TL;DR
CoSlice is a hierarchical multi-agent reinforcement learning (MARL) framework designed to allocate radio resources across 5G network slices. It combines fast, preference-conditioned local agents in the Near-Real-Time (Near-RT) control loop with a Large Language Model (LLM) coordinator in the Non-Real-Time (Non-RT) loop to dynamically balance Service Level Agreement (SLA) satisfaction against resource efficiency.
What Problem Does This Paper Address?
5G networks use "slicing" to run multiple virtual networks—like high-speed video (eMBB) and ultra-low-latency industrial control (URLLC)—on the same physical hardware. This creates a two-fold challenge:
- Intra-slice trade-offs: Each slice must balance its own performance (SLA) against how many Physical Resource Blocks (PRBs) it consumes.
- Inter-slice conflict: All slices compete for a fixed pool of PRBs. If one slice over-requests, others suffer.
Existing solutions often use a single centralized agent (which struggles with the diverse needs of different slices) or independent agents (which fail to coordinate and cause resource contention). Furthermore, network control happens at millisecond scales, while complex global coordination requires heavy computation that can't fit into such tight windows.
Methods
CoSlice introduces a hierarchical architecture aligned with the Open Radio Access Network (O-RAN) standard:
- Preference-Conditioned Agents (Near-RT RIC): Each slice has its own lightweight agent (a Multi-Layer Perceptron). Instead of learning a single fixed policy, these agents are "preference-conditioned." They take a weight vector as input that tells them how much to prioritize SLA satisfaction versus resource saving.
- LLM Coordinator (Non-RT RIC): A Large Language Model (Qwen3-4B) acts as the "brain." Every 10 seconds, it analyzes aggregated network reports and adjusts the preference weights for each slice agent. This allows the system to use the LLM's reasoning capabilities without blocking the millisecond-level resource decisions.
- Theoretical Framework: The authors formulate the interaction as a multi-objective partially observable stochastic game (mo-POSG). They prove the existence of a Preference Correlated Equilibrium (PCE) and demonstrate that their learning procedure achieves a sublinear regret bound, meaning it reliably converges toward optimal behavior.
- Two-Stage Training:
- MORL Pretraining: Agents learn to respond to a wide range of preference signals, effectively mapping out a Pareto front of possible behaviors (see Figure 3/4).
- Joint Optimization: Agents are fine-tuned together using a Centralized Training, Decentralized Execution (CTDE) approach to handle resource competition.
Datasets
The researchers did not use a static dataset but rather a high-fidelity emulation environment:
- ORANSlice: An open-source 5G environment based on the OpenAirInterface (OAI) protocol stack.
- RFsim: A radio-frequency simulator that transports time-domain baseband samples, emulating real-world wireless channel behavior.
- Workloads: They tested two scenarios: a 3-slice setting (eMBB, URLLC, mMTC) and a more complex 6-slice setting with differentiated traffic like live streaming and bulk transfers.
Key Insights and Ideas
- Decoupling Logic and Speed: By placing the LLM in the Non-RT loop and simple MLPs in the Near-RT loop, CoSlice achieves "intelligent" coordination without violating the strict latency requirements of cellular networks (see Figure 2).
- LLMs as Network Reasoners: The LLM can interpret high-level statistics (e.g., "eMBB is meeting its goals but URLLC is failing") and provide interpretable reasons for its preference adjustments, which is a step toward "Explainable AI" in networking.
- Pareto-Optimal Initialization: Pretraining agents to understand the trade-off between performance and cost (the Pareto front) provides a much better starting point for multi-agent coordination than random initialization.
Key Results
CoSlice was compared against several baselines, including Zipper (Model Predictive Control), xSlice (GNN-based RL), and COQRA (Correlated Q-learning).
- SLA Violation Reduction: In the 6-slice setting, CoSlice reduced mean SLA violations by 37.9% and tail (P95) violations by 37.4% compared to the best-performing baseline.
- Ablation Success: Adding the LLM coordinator alone improved performance, but combining it with MORL pretraining yielded the best results, reducing violations by nearly 50% in some metrics compared to a basic multi-agent setup.
- Latency: The local sub-agents have an end-to-end inference latency of only 6.5 ms, well within the requirements for Near-RT control, while the LLM coordinator averages around 508 ms per update cycle.
Limitations and Future Work
- LLM Latency: While the LLM doesn't block the fast loop, its ~500ms inference time limits how quickly the system can react to sudden, massive shifts in global network state.
- Scalability: The experiments were conducted with up to 6 slices. Future work could explore how the LLM coordinator handles dozens or hundreds of slices in a large-scale macro-cell environment.
- Safety Constraints: While the system reduces violations, it does not provide "hard" mathematical guarantees that an SLA will never be violated, which is a common requirement in mission-critical industrial 5G.
Summary updated