BeamAgent: LLM-Aided MIMO Beamforming with Decoupled Intent Parsing and Alternating Optimization for Joint Site Selection and Precoding

Xiucheng Wang; Yue Zhang; Nan Cheng

cs.ITMar 19, 2026

BeamAgent: LLM-Aided MIMO Beamforming with Decoupled Intent Parsing and Alternating Optimization for Joint Site Selection and Precoding

Authors: Xiucheng Wang, Yue Zhang, Nan Cheng

Organizations: State Key Laboratory of ISN and School of Telecommunications Engineering, Xidian University, Xi’an 710071, China

Abstract

Integrating large language models (LLMs) into wireless communication optimization is a promising yet challenging direction. Existing approaches either use LLMs as black-box solvers or code generators, tightly coupling them with numerical computation. However, LLMs lack the precision required for physical-layer optimization, and the scarcity of wireless training data makes domain-specific fine-tuning impractical. We propose BeamAgent, an LLM-aided MIMO beamforming framework that explicitly decouples semantic intent parsing from numerical optimization. The LLM serves solely as a semantic translator that converts natural language descriptions into structured spatial constraints. A dedicated gradient-based optimizer then jointly solves the discrete base station site selection and continuous precoding design through an alternating optimization algorithm. A scene-aware prompt enables grounded spatial reasoning without fine-tuning, and a multi-round interaction mechanism with dual-layer intent classification ensures robust constraint verification. A penalty-based loss function enforces dark-zone power constraints while releasing optimization degrees of freedom for bright-zone gain maximization. Experiments on a ray-tracing-based urban MIMO scenario show that BeamAgent achieves a bright-zone power of 84.0,dB, outperforming exhaustive zero-forcing by 7.1 dB under the same dark-zone constraint. The end-to-end system reaches within 3.3 dB of the expert upper bound, with the full optimization completing in under 2 s on a laptop.

Figures & tables

Method	LLM Role	Optimization Layer	NL Interface	Fine-Tuning	Solver
ComAgent [ 14 ]	Code generator	Task-dependent	No	No	LLM-generated code
WirelessAgent [ 15 ]	Decision engine	Network (slicing)	Partial	No	LLM reasoning loop
PE-RTFV [ 17 ]	Black-box solver	Physical (constellation)	No	No	LLM output
NetLLM [ 19 ]	End-to-end model	Network (scheduling)	No	Yes	Fine-tuned LLM
TelecomGPT [ 24 ]	Domain-adapted LLM	Telecom tasks	No	Yes	Fine-tuned LLM
TelePlanNet [ 37 ]	Intent parser + RL	Site planning	Yes	Yes (RL)	RL policy

TABLE I : Comparison of BeamAgent with representative existing methods.

Fig. 1 : Overview of the BeamAgent framework. The system consists of three functional stages separated into two domains by the structured constraint interface $\mathcal{C}=(\mathcal{B},\mathcal{D},\mathcal{R}_{\mathrm{tx}})$ . In the semantic domain (left), an LLM parses natural language input into spatial constraints via a scene-aware prompt, followed by multi-round interaction with dual-layer intent classification for constraint verification. In the numerical domain (right), an alternating optimization algorithm jointly solves BS site selection (Phase A) and precoding design (Phase B), with a threshold-based penalty that releases optimization degrees of freedom once the dark-zone constraint is satisfied.

Method	$\bar{P}_{\mathcal{B}}$ (dB)	$\max P_{\mathcal{D}}$ (dB)	Scale (dB)	Contrast (dB)
Random	75.9	30.0 †	19.9	43.2
Exh+MRT	77.8	30.0 †	20.3	48.8
Exh+MMSE	78.6	30.0 †	1.0	57.7
Exh+ZF	76.9	14.5	—	70.1
SLNR	76.4	14.3	—	80.9
BeamAgent	84.0	30.0 †	0.9	54.2

TABLE II : Comparison of constraint-aware precoding methods in a 4-Site scenario with dark-zone threshold

T=30

dB. Methods marked with † are post-processed by scaling

\|\mathbf{W}\|

to enforce

\max_{s\in\mathcal{D}}P_{s,\text{dB}}\leq T

. The Scale column indicates the dB reduction applied during post-processing.

Fig. 2 : Bright-zone power

\bar{P}_{\mathcal{B}}

versus maximum dark-zone power

\max P_{\mathcal{D}}

for all methods after post-processing. The dashed line marks the dark-zone threshold

T=30

dB. Points below this line are feasible. BeamAgent achieves the highest

\bar{P}_{\mathcal{B}}

in the feasible region while operating near the constraint boundary.

Fig. 3 : Per-site received power (bars) and SVD bounds (triangles) at the optimized BS location. Black triangles (

\blacktriangledown

) denote

\sigma_{\max}^{2}

and gray triangles (

\vartriangle

) denote

\sigma_{\min}^{2}

. The dashed line marks

T=30

dB. Bright Sites approach their upper bounds; dark Sites are suppressed to the threshold.

Fig. 4 : Spatial distribution of received power (dB) in the 224 m $\times$ 222 m urban scenario after joint optimization. Green circles indicate bright Sites (7, 8); red squares indicate dark Sites (6, 12); the star marks the optimized BS location. The beam pattern steers energy toward the north while suppressing the south and center.

Method	Contrast (dB)	Std (dB)	Note
Expert+Optimizer	55.3	0.0	Expert constraints
BeamAgent	52.0	3.6	LLM parsing + optimizer
LLM Direct W	49.4	0.0	LLM outputs $\mathbf{W}$ directly
LLM Gen Code	80.9 ‡	0.0	LLM generates SLNR solver
‡ $\bar{P}_{\mathcal{B}}=76.4$ dB, lower than BeamAgent (84.0 dB); no site selection.

TABLE III : Comparison of four LLM integration modes using Claude Sonnet 4.6. Expert+Optimizer serves as the upper bound with manually specified constraints. BeamAgent uses LLM-parsed constraints with the same optimizer. LLM Direct W and LLM Gen Code represent alternative paradigms where the LLM directly participates in numerical computation.

Fig. 5 : Bright-dark contrast (dB) under four LLM integration modes using Claude Sonnet 4.6. BeamAgent (decoupled parsing + dedicated optimizer) achieves within 3.3 dB of the expert upper bound. LLM Direct W, which outputs the precoding vector without channel information, yields the lowest contrast among LLM-involved modes.

Fig. 6 : Effect of dark-zone threshold

T

on bright-zone power

\bar{P}_{\mathcal{B}}

(green squares) and actual dark-zone power

\max P_{\mathcal{D}}

(red triangles). The dashed line shows the threshold value. As

T

increases, the dark-zone power tracks the threshold and bright-zone power rises accordingly, confirming that relaxing the constraint releases optimization degrees of freedom.

Configuration	$\bar{P}_{\mathcal{B}}$ (dB)	$\max P_{\mathcal{D}}$ (dB)	Contrast (dB)
Full (Ours)	85.8	30.6	55.1
No Cosine Annealing	85.9	30.7	55.2
No Gradient Clipping	85.8	30.6	55.2
No Cosine + No Clip	85.9	30.7	55.3
$R=1$ (No Alternating)	77.6	30.9	46.7
Tradeoff (no penalty)	86.0	25.2	60.7

TABLE IV : Ablation study on optimization components in the 4-Site scenario (

T=30

dB). Each row removes one component from the full configuration. Disabling alternating optimization (

R=1

) causes the only significant performance degradation (

-8.2

dB in

\bar{P}_{\mathcal{B}}

Stage	Time (s)
Constraint preparation (tensor stacking)	1.350
Phase A: Tx scan (per round)	0.006
Phase B: 400 gradient steps	0.308
Full optimization (5 rounds)	1.668
Exhaustive MRT baseline (all Tx)	0.124

TABLE V : Wall-clock time for each optimization stage measured on an Apple M2 MacBook. LLM API latency is excluded. The full pipeline completes in under 2 s, with vectorized Tx scanning taking 6 ms per round.

Wireless Resource Allocation

Wireless resource allocation assigns spectrum, power, beams, and computing capacity to meet throughput, latency, fairness, and energy goals.

See all 174 papers

Mar 17, 2026

Structure-Aware Multimodal LLM Framework for Trustworthy Near-Field Beam Prediction

In near-field extremely large-scale multiple-input multiple-output (XL-MIMO) systems, spherical wavefront propagation expands the traditional beam codebook into the joint angular-distance domain, rendering conventional beam training prohibitively inefficient, especially in complex 3-dimensional (3D) low-altitude environments. Furthermore, since near-field beam variations are deeply coupled not only with user positions but also with the physical surroundings, precise beam alignment demands profound environmental understanding capabilities. To address this, we propose a large language model (LLM)-driven multimodal framework that fuses historical GPS data, RGB image, LiDAR data, and strategically designed task-specific textual prompts. By utilizing the powerful emergent reasoning and generalization capabilities of the LLM, our approach learns complex spatial dynamics to achieve superior environmental comprehension...

Mengyuan Li, Qianfan Lu, Jiachen Tian +5

School of Information Science and Engineering, Southeast University, Nanjing 210096, China · Institute of Communications Engineering, National Sun Yat-sen University, Kaohsiung 804, Taiwan

Wireless Communications

Ask PDF

May 7, 2026

AgenticPrecoding: LLM-Empowered Multi-Agent System for Precoding Optimization

Precoding is a key technique for interference management and performance improvement in multi-antenna wireless systems. However, existing precoding methods are typically developed for specific system models, objectives, and constraint sets, which limits their adaptability to the heterogeneous and evolving scenarios expected in future 6G networks. To address this limitation, we propose AgenticPrecoding, a universal multi-agent framework that automates end-to-end precoding derivation directly from user-level communication requirements. Specifically, AgenticPrecoding decomposes the derivation process into four coordinated stages: problem formulation, solver selection, prompt upsampling, and code generation, assigning each stage to a specialized agent tailored to its specific reasoning demands. We employ two LoRA-adapted reasoning agents to inject precoding-specific domain knowledge for problem formulation and solver selection, while two general-purpose Large Language Models (LLMs) handle prompt refinement and executable code generation. Furthermore, a feedback-driven refinement mechanism is incorporated to enhance code executability, constraint feasibility, and solution quality. Extensive experiments across 10 representative precoding scenarios demonstrate that AgenticPrecoding achieves superior cross-scenario adaptability compared to conventional optimization-based and LLM-based baselines.

Zijiu Yang, Zixiang Zhang, Shunpu Tang +2

College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China

Multi-Agent LLM Systems Large Language Model-Guided Optimization

Ask PDF

Feb 28, 2026

Wireless Power Control Based on Large Language Models

This paper investigates the power control problem in wireless networks by repurposing pre-trained large language models (LLMs) as relational reasoning backbones. In hyper-connected interference environments, traditional optimization methods face high computational cost, while standard message passing neural networks suffer from aggregation bottlenecks that can obscure critical high-interference structures. In response, we propose PC-LLM, a physics-informed framework that augments a pre-trained LLM with an interference-aware attention bias. The proposed bias tuning mechanism injects the physical channel gain matrix directly into the self-attention scores, enabling explicit fusion of wireless topology with pre-trained relational priors without retraining the backbone from scratch. Extensive experiments demonstrate that PC-LLM consistently outperforms both traditional optimization methods and state-of-the-art graph neural network baselines, while exhibiting exceptional zero-shot generalization to unseen environments. We further observe that topology-relevant relational reasoning is concentrated in shallow layers, whereas deeper layers encode task-irrelevant semantic noise. Motivated by this finding, we develop a lightweight adaptation strategy that reduces model depth by 50%, significantly lowering inference cost while preserving state-of-the-art spectral efficiency.

Jiacheng Wang, Yucheng Sheng, Le Liang +2

School of Information Science and Engineering, Southeast University, Nanjing 210096, China · Department of Electrical and Computer Engineering, University of California, Santa Cruz, CA 95064, USA

Wireless Resource Allocation Language Model-Based Control LLM Reasoning with Graphs

Ask PDF

Abstract

Wireless Resource Allocation

Explore similar work

Structure-Aware Multimodal LLM Framework for Trustworthy Near-Field Beam Prediction

AgenticPrecoding: LLM-Empowered Multi-Agent System for Precoding Optimization

Wireless Power Control Based on Large Language Models