LLM-RIMSA: Transformer Control for 6G Metasurfaces

• LLM-RIMSA is a wireless control framework that employs a pretrained transformer to convert uplink pilot observations into analog beamforming and digital precoding matrices for reconfigurable intelligent metasurfaces.
• It integrates Conv1D, BiLSTM, and spatio-temporal attention with a frozen GPT-2 model to directly map pilots to control outputs, eliminating the need for explicit channel estimation during inference.
• Experimental results show that LLM-RIMSA outperforms DRL and GNN baselines by reducing training parameters and inference time while achieving state-of-the-art sum rate performance in dense 6G scenarios.

LLM-RIMSA denotes a wireless control framework in which a pretrained LLM backbone is adapted to control a reconfigurable intelligent metasurface antenna system (RIMSA) for multi-user 6G radio environments. In the formulation introduced in "LLM-RIMSA: LLMs driven Reconfigurable Intelligent Metasurface Antenna Systems" (Huang et al., 18 Aug 2025), the central task is to map uplink pilot observations directly to a RIMSA analog beamforming matrix and a digital precoding matrix, thereby bypassing explicit sub-channel estimation at inference time. The framework is motivated by the claim that conventional optimization, deep reinforcement learning, and graph neural networks become increasingly inadequate when the radio front end evolves from a phase-only reflective RIS into a large-scale radiative metasurface antenna with independently controllable elements (Huang et al., 18 Aug 2025).

1. Conceptual position and lineage

LLM-RIMSA is situated within a broader shift from passive RIS-assisted propagation control toward metasurface-native transceiver architectures. In earlier RIMSA work, the metasurface is treated as the antenna aperture itself, rather than as an external reflector, and the key architectural move is the use of parallel coaxial feeding so that all metamaterial elements are excited simultaneously (Wei et al., 23 Jun 2025). Related work also studies RIMSA as a programmable sensing front end, including distributed anti-jamming sensing with frame-wise beam control and maximum ratio combining across multiple RIMSA receivers (Wang et al., 7 Aug 2025). LLM-RIMSA inherits this hardware trajectory but replaces explicit optimization over the metasurface state with a pretrained transformer-based controller (Huang et al., 18 Aug 2025).

The problem setting is specific to dense 6G scenarios requiring ultra-massive connectivity, high spectral efficiency, sub-millisecond latency, and 3D coverage. The paper argues that reflective RIS suffers from double-path fading, active RIS raises power and hardware costs, and serial-fed radiative variants such as DMA and RHS inherit frequency selectivity or reduced control flexibility. RIMSA is proposed as an alternative because it uses parallel coaxial feeding and compact 2D metasurface integration, while LLM-RIMSA is the corresponding control stack for that hardware (Huang et al., 18 Aug 2025).

A common misconception is that LLM-RIMSA is a prompt-engineered language agent for wireless networking. The implemented system is instead a GPT-2-based neural controller embedded inside a signal-processing pipeline. The paper explicitly provides no natural-language prompt templates, no retrieval module, no tool use, and no text generation step in the conventional agentic sense; its "LLM" is a pretrained transformer backbone adapted to pilot-derived wireless features rather than a chat-style reasoning model (Huang et al., 18 Aug 2025).

2. RIMSA hardware and communication model

In LLM-RIMSA, RIMSA is introduced as a radiative metasurface antenna architecture that integrates antenna radiation and metasurface reconfiguration into a compact structure. The architecture includes phase-shifting circuits, metamaterial elements, a power distribution network, and feed ports. Each metamaterial element contains a varactor, transmission line, and short stub, and the phase response is controlled by changing the varactor capacitance through a bias voltage. RF power is distributed through a microstrip-line power divider network, and all elements are excited simultaneously through the parallel feed network (Huang et al., 18 Aug 2025).

A single RIMSA contains

$N_E = N_{E_x} \times N_{E_y}$

subwavelength metamaterial elements. A RIMSA array contains

$N_R = N_{R_x} \times N_{R_y}$

RIMSAs, each connected to an RF chain, and the total number of elements is

$$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$

The system model is a multi-user MISO downlink/uplink reciprocity setup with a base station equipped with a RIMSA array and $K$ single-antenna users (Huang et al., 18 Aug 2025).

The uplink pilot receive model is

$$\mathbf{y}(\ell)=\sum_{k=1}^{K}\mathbf{V}^{H}\mathbf{h}_{k}x_{k}(\ell)+\mathbf{V}^{H}\mathbf{n}(\ell),$$

with pilot observation matrix

$$\mathbf{Y}=[\mathbf{y}(1),\mathbf{y}(2),\cdots,\mathbf{y}(L)].$$

The downlink received signal at user $k$ is

$$r_k = \mathbf{h}_k^H \mathbf{V}\mathbf{w}_k s_k + \sum_{i\neq k} \mathbf{h}_k^H \mathbf{V}\mathbf{w}_i s_i + n_k,$$

where analog metasurface control is represented by $\mathbf V$ and digital precoding by $\mathbf W=[\mathbf w_1,\dots,\mathbf w_K]$ (Huang et al., 18 Aug 2025).

The beamforming matrix is block diagonal, with per-element phase coefficients

$N_R = N_{R_x} \times N_{R_y}$0

and the paper imposes

$N_R = N_{R_x} \times N_{R_y}$1

The channel model is Rician, with geometry-based LoS and NLoS components. The SINR is defined as

$N_R = N_{R_x} \times N_{R_y}$2

and the sum rate is

$N_R = N_{R_x} \times N_{R_y}$3

Although the architectural motivation repeatedly invokes independent amplitude-phase control, the formal signal model and control parameterization mainly use phase-only coefficients, a distinction the paper itself leaves unresolved (Huang et al., 18 Aug 2025).

3. Neural architecture and pilot-to-control mapping

The core control objective is to learn a direct mapping

$N_R = N_{R_x} \times N_{R_y}$4

from received pilots to digital-plus-analog beamforming decisions. The pipeline begins by converting the complex pilot matrix $N_R = N_{R_x} \times N_{R_y}$5 into a real-valued tensor by separating real and imaginary parts and reshaping it to $N_R = N_{R_x} \times N_{R_y}$6. This tensor is then processed by a Conv1D layer, GELU, batch normalization, max pooling, and a BiLSTM to extract local temporal structure and denoise the pilots (Huang et al., 18 Aug 2025).

A spatio-temporal attention module follows. Temporal attention uses scaled dot-product attention, and spatial attention applies a learned projection over the BiLSTM output. The fused representation is projected to the GPT-2 hidden dimension and augmented with learnable positional encodings before entering the transformer backbone. The paper uses standard GPT-2 and reports main experiments with a 6-layer configuration (Huang et al., 18 Aug 2025).

The pretrained GPT-2 multi-head attention and feedforward layers are frozen. Training is restricted to position encodings, normalization-related components, adapters or extensions, and the output heads. This freezing strategy is one of the paper's main arguments for reducing trainable parameter count while retaining a large pretrained backbone (Huang et al., 18 Aug 2025).

Three output branches are produced from the shared latent representation. A phase optimization head generates the RIMSA control variables and enforces physical phase limits with Hardtanh:

$N_R = N_{R_x} \times N_{R_y}$7

A second head generates the digital precoding matrix $N_R = N_{R_x} \times N_{R_y}$8. A third head outputs an auxiliary channel estimate $N_R = N_{R_x} \times N_{R_y}$9, which is used only during training. This makes the system a multitask controller in which the LLM backbone supplies a shared latent space for analog control, digital precoding, and supervised channel-estimation regularization (Huang et al., 18 Aug 2025).

The paper describes this as "cross-modal reasoning," but the implemented modalities are internal wireless representations: raw complex pilot signals, temporal and spatial correlations, latent transformer features, physical control outputs, and an auxiliary channel estimate. This suggests that the term is being used more loosely than in multimodal language-model research (Huang et al., 18 Aug 2025).

4. Learning objective and inference regime

LLM-RIMSA is trained offline with a hybrid multitask loss:

$$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$0

The channel-estimation term is

$$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$1

the rate term is

$$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$2

and the precoding-alignment term is

$$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$3

where $$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$4 is the ZF precoding matrix (Huang et al., 18 Aug 2025).

This training strategy is structurally important. The paper is explicit that explicit channel estimation is used only during training; at inference time, the controller maps pilot signals directly to control variables. LLM-RIMSA therefore belongs to the class of amortized optimization methods: the high-dimensional beamforming problem is not solved online by alternating optimization or policy search, but is replaced by a single forward pass of a pretrained transformer-based network (Huang et al., 18 Aug 2025).

The method is compared against DRL/DDPG, GNN/GAT, and random or no-optimization baselines. The paper's critique is that DRL requires extensive online interaction and explores the very large action space inefficiently, while GNNs incur growing computational cost with larger user and element counts and rely more heavily on offline data. LLM-RIMSA is presented as a lower-overhead alternative because it needs no explicit CSI estimation at inference and no trial-and-error online exploration (Huang et al., 18 Aug 2025).

5. Experimental configuration and reported performance

The simulated topology places the base station at $$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$5 m and three users uniformly in $$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$6 on the $$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$7-plane with user height $$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$8 m. The RIMSA hardware configuration uses 1024 total elements, 128 RF chains, a $$N_t = N_{E_x}N_{R_x} \times N_{E_y}N_{R_y} = N_x \times N_y.$$9 planar array, and a $K$0 RF-chain arrangement. Uplink pilot power is 15 dBm, downlink transmit power is 20 dBm, uplink noise power is $K$1 dBm, and downlink noise power is $K$2 dBm. The training set is reported as 10,000 samples, while the validation and test counts are inconsistent across sections, appearing once as 2,000 and 5,000 and elsewhere as 1,000 and 2,000 (Huang et al., 18 Aug 2025).

The main deployment comparison reported in Table III is as follows.

Method	Trainable / total parameters	Training time	Inference time
LLM	1.73M / 84.78M	2.6 h	20 ms
GNN	2.12M / 2.12M	4.8 h	50 ms
DRL	2.36M / 2.36M	7.3 h	120 ms

Despite the largest total model size, LLM-RIMSA has the fewest trainable parameters among the compared learned baselines and the lowest reported training and inference times (Huang et al., 18 Aug 2025).

The GPT-2 depth ablation in Table IV reports: 2 layers yielding 7.01 bps/Hz with 13 ms inference, 4 layers yielding 9.62 bps/Hz with 17 ms inference, 6 layers yielding 9.79 bps/Hz with 20 ms inference, and 8 layers yielding 9.73 bps/Hz with 28 ms inference. Within this tested range, 6 layers is the best configuration (Huang et al., 18 Aug 2025).

Across pilot-length, transmit-power, fairness, and user-count experiments, the paper reports that LLM-RIMSA consistently outperforms the compared DRL and GNN baselines in sum rate and max-min rate. The abstract summarizes these simulations as demonstrating "state-of-the-art performance" and reduced training overhead, but the empirical scope is limited to the selected DRL, GNN, and random baselines rather than the full optimization literature (Huang et al., 18 Aug 2025).

6. Interpretation, misconceptions, and limitations

LLM-RIMSA's main architectural lesson is that a pretrained transformer can be repurposed as a wireless sequence model over pilot-derived spatio-temporal representations. In the paper's interpretation, the gains arise from five interacting components: pilot-to-control mapping rather than pilot-to-CSI-to-control, hierarchical feature extraction through Conv1D and BiLSTM, decoupled spatio-temporal attention, a pretrained GPT-2 contextual prior, and multitask regularization through the auxiliary channel-estimation head (Huang et al., 18 Aug 2025).

Several limitations are explicit. First, the "LLM" is unconventional: the system does not use textual prompts, in-context exemplars, or language outputs. Second, the architecture is motivated by independent amplitude and phase control, but the formal model mainly instantiates phase-only control. Third, the evaluation is simulation-based under a specific Rician geometry and does not deeply test distribution shift, hardware mismatch, quantized control, mutual coupling, nonlinear varactor effects, synchronization errors, pilot corruption, or adversarial robustness. Fourth, there is no formal approximation guarantee, convergence proof for the end-to-end controller, or analytical account of why GPT-style pretraining should be uniquely advantageous for this task (Huang et al., 18 Aug 2025).

A further limitation is interpretive rather than empirical. The paper's language about "few-shot learning" and "cross-modal reasoning" is stronger than what is concretely implemented. The deployed system is better understood as a frozen-backbone transformer controller with wireless-specific preprocessing and multitask decoding than as a general reasoning LLM (Huang et al., 18 Aug 2025).

Within the RIMSA literature, LLM-RIMSA is most usefully read as a control-layer counterpart to hardware- and optimization-centric RIMSA studies. Earlier work formulates joint digital and metasurface optimization for MU-MISO and MU-MIMO transceivers using alternating optimization, fractional programming, WMMSE, and product manifold optimization (Wei et al., 23 Jun 2025), while related sensing work shows that distributed RIMSA receivers can support anti-jamming occupancy detection by learning frame-wise beam patterns under a combined cross-entropy and SINR objective (Wang et al., 7 Aug 2025). LLM-RIMSA departs from both by amortizing the control problem into a forward model from pilots to metasurface states and digital precoders (Huang et al., 18 Aug 2025).

In that sense, LLM-RIMSA is less a language-centric system than a proposal for pretrained-transformer-native wireless control. Its importance lies in the claim that, for large radiative metasurface arrays, the control bottleneck may be attacked more effectively by frozen-backbone sequence modeling than by explicit CSI-driven optimization, DRL exploration, or fully trainable graph architectures (Huang et al., 18 Aug 2025).