⚡ electrical engineering

MAxLM: Multi-Agent Language Model-Based Scheduling and Resource Allocation in MU-MIMO-OFDMA-Enabled Wireless Networks

This paper proposes MAxLM, a multi-agent framework leveraging a pretrained small-to-medium language model within the AI-assisted WiSER platform to optimize user scheduling and resource allocation in uplink MU-MIMO-OFDMA WLANs, demonstrating superior throughput performance compared to benchmark techniques.

Original authors: Adnan Quadri, Hongxiang Li

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adnan Quadri, Hongxiang Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy Wi-Fi router (the Access Point) trying to talk to many different devices (like phones and laptops) at the same time. In the old days, the router had to take turns talking to each device, which was slow. But modern technology (called MU-MIMO and OFDMA) allows the router to talk to several devices simultaneously, like a conductor leading an orchestra where every musician plays a different instrument at the same time.

However, this is a very difficult job for the router. It has to decide:

Who gets to talk right now?
Which "frequency lane" (Resource Unit) should they use?
How should they share the airwaves so they don't crash into each other?

Doing this math perfectly is incredibly hard and slow, like trying to solve a giant, shifting Sudoku puzzle every second.

The Solution: A Team of "Smart Assistants"

The authors of this paper, Adnan Quadri and Hongxiang Li, propose a new way to solve this puzzle. Instead of the router doing all the heavy math itself, they use a Multi-Agent Language Model (MAxLM) system.

Think of this system as a team of smart, specialized assistants (Agents), where each assistant represents the connection between the router and one specific device.

Here is how their system, called WiSER (Wireless Systems Engineering and Research), works using a creative analogy:

1. The "Context Manager" (The Briefing Room)

Before the assistants make a decision, they need to know the current situation. The WiSER platform gathers data about the "weather" of the Wi-Fi signal.

The Analogy: Imagine the assistants are pilots. Before takeoff, they get a briefing that says, "Your engine is strong, but the wind is blowing from the left, and your neighbor's plane is a bit shaky."
The Paper's Claim: The system uses a special "Adaptive Context Management" procedure to translate complex signal data into a simple story (a "prompt") that a small, pre-trained AI language model (xLM) can understand. It tells the AI: "You are strong, but you are compatible with these other devices, and not those ones."

2. The "Scheduler" (The Decentralized Decision)

Instead of one big brain trying to figure out the schedule for everyone at once, each assistant (Agent) makes its own recommendation based on the briefing.

The Analogy: Instead of one conductor shouting out every note, every musician in the orchestra looks at their sheet music and the weather, then whispers, "I think I should play my solo on the 3rd and 5th beats."
The Paper's Claim: The AI models (like Mistral-NeMo or Llama) read the prompts and output a plan in a structured format (JSON). They decide which "Resource Units" (frequency lanes) to assign to their specific device.

3. The "Parser" and "Self-Correction" (The Safety Net)

Sometimes, the assistants might get a little too excited and suggest a plan that breaks the rules (like trying to put 10 people in a car that only holds 8).

The Analogy: A safety inspector checks the orchestra's plan. If the plan says "Everyone play the 3rd beat," but the room only has space for 4 people, the inspector says, "Wait, that's too many! Let's remove the extra people from that beat."
The Paper's Claim: The system automatically checks the AI's suggestions against the physical limits of the router (the MIMO constraint). If the AI suggests too many devices for one frequency lane, the system "revokes" those assignments and fixes the plan before sending it out.

What Did They Find?

The researchers tested this system in a simulated environment with different numbers of devices (10 or 20) and different router strengths (4 or 8 antennas).

The "Prompt" Matters: They found that how they asked the AI the question was crucial. If they just gave the AI raw numbers (like a spreadsheet), it made mistakes about 40% of the time. But if they gave the AI a story describing the device's strength and how well it fits with others (the "Semantic Analysis"), the AI made almost no mistakes.
Better Than the Old Way: They compared their AI system to the standard "Best Channel Quality" (BCQ) method, which simply picks the devices with the strongest signals.
- In many scenarios, their AI system (specifically using the Mistral-NeMo model) achieved up to 30% more data throughput than the standard method.
- This means more data gets sent, faster, and more efficiently.

The Bottom Line

The paper claims that by using a team of small, smart AI agents that "talk" to each other through a structured platform (WiSER), we can solve the complex puzzle of Wi-Fi scheduling much better than current methods. The key is not just using AI, but telling the AI the right story about the network environment so it can make smart, decentralized decisions that the router can then verify and execute.

The result is a Wi-Fi network that is faster and smarter at handling many users at once, without needing a supercomputer to do the math.

Technical Summary: MAxLM: Multi-Agent Language Model-Based Scheduling and Resource Allocation in MU-MIMO-OFDMA-Enabled Wireless Networks

Problem Statement
In joint Multi-User Multiple-Input Multiple-Output (MU-MIMO) and Orthogonal Frequency-Division Multiple Access (OFDMA) enabled Wireless Local Area Networks (WLANs), optimizing User Scheduling and Resource Allocation (SRA) for Uplink Scheduled Access (UL-SA) is a complex combinatorial problem. The Access Point (AP) must simultaneously schedule multiple stations (STAs) across time, frequency (Resource Units or RUs), and space (antennas) to maximize throughput. Traditional optimization methods require computationally expensive exhaustive searches to determine spatially compatible groups of STAs and their power distributions, particularly because the AP cannot fully evaluate the impact of its scheduling strategy on uplink channel conditions and interference prior to transmission. While Deep Learning and Multi-Agent Reinforcement Learning (MARL) have been proposed, they often lack practical amenability or rely on computationally intensive frontier models.

Methodology
The authors propose MAxLM, a framework that leverages openly available, pretrained small-to-medium-sized Language Models (xLMs) to perform decentralized SRA. The core approach involves modeling the wireless channel between the AP and each STA as an independent agent within a Multi-Agent system.

Multi-Agent Framework: Instead of the AP making a single complex decision for all users, each "agent" (representing an AP-STA connection) independently determines its RU assignments. This reduces the action space complexity compared to a centralized approach where the AP must select from all possible MIMO user group combinations.
State Representation: Each agent's state includes:
- Channel Gains ( $\zeta$ ): The Frobenius norm of channel states over available RUs.
- Impact Factors ( $\eta$ ): A measure of spatial compatibility, indicating how uncorrelated an agent's channel is with others.
- Feedback ( $F$ ): System notifications regarding parsing success or errors.
The WiSER Platform: To facilitate autonomous execution, the authors introduce the AI-assisted Wireless Systems Engineering and Research (WiSER) platform. Its workflow includes:
- Context Manager: Uses an Adaptive Context Management (ACM) procedure to construct prompts for the xLM. It translates raw channel data into semantic analyses (e.g., "Agent is one of the M strongest agents") and numerical data, ensuring the xLM understands the dynamic environment.
- Scheduler: Asynchronously loads prompts for all agents into the xLM's context window.
- Parser: Converts the xLM's natural language or JSON-formatted intent into binary RU assignment matrices.
- Environment Manager & Self-Correction: Simulates the transmission and validates assignments. Crucially, it includes a self-correction step that revokes any RU assignments violating the MIMO spatial constraint (i.e., assigning more STAs to a single RU than the number of AP antennas, $M$ ).
Prompt Engineering: The study compares two prompt templates: one providing semantic analysis of agent attributes (Template 1) and one providing raw numerical arrays (Template 2). Results indicate that semantic prompts significantly reduce inference errors.

Key Contributions

Decentralized SRA via xLM: The paper adopts a multi-agent framework using small/medium-sized Language Models to simplify the combinatorial SRA optimization problem, moving away from computationally intensive frontier AI models.
Adaptive Context Management (ACM): A procedure developed to effectively prompt the xLM with updated WLAN environmental information, enabling the model to anticipate effective SRA strategies before UL-SA occurs.
WiSER Platform: The introduction of a modular platform (available at [9]) that orchestrates the workflow of xLM-based SRA, including context management, asynchronous scheduling, parsing, and self-correction.
Empirical Validation: The study evaluates the performance of MAxLM across varying numbers of STAs (10 and 20) and AP antenna configurations (4 and 8 antennas).

Results
The performance of MAxLM-optimized SRA was evaluated against a benchmark Best Channel Quality (BCQ) scheme, which assigns RUs to STAs with the strongest channel gains.

Prompt Efficiency: Using semantic prompt templates (Template 1) significantly reduced assignment errors compared to raw numerical prompts (Template 2). For instance, with a 4-antenna AP, error rates dropped from ~43% (Template 2) to <1% (Template 1) for the Mistral-NeMo model.
Throughput Gains:
- In scenarios with 10 stations and an 8-antenna AP, the Mistral-NeMo:12b model achieved a 10% performance gain in UL rate-sum over the BCQ benchmark.
- In scenarios with 20 stations and a 4-antenna AP, Mistral-NeMo achieved a 30% performance gain over the benchmark.
- The Gemma model showed high variance; while it achieved gains in some cases, it frequently violated MIMO spatial constraints, leading to self-correction steps that revoked assignments and resulted in zero rate-sums for specific transmission slots.
Consistency: Mistral-NeMo demonstrated consistent performance across all tested scenarios, outperforming both the BCQ benchmark and other xLMs (Llama, Gemma).

Significance and Claims
The paper claims that the proposed MAxLM framework successfully addresses the complexity of combinatorial SRA in joint MU-MIMO-OFDMA WLANs by leveraging generative AI. The significance lies in demonstrating that small/medium-sized, openly available Language Models can effectively perform complex network resource management tasks when guided by a structured multi-agent framework and adaptive prompting. The authors assert that their technique achieves higher uplink throughput than traditional benchmark techniques (specifically BCQ) without the computational overhead of training custom deep reinforcement learning models. The work positions the WiSER platform as a viable tool for autonomous, intent-based network management in dynamic wireless environments.