TLDR: A new agentic controller uses a Large Language Model (LLM) to manage Wi-Fi and 5G NR-U coexistence in the 6 GHz band. The LLM proposes high-level policy “knobs” like fairness index and duty-cycle caps based on network telemetry. A deterministic optimizer then enforces these policies safely, balancing throughput, energy efficiency, and service-level agreements. Simulations show LLM-assisted policies significantly improve energy efficiency and maintain competitive throughput compared to traditional rule-based systems, offering a transparent and adaptive solution for complex wireless environments.
The rapid expansion of wireless communication has made the unlicensed 6 GHz spectrum a crucial resource for high-capacity access. In this band, technologies like Wi-Fi 6E/7 and 5G NR-U (New Radio Unlicensed) are designed to coexist, sharing the same channels under specific rules known as listen-before-talk (LBT). This shared environment presents a complex challenge: how to manage these competing technologies to optimize for various objectives such as data throughput, energy efficiency, and meeting service-level agreements (SLAs), all while ensuring the system remains stable and auditable.
Traditional wireless schedulers often focus on a single objective, like maximizing data rates, and use static rules to prevent collisions. While effective in certain situations, these methods struggle to adapt when network conditions change or when different priorities, such as energy saving or urgent latency demands, arise. On the other hand, advanced learning systems like deep reinforcement learning (DRL) can learn complex strategies but are often “black boxes,” making their decisions hard to understand, verify, or trust in real-world networks. They also require extensive training.
A new research paper introduces an innovative solution: an agentic controller that uses a Large Language Model (LLM) to manage this complex coexistence. This system separates high-level strategic decision-making from low-level execution. At the beginning of each scheduling period, the agent collects detailed network information, including channel activity, LBT failure rates, and per-user data like signal quality (CQI), data backlog, latency targets, battery status, and priority. This telemetry is then summarized into a compact JSON format.
The LLM, acting as a “policy layer,” analyzes this summary and proposes a small set of understandable control “knobs.” These include a fairness index (alpha), duty-cycle caps for Wi-Fi and NR-U on each channel, and priority weights for different traffic classes (e.g., emergency, high, normal, bulk). This approach avoids the rigidness of fixed rules and the opacity of DRL, offering a system that is adaptive, transparent, and capable of intelligently balancing conflicting objectives dynamically.
Crucially, the raw policy proposed by the LLM is never directly implemented. Instead, a deterministic optimizer takes these proposed knobs and enforces feasibility and safety. This means that the fairness index is restricted to safe values, duty caps are adjusted to prevent over-utilization and ensure headroom, and priority weights are kept within reasonable bounds. If the LLM fails or provides an invalid policy, the system automatically falls back to a reliable rule-based policy, guaranteeing continuous and safe operation. This separation ensures that the LLM is responsible only for high-level, transparent decisions, while a verifiable executor handles hard constraints.
The optimizer then proceeds in two stages. First, it assigns each user to a single channel by evaluating a utility score that considers data throughput, energy cost, deadline pressure, and user priority/battery levels. Second, within each assigned channel and technology stack, it first grants minimum airtime to meet urgent latency targets for critical users. The remaining airtime budget is then distributed using a weighted alpha-fair rule. After these allocations, the system recalculates LBT losses and determines the final data throughput and energy consumption for each user.
The researchers, Qun Wang, Yingzhou Lu, Guiran Liu, Binrong Zhu, and Yang Liu, evaluated this LLM-assisted agent in a simulated 6 GHz environment with two 160 MHz channels and a mix of Wi-Fi and NR-U users. They compared its performance against a strong rule-based baseline and found significant improvements.
Under moderate network load, LLM-assisted policies consistently reduced cumulative energy consumption and improved energy efficiency (bits per Joule) compared to the rule baseline. One LLM variant (GPT-4o-Mini) lowered total energy by over a third, albeit with a slight reduction in throughput. Another variant (GPT-5-Mini) achieved the best overall balance, delivering higher total data bits (+3.5%) and significantly higher energy efficiency (+12.2%) than the baseline.
In scenarios with high network load, the advantages of the LLM-assisted policies became even more pronounced. GPT-5-Mini maintained the fastest cumulative throughput growth and finished with the highest total bits, while GPT-4o-Mini continued to offer substantial energy savings. Both LLM policies kept cumulative energy lower than the rule baseline, which tended to use aggressive duty cycles leading to higher LBT losses. The energy efficiency curves consistently showed LLM-assisted control outperforming the baseline.
The key takeaway is that the LLM’s ability to propose headroom-aware duty caps and fairness regimes closer to proportional fairness helps keep the system away from congested operating points where increased airtime yields little useful data but consumes significant energy. This intelligent, high-level guidance from the LLM, combined with a robust, deterministic optimizer, demonstrates a promising path for safely improving wireless coexistence in complex shared spectrum environments.
Also Read:
- Distributed AI Agents Enhance Autonomy and Safety in Future RANs
- Optimizing UAV-Assisted Communication for Search and Rescue with Reinforcement Learning
You can find more details about this research in the full paper available at arXiv:2510.17814.
