Source-linked AI summary
Smart Adaptive Computing Across the Continuum: LLMs in IoT-Edge-Cloud Resource Management
Antonino Vaccarella, Lanpei Li, Vincenzo Lomonaco, Massimo Coppola
TL;DR
Cloud Continuum resource management requires adaptive decisions under changing constraints, while the architectural relationship between DRL and LLM components remains unclear. The paper extends a DRL taxonomy with LLM-role and feedback dimensions, applies it to six systems, and identifies a missing combination of full LLM orchestration and agent-layer feedback in the Cloud Continuum.
Problem
The architectural role of LLMs in DRL-based Continuum resource management and the feedback path to the LLM remain underspecified.
Method
The paper extends Wang et al.’s taxonomy with AI Augmentation Paradigm and Feedback dimensions and applies it to six recent system architectures.
Results
None of the surveyed systems combines full LLM orchestration with full agent-layer feedback in a Cloud Continuum setting.
Takeaways & Limitations
The gap points to a missing cross-tier feedback abstraction that reconciles heterogeneous execution signals for the LLM orchestrator.
Takeaways & Limitations
The surveyed literature lacks a demonstrated LLM2×MA×Standard Cloud Continuum architecture, and execution outcomes must be compressed into prompts while controlling inference cost.
Abstract
from arXiv · showhide
Managing resources across IoT, edge, and cloud layers calls for continuous, context-aware decisions under constraints that rarely stay fixed. Deep reinforcement learning (DRL) handles this class of problems well, and large language models (LLMs) are increasingly used to augment DRL pipelines, yet the architectural relationship between the two is seldom made explicit. We build on Wang et al.'s taxonomy of Continuum Orchestration Systems employing DRL techniques and extend it with two further dimensions. The AI Augmentation Paradigm measures how LLMs are exploited, while the Feedback channel captures whether and through which system path the execution feedback returns to the LLM in order to close the MAPE control loop at the LLM Orchestration layer. We apply this taxonomy to six recent system architectures and find a common gap, as none combines full LLM orchestration with full agent-layer feedback in a Cloud Continuum setting. We relate this gap to a missing cross-tier feedback abstraction, bridging the incommensurable per-tier signals and the LLM Orchestrator.
1 Introduction
Cloud Continuum resource management requires adaptive, multi-objective decisions under changing system conditions. The paper extends an existing DRL taxonomy to make LLM roles and feedback paths explicit.
- Cloud Continuum management spans placement, scheduling, scaling, and routing under changing load, mobility, network, hardware, policy, and security constraints.
- DRL fits this setting because it optimizes temporal decision sequences and learns from system interaction without a hand-crafted dynamics model.
- Existing systems use LLMs as decision-makers, orchestrators, or closed-loop components, but their architectural relationship with DRL remains underspecified.
- The paper extends Wang et al.’s taxonomy with AI Augmentation Paradigm and Feedback dimensions, then applies it to six recent works.
- LLM roles and control-loop design condition the capabilities of an LLM as an orchestrator, making feedback classification central to current and future systems.
2 Cloud Continuum Orchestration
Cloud Continuum orchestration extends cloud management across heterogeneous, distributed devices and introduces dynamic, multi-factor constraints. The existing DRL taxonomy describes controller organization and training, but not the architectural role of LLMs.
- Continuum platforms combine data-center, local cloud, vehicle, IoT, mobile, and other devices across distributed layers.
- Continuum orchestration must optimize across complex constraints including user intent, security vulnerabilities, and computing, networking, and storage configuration.
- Reinforcement-learning agents address the need to react quickly and adapt to complex, unanticipated behavior.
- Wang et al.’s taxonomy separates single-agent from multi-agent RL and standard from federated training.
- The bidimensional taxonomy captures DRL controller organization and training but does not describe an LLM’s architectural role in the control loop.
3 Generic architecture and AI Augmentation Paradigm
The paper models Continuum orchestration as layered control: an LLM orchestration layer directs autonomous agents that execute in environments and return feedback. It classifies LLM integration by delegation and feedback by the execution path returning information to the LLM.
- Orchestrator and architecture: A Continuum orchestrator devises allocation plans for distributed applications and initiates actions as application needs and device availability evolve.
- Orchestrator and architecture: The generic architecture contains an orchestration layer of LLMs, an agentic layer of specialized agents, and controlled environments.
- Orchestrator and architecture: The layer cardinalities x, y, and z need not match, and the architecture leaves the entities’ interaction graph unspecified.
- Orchestrator and architecture: In one concrete LLM2 configuration, a single orchestrator coordinates n agents in a one-to-one relationship with n environments, while agents return observations and data.
- Feedback channel: A closed loop updates controller decisions from execution feedback, enabling dynamic adaptation; the paper focuses on feedback reaching the topmost LLM layer.
- Feedback channel: Feedback may come directly from controlled systems, through agents, or through dedicated mediators that filter and process platform and agent outputs.
- Feedback channel: F0 excludes post-execution environmental feedback, while F1, F2, and F3 route feedback through platform APIs, executing sub-agents, or dedicated evaluation components.
- AI Augmentation Paradigm: LLM0 has no LLM, LLM1 makes and enacts decisions directly, and LLM2 delegates execution to autonomous downstream agents.
4 Extending the Taxonomy towards Upcoming LLM Approaches
The extended taxonomy adds LLM role and feedback dimensions to classify how LLMs and downstream agents participate in continuum orchestration. Across six systems, the classification distinguishes direct, mediated, and absent feedback while revealing sparsely populated LLM1–LLM2 regions.
- Taxonomy extension: The taxonomy retains Control Scope and Training Paradigm while adding AI Augmentation Paradigm and Feedback Channel dimensions.The AI Augmentation Paradigm captures LLM levels, while the Feedback Channel captures how execution data returns to the LLM.
- Taxonomy extension: LLM0 corresponds to SARL and MARL, whereas LLM1–LLM2 use broader SA/MA labels because their controllers may be LLMs or RL agents.The LLM level identifies the kind of controller or augmentation present in the architecture.
- Taxonomy findings: LLM1–LLM2 occupy largely unpopulated territory, with the LLM2-MA row containing the most significant structural gaps.The taxonomy is intended to expose unexplored architectural combinations, not rank system performance.
- Taxonomy findings: Among six classified works, three are LLM1 and three are LLM2; only two LLM2 systems close the feedback loop at the orchestration level.None uses federated learning, and only the three LLM2 systems train downstream RL agents.
- System classifications: AgentEdge is LLM1×F0 because its four sequential LLM stages lack an autonomous Agentic Layer and receive no post-execution signal.Its planning loop operates on virtual state through simulation and critique rather than execution feedback.
- System classifications: IntentContinuum and SALLMA are LLM1×F1, while Habib et al. is LLM2×F0 because feedback is direct in the former systems but absent after delegation in the latter.These classifications distinguish deterministic orchestration from autonomous downstream agents and open-loop handoffs.
5 Gap Analysis and Research Directions
The surveyed literature reveals gaps in feedback, Cloud Continuum coverage, federated training, latency, cost, and cross-tier feedback abstraction. In particular, no classified system combines full LLM orchestration with agent-layer feedback in a Cloud Continuum setting.
- Feedback gaps: Four of six classified works provide no post-execution environmental feedback or only platform-level signals to the LLM.Only AURA and CyberOps-Bots close the agent-layer feedback loop, and both are LLM2 systems.
- Continuum coverage: AURA and CyberOps-Bots combine LLM2 orchestration with agent-layer feedback but remain confined to single deployment tiers.Their patterns do not address multi-tier heterogeneity, placement, SLOs, tenancy, cost models, or SDN routing in the Cloud Continuum.
- Federated training: No work combines LLM2 orchestration with federated training of the agent layer, although privacy and locality concerns motivate that direction.AgentEdge identifies federated orchestration as a future direction, while CyberOps-Bots names LLM centralization as a scalability concern.
- Engineering challenges: Direct LLM deployment for fine-grained real-time control remains infeasible because computational overhead and latency would stall the control loop.The proposed structural response is asynchronous LLM operation over episode summaries while faster local agents continue under the last directive or shaping signal.
- Engineering challenges: LLM involvement also introduces monetary cost because each orchestration step is a billed API call whose total expense scales with control frequency and agent count.Threshold-triggered consultation can reduce paid calls by invoking the LLM only when feedback deviates beyond a defined margin.
- Feedback abstraction: Cloud Continuum agents expose heterogeneous, often incommensurable metrics, and no established cross-tier feedback abstraction yet unifies them for directive revision.This missing abstraction is identified as a remaining engineering challenge for closing the orchestration loop.
6 Toward Closed-Loop LLM Orchestration in the Cloud Continuum
The paper distinguishes semantic front-ends, workflow pipelines, and strategic orchestrators by how the LLM remains connected to execution. Only strategic LLM2 orchestrators revise subgoals or reward-shaping signals from agent feedback, yet current F2/F3 systems do not span the Cloud Continuum; the proposed direction is a cross-tier mediator.
- Architectural patterns: Semantic front-ends compile operator intent into executable goals, while workflow pipelines produce structured deployment plans before handing execution downstream.IntentContinuum and Habib et al. exemplify the first pattern; SALLMA and AgentEdge exemplify the second.
- Architectural patterns: Strategic orchestrators issue subgoals or reward-shaping signals and revise them as delegated agents report back.This is the only surveyed pattern that keeps the LLM coupled to the agent layer after the first directive.
- Feedback architecture: Every F2/F3 system in the survey is an LLM2 orchestrator, whereas no front-end or pipeline exceeds F1 feedback.Thus agent-mediated orchestration-level control loops require the strategic orchestrator pattern.
- Open gap: No LLM2 system with F2 or F3 feedback addresses the Cloud Continuum because heterogeneous per-tier signals lack a unified abstraction the orchestrator can use.AURA and CyberOps-Bots build such abstractions within single tiers, where cross-tier translation is unnecessary.
- Proposed architecture: The reference architecture introduces a Mediator that collects diverse outcomes from tier-specialised RL agents supervised by the orchestration layer.The mediator extends AURA’s single-tier alignment controller toward cross-tier feedback.
- Proposed architecture: The cross-tier mediator ingests heterogeneous signals and produces structured representations pairing performance scores with the context needed to revise directives.Without it, the orchestrator cannot determine which tier or policy requires adjustment.
7 Conclusion
The taxonomy separates LLM augmentation from feedback channels and reveals that no surveyed system closes the loop across IoT, edge, and cloud. The paper characterizes a mediator as the needed abstraction for reconciling heterogeneous cross-tier signals.
- The extended taxonomy adds AI Augmentation Paradigm and Feedback Channel dimensions to classify how LLMs orchestrate agents and receive execution feedback.It was applied to six recent systems to identify architectural patterns and literature gaps.
- None of the six surveyed systems combines multi-agent LLM orchestration with closed-loop feedback across IoT, edge, and cloud tiers.Closed-loop orchestration remains confined to a single resource tier, while intent-driven Continuum management is open-loop or uses platform-level feedback.
- A mediator function is needed to reconcile heterogeneous signal sources before feedback reaches the LLM orchestrator in the LLM2-MA-Standard Cloud Continuum cell.The paper presents this as a missing feedback abstraction rather than a settled negative result.
- Extending the taxonomy to future LLM orchestrators with distributed, federated, or continual-learning agent layers remains future work.