Conversational Orchestration for Organic 6G

The paper proposes a lightweight, decentralized conversational orchestration framework for service provisioning in Organic 6G networks, based on Large Language Model (LLM)-driven domain agents. The core problem addressed is optimizing service provisioning—including placement, user–instance binding, scaling, and migration—over a heterogeneous edge–cloud–NTN continuum composed of independently administered resource domains, under stringent QoS requirements and high dynamism, while satisfying three system-level requirements: scalability, simplicity, and agility.

The proposed architecture places one LLM-based agent at the top of each administrative resource domain. Each agent observes local state via tools, reasons in a closed loop, and exchanges summaries with neighboring agents over an Agent-to-Agent (A2A) overlay aligned with data-plane coupling. The agent has five internal capabilities—goal setting, decision synthesis, memory, action execution, and self-correction—realized through three modules: the LLM itself, a memory store, and a tool layer. An adapter layer connects agents to both modern software-controlled resources (e.g., via MCP) and legacy NMS/API stacks.

Cross-domain coordination uses two complementary message-passing modes: (1) change-triggered dissemination, where agents periodically share compact resource reachability advertisements (combining end-to-end latency, path bottleneck bandwidth, and available compute capacity) with neighbors, who perform routing-style updates to build a distributed resource reachability table; and (2) on-demand request/negotiation, where agents handle per-request provisioning by consulting the reachability table to select feasible remote destinations, forwarding requests hop-by-hop with soft reservations, and use event-driven negotiation for safe re-optimization, scaling, and migration.

To meet real-time constraints, each domain agent is instantiated with a compact Small Language Model (SLM) (DeepSeek-R1-Distill-Qwen-7B in evaluation), specialized via offline self-verification training using a stronger verifier LLM (DeepSeek-R1) that assigns multi-objective rewards (reasoning validity, optimization quality, QoS feasibility) driving Group reward-Decoupled Normalization Policy Optimization (GDPO) updates. Online, the SLM is periodically refined via shadow updates: inference traces are batched, a separate copy is updated offline, and the updated copy periodically replaces the working model.

Simulations evaluate two complementary aspects. Scenario A assesses control-plane overhead: with N ∈ 10, 20, 30 domains and average degree 4, message-passing overhead remains manageable, increases approximately linearly with domain count, and shows a smaller transient after a domain join at time slot 60. Scenario B validates training and refinement: offline RL specialization brings the SLM close to verifier-level performance on a load-balance objective (epochs 0–50); after switching to a min-latency objective (epochs 50–100), only the variant with online refinement recovers toward the baseline, while the offline-only variant drops and remains lower.

The paper concludes by outlining future research directions: formal cross-domain agentic orchestration models, uncertainty-aware decision theory for tail-risk QoS guarantees, security for agentic control planes (e.g., malicious agents injecting false advertisements), multi-agent stability and conflict resolution, mechanism design for truthful coordination, and learning/adaptation under real-time constraints.

Improvements to AI Systems:

1. Hierarchical Multi-Agent Orchestration with Localized Reasoning: Implement a decentralized system where each administrative domain (edge, cloud, NTN) runs a dedicated SLM-based agent. Each agent only observes local state (via tool APIs) and exchanges compact summaries with neighbors, avoiding global state aggregation. This enables scalable coordination across hundreds of domains without a central bottleneck.

2. Dual-Mode Communication Protocol for Dynamic Resource Management: Integrate (a) periodic, change-triggered dissemination of resource reachability advertisements (latency, bandwidth, compute) to build a distributed routing table, and (b) on-demand request/negotiation for per-task provisioning with soft reservations. This allows the system to handle both steady-state optimization and sudden spikes (e.g., domain joins, traffic surges) with minimal overhead.

3. Self-Correcting Agent Loop with Memory and Tool Integration: Each agent uses a closed-loop cycle: goal setting → decision synthesis → memory retrieval → action execution → self-correction. The memory store retains past decisions and outcomes, enabling the agent to avoid repeated mistakes and adapt to changing QoS conditions. The tool layer (MCP for modern resources, legacy NMS/API adapters) ensures compatibility with heterogeneous infrastructure.

4. Offline RL Specialization with Online Shadow Refinement: Train a compact SLM (e.g., 7B parameters) using Group reward-Decoupled Normalization Policy Optimization (GDPO), where a stronger verifier LLM provides multi-objective rewards (reasoning validity, optimization quality, QoS feasibility). Then, continuously refine the SLM online via shadow updates: batch inference traces, retrain a copy offline, and periodically swap it in. This maintains high performance under shifting objectives (e.g., from load balancing to min-latency) without retraining from scratch.

5. Distributed Reachability Table with Soft Reservations: Implement a routing-style update mechanism where each agent maintains a table of feasible remote destinations (end-to-end latency, bottleneck bandwidth, available compute). When a request arrives, the agent selects the best remote domain, forwards the request hop-by-hop with soft reservations (temporary holds), and uses event-driven negotiation for safe re-optimization, scaling, or migration—preventing resource conflicts and over-commitment.

6. Proactive Anomaly Handling and Conflict Resolution: Add a negotiation layer that triggers when QoS violations are predicted (e.g., latency threshold breach). Agents exchange compensation offers (e.g., migrate a workload to a neighbor with spare capacity) and resolve conflicts via priority-based arbitration, ensuring stability even when multiple agents act concurrently.

What the Improved AI System Can Do:

• Autonomous Service Provisioning Across Heterogeneous Networks: Dynamically place, bind, scale, and migrate services (e.g., VNFs, AI inference tasks) across edge, cloud, and satellite domains, meeting strict latency and bandwidth SLAs without human intervention.

• Real-Time Adaptation to Network Dynamics: Respond to domain failures, load spikes, or new resource joins within seconds by updating reachability tables and triggering negotiations, while keeping control-plane message overhead linear with domain count (e.g., <10% extra traffic for 30 domains).

• Continuous Learning Under Changing Objectives: Switch optimization goals (e.g., from energy efficiency to latency) on the fly, recovering performance within 10–20 epochs via online refinement, unlike static models that degrade permanently.

• Interoperability with Legacy and Modern Infrastructure: Seamlessly control both software-defined resources (via MCP) and legacy NMS/API stacks, enabling deployment in real-world telecom operators without forklift upgrades.

• Fault-Tolerant and Secure Coordination: Detect and mitigate malicious agents (e.g., false advertisements) via cross-validation of reachability data and reputation-based trust, while maintaining graceful degradation if a domain goes offline.

• Scalable to Hundreds of Domains: Operate with a fully distributed architecture, where each agent only communicates with 3–5 neighbors, ensuring that coordination overhead remains manageable even as the network grows to 100+ administrative domains.