and self-destructing instances. System-level mechanisms include intelligent input checkers, reputation schemes, topology-aware risk analysis, blockchain-backed interaction logs, threshold ECDSA, network slicing, graph-based anomaly detectors, and meta-LLM monitors (Liu et al., 27 Aug 2025, Luo et al., 1 Jul 2025).

The open research agenda is correspondingly broad. Surveys identify federated world-model learning, meta-learning for rapid adaptation, hierarchical planning, integration with 6G, adaptive and efficient collective intelligence, privacy-preserving federated agent systems, cross-domain adaptation and migration, explainable and trustworthy world models, and standardization through common APIs, versioning schemes, and benchmark datasets. Wireless context engineering adds unresolved questions on regret and sample-complexity bounds for RL-based context selection, as well as information-theoretic limits of context utility under finite token budgets. Personalized federated intelligence contributes concerns about communication cost, heterogeneity, hallucinations, backdoor and poisoning attacks, and fairness-aware objectives (Zheng et al., 18 Mar 2026, Zhang et al., 26 Aug 2025, Zhao et al., 7 Feb 2026, Qiao et al., 11 May 2025).

A plausible implication is that EGI is best understood not as a single model family but as a systems-level convergence. In current formulations, generality at the edge emerges from the interaction of compact world models, language-model reasoning, federated adaptation, semantic and context compression, secure orchestration, and hardware-aware deployment. Whether that convergence yields truly general edge autonomy, robust personalized intelligence, or a family of specialized but interoperable edge-cognitive stacks remains one of the central unresolved questions in the field.