When everyone started stuffing LLMs into control loops, the actual deployments started revealing something: in most setups, the feedback the model receives is just the prompt it just sent. A recent paper from the Italian National Research Council's ISTI team, presented at the FRAME 2026 workshop co-located with Euro-Par 2026, makes this gap brutally clear. They audited six systems that pair LLMs with deep reinforcement learning (DRL) for Cloud Continuum orchestration, and the verdict is stark: four out of six give the LLM no feedback at all, two manage closed loops but only within a single resource tier, and the cross-tier loop spanning IoT-edge-cloud remains entirely blank.
Two new dimensions: what the LLM does, and how feedback travels
Existing literature classifies DRL orchestration systems along two axes: control scope (SARL vs MARL) and training paradigm (standard vs federated). This taxonomy is completely blind to LLMs. Vaccarella et al. add two new axes: the AI Augmentation Paradigm (L0–L2) captures whether the LLM is a semantic front-end that interprets intent, a workflow step that issues instructions, or a true strategic orchestrator over a multi-agent layer; the Feedback Channel (F0–F3) captures the path execution results take to return to the LLM. F0 is no feedback, F1 is direct platform API delivery, F2 is an agent-layer relay, F3 is an independent Mediator producing a feedback abstraction.
How the six systems actually stack up
Mapped onto the two new axes, the distribution of the six representative systems is striking. AgentEdge strings four LLM stages (Intent, Observability, Planning, Infrastructure Action) in the orchestration layer, with no independent agent layer, and no execution results flow back: L1×F0, really a single-layer LLM pipeline. IntentContinuum has GPT-4o read structured JSON from three platform APIs (Kubernetes, ONOS, sFlow) to diagnose root causes: L1×F1. SALLMA presses the entire LangChain-orchestrated workflow into the orchestration layer with a passive knowledge layer, L1×F1. Habib et al. for 5G RAN automation has ALBERT parse intent then hand off to a hierarchical DQN, but the DQN never reports back: L2×F0, an open loop. The only systems that actually close the loop are AURA and CyberOps-Bots: AURA uses Claude Sonnet 4 to orchestrate at base stations, with signals flowing back to an independent Centralised Alignment Controller that returns a delayed reward in [-1, +1], L2×F3; CyberOps-Bots uses Qwen3-8B with four classes of RL agents for cloud-network resilience, L2×F2. Both work, but both stay inside a single resource tier. Nobody has closed the loop across IoT-edge-cloud heterogeneity.
What's blocking the multi-tier loop: latency mismatch and feedback abstraction
Why is the cross-tier loop still empty? The paper identifies three engineering constraints, and the first two are brutally real. First, latency mismatch: putting an LLM on the critical path of a real-time control loop kills the loop, since even the API latency of a frontier model overwhelms the response budget. The proposed workaround is to have the LLM run asynchronously on episode summaries, while local agents continue under the last issued subgoal. Second, money: AURA's Claude Sonnet 4 is priced at $3 per million input tokens and $15 per million output tokens. The paper estimates a 10K input + 1K output call at roughly $0.05, and at continuous control frequencies, daily cost can hit hundreds of dollars. The fix is threshold-triggered invocation, where the LLM only re-plans when feedback deviates from the current directive beyond a defined margin. Third, feedback abstraction: IoT, edge, and cloud agents expose incommensurable metrics (latency, energy, container SLOs, SDN path costs), and someone needs to compress these into a prompt-sized representation the LLM can act on. AURA's Centralised Alignment Controller does this in a single tier, but no one has extended it across the Continuum.
What to watch next: the Mediator and L2×MA×Federated
The paper names the cross-tier closed loop as the L2×MA×Federated cell, with the missing engineering interface called the Mediator, a cross-tier, independent feedback aggregator that compresses each tier's heterogeneous KPIs into a unified score-plus-context pair, which the orchestration LLM consumes to revise the next round of subgoals. It is structurally the same idea as AURA's CAC, but extended from one tier to three. CNR's generic reference architecture (Figure 1(a) of the paper) shows the Mediator standing between the LLM Orchestration layer and the Agentic/Environments layers, collecting structured feedback from specialised agents and producing a unified representation. The next iteration will appear in the formal Euro-Par 2026 proceedings; if the team or a follow-up paper actually runs the L2×F3 cross-tier benchmark, the Cloud Continuum orchestration line will finally move from "LLM as semantic front-end" to "LLM as control core."
For engineering teams working on cloud-edge orchestration today, the immediate takeaway is: don't rush to make the LLM close the loop yet. First build the feedback abstraction layer. Single-tier CACs already exist; cross-tier Mediators are the research direction worth watching over the next 12 months.