Meta-Transfer Learning for mmWave Beam Alignment
arXiv:2607.00860v1 [eess.SP] 1 Jul 2026
A. Nuri Cevik and Sinem Coleri
Abstract—Millimeter-wave (mmWave) beam alignment plays a critical role in next-generation wireless systems, yet its efficient implementation remains challenging. Meta-learning and transfer learning have been explored to enable deep learningbased beam prediction models to rapidly adapt to unseen environments; however, existing meta-learning approaches adapt the entire network and are trained from random initialization, leading to a large number of updated parameters and a high meta-training cost, while transfer learning approaches restrict adaptation to part of the network but do not exploit episodic meta-learning, which explicitly trains the model over multiple tasks, to optimize the adaptation process itself. To overcome these limitations, we propose MTL-BA, a meta-transfer learning framework for beam alignment in millimeter-wave multipleinput single-output (MISO) systems that freezes a pre-trained convolutional backbone and meta-learns only lightweight Scaleand-Shift (SS) adapters together with a classifier head. Warmstarting from the pre-trained model and restricting adaptation to the SS adapters and classifier head reduce both the adaptation cost and the meta-training budget without sacrificing prediction performance. Simulation results on the DeepMIMO raytracing dataset show that MTL-BA matches the accuracy and spectral efficiency of full fine-tuning across various SNR levels despite updating approximately 17× fewer parameters than both full fine-tuning and Model-Agnostic Meta-Learning (MAML), outperforms last-layer fine-tuning while updating a comparable number of parameters, and approaches MAML’s performance while requiring 60% fewer meta-training epochs. Index Terms—Beam alignment, meta-learning, transfer learning, millimeter-wave, MISO, adaptive beam prediction, MAML.
I. I NTRODUCTION Millimeter-wave (mmWave) communication has emerged as a cornerstone technology for fifth-generation (5G) and beyond wireless systems, offering multi-gigahertz bandwidth that enables unprecedented data rates. However, the high path loss and severe susceptibility to blockage inherent to mmWave frequencies necessitate large antenna arrays at the base station (BS), which in turn demands precise beam alignment between the BS and each user equipment (UE). Traditionally, beam alignment is accomplished via exhaustive search over a predefined narrow-beam codebook, but this incurs a sweeping overhead that scales linearly with the codebook size and quickly becomes prohibitive in large-array deployments. To mitigate this overhead, deep learning (DL)-based beam prediction methods have been extensively investigated. A prominent approach replaces exhaustive sweeping with a small set of wide probing beams drawn from a DFT codebook, whose received power measurements are fed into a neural network that directly predicts the optimal narrow beam from an oversampled DFT (O-DFT) codebook, significantly reducing
beam training overhead [1], [2]. While such methods achieve strong prediction accuracy in controlled settings, they rely on the assumption that training and deployment environments share the same data distribution. In practice, however, channel characteristics vary across locations, carrier frequencies, and time, causing a distribution shift that significantly degrades a statically trained model’s performance. Addressing this distribution shift through fast adaptation to new environments using only a handful of labeled samples from the target domain has therefore become a central research challenge. One approach is to train the entire network through episodic meta-learning so that the model can be rapidly adapted to new tasks. Yuan et al. [3] were among the first to explore MAML-based meta-learning for beamforming adaptation, training a convolutional neural network (CNN) from random initialization across multiple channel distributions. Xu et al. [4] adopted a channel state information (CSI)-free formulation by casting beam prediction as a classification problem over probing beam measurements, and introduced a scenario-adaptive task-splitting strategy to improve MAML-based convergence. While these approaches benefit from episodic optimization, they update all network weights during both meta-training and adaptation, resulting in high parameter cost, and training from random initialization requires many meta-training epochs to converge. An alternative is to reduce the adaptation cost by restricting the set of updated parameters. Yuan et al. [3] proposed a transfer learning approach that pre-trains a CNN and fine-tunes only the last fully-connected layer. To make this approach more effective, hybrid methods that incorporate meta-learning have also been developed. Zhang et al. [5] trained a fixed embedding model and fitted a support vector regressor on the target domain. Yang et al. [6] designed a meta-learning framework for dual-band beam prediction, where the encoder is optimized during meta-training and only the decoder is updated during adaptation. More recently, Mu et al. [7] proposed a simplified meta-training mechanism for beam alignment, where the network is meta-trained but only the classifier head is updated during deployment. Although these hybrid methods achieve lower adaptation cost, they do not modulate the intermediate feature representations to the target domain. Furthermore, those that employ meta-learning are still trained from random initialization, requiring many epochs to converge. In this paper, we propose MTL-BA, a meta-transfer learning framework for adaptive beam alignment in mmWave systems that unifies transfer learning and meta-learning by integrating a frozen pre-trained CNN backbone with lightweight Scale-
and-Shift (SS) adapters. (i) MTL-BA freezes the pre-trained backbone and adapts only the lightweight SS adapters and classifier head, reducing the number of updated parameters. (ii) Warm-starting from a pre-trained model accelerates meta-training convergence compared to random initialization. (iii) SS adapters and classifier head are jointly meta-learned through episodic training, combining the reuse of transfer learning features with the optimization of meta-learning adaptation. We evaluate MTL-BA on DeepMIMO ray-tracing [8] scenarios with a non-trivial domain shift between training and deployment environments. The remainder of this paper is organized as follows. Section II presents the system model. Section III formulates the beam prediction problem and details the proposed MTL-BA framework. Section IV provides simulation results. Section V concludes the paper.
where PT denotes the transmit power, nu ∼ CN (0, σn2 ) is complex Gaussian noise, and (·)H denotes the conjugate transpose. Under the unit-power symbol assumption, the resulting received signal-to-noise ratio (SNR) at UEu for the i-th beam is given by 2 PT hH u vi . (5) SNRu,i = σn2
II. S YSTEM M ODEL
We propose MTL-BA, a meta-transfer learning framework for adaptive beam alignment in mmWave systems that integrates transfer learning and meta-learning to enable efficient and rapid adaptation to new deployment environments. A CNN is trained to predict the optimal narrow beam from a small set of wide probing beam measurements, replacing the exhaustive codebook sweep. To address performance degradation under distribution shift, we further introduce a meta-transfer learning strategy that freezes the pre-trained backbone and meta-learns lightweight Scale-and-Shift adapters for rapid adaptation.
We consider a mmWave multiple-input single-output (MISO) system where the BS is equipped with a uniform linear array (ULA) with NBS antennas. The BS communicates with NU single-antenna UEs, and beam alignment is performed at the BS. We concentrate on the scenario of multi-user beamforming, where the BS communicates with each UE using only a single stream. The channel from the BS to UEu can be expressed based on geometric channel modeling as hu =
L X
αu,l b(ϕu,l )
(1)
l=1
where L denotes the number of paths, αu,l represents the complex path gain for the l-th path, ϕu,l is the angle of departure for the l-th path, and b(ϕu,l ) represents the array response vector, which is given by T 1 b(ϕu,l ) = √ 1 ejψ · · · ej(NBS −1)ψ , (2) NBS where ψ = 2π λ d sin(ϕu,l ), λ denotes the signal wavelength, and d = λ2 is the antenna spacing. We consider an analog beamformer in the radio frequency (RF) domain. Each antenna element, indexed by m ∈ {1, . . . , NBS }, is connected to an analog phase shifter capable of applying a phase shift of ζm . The beamformer satisfies both the power constraint and the constant modulus constraint on each antenna element, and can be expressed as T 1 jζ1 f=√ (3) e ejζ2 · · · ejζNBS . NBS During the beam sweeping phase, the BS transmits a complex symbol s to the UE using a beamformer vi selected from a predefined codebook V ∈ CNBS ×NV , which contains NV candidate beams. Specifically, vi = Vei , where ei is the standard basis vector whose i-th element is 1. Without loss of generality, the transmit symbol satisfies |s|2 = 1. The received signal at user u corresponding to the i-th beam is given by p ru,i = PT hH (4) u vi s + n u ,
Therefore, for a given BS–UE pair, the optimal beam index is determined by selecting the beam that maximizes the received SNR, i.e., i∗u = arg
max
i∈{1,...,NV }
SNRu,i = arg
max
i∈{1,...,NV }
2
hH u vi . (6)
III. D EEP N EURAL N ETWORK F RAMEWORK FOR B EAM A LIGNMENT
A. DL-based Beam Prediction Formulation To reduce the beam training overhead associated with exhaustive beam sweeping, we employ a small set of probing beams, denoted by Mp , instead of sweeping the entire narrowbeam codebook. During the probing phase, the BS sequentially transmits pilot symbols over these Mp probing beams, each in a separate time slot. For each user u, the received signal corresponding to the p-th probing beam is given by p p = 1, . . . , Mp , (7) ru,p = PT hH u wp sp + nu,p , where wp denotes the p-th probing beam, sp is the transmitted pilot symbol, and nu,p is the corresponding noise term. Based on these probing transmissions, each UE measures the received signal strength over the Mp probing beams and forms the input feature vector as T xu = |ru,1 |2 , |ru,2 |2 , · · · , |ru,Mp |2 . (8) It is assumed that the probing, measurement, and feedback operations are completed within the channel coherence time, such that the channel remains approximately constant during this process. We formulate the beam prediction problem as a supervised learning task that maps the probing measurements to the optimal narrow beam index. Specifically, given the probing feature vector xu ∈ RMp , the objective is to predict the optimal narrow beam index i∗u ∈ {1, . . . , NV }. To this end, we define a parametric function f (·; θ) : RMp → RNV , where θ denotes the learnable parameters of
the model. The function f outputs a probability distribution over the candidate narrow beams:
Input RMp
q̂u = f (xu ; θ),
Convolution
(9)
where q̂u ∈ RNV and its i-th element q̂u,i represents the predicted probability that beam i is optimal. The groundtruth label is expressed as a one-hot probability distribution ∗ ∗ q∗u ∈ RNV , where qu,i = 1 if i = i∗u and qu,i = 0 otherwise. The learning problem can then be formulated as the following empirical risk minimization: min Eu [L(xu , q∗u ; θ)] , θ
L(xu , q∗u ; θ) = −
∗ qu,i log q̂u,i .
Scale-Shift γ, β
Convolution ReLU
(10)
where the loss function is defined as the cross-entropy between the predicted and true distributions: NV X
ReLU
(11)
Scale-Shift γ, β flatten
Fully Connected
i=1
ReLU
Once trained, the predicted optimal beam index is obtained as îu = arg
max
i∈{1,...,NV }
[f (xu ; θ)]i .
(12)
Fully Connected (head)
where [·]i denotes the i-th element of the output vector. B. Meta-Transfer Learning for Beam Alignment
Output {0, 1}NV
Although the DL-based beam prediction framework described above significantly reduces probing overhead, its performance degrades when the deployment environment differs from the training distribution, as channel conditions vary across locations, carrier frequencies, and time. To overcome this limitation and enable fast adaptation of the beam predictor to unseen environments using only a small number of labeled samples, we adopt a meta-transfer learning (MTL) strategy inspired by Sun et al. [9], originally proposed for few-shot image classification. The key idea is to freeze a pre-trained backbone and meta-learn a set of lightweight Scale-and-Shift (SS) parameters Φ = {ϕγ , ϕβ } that modulate the frozen feature representations, together with a task-adaptive classifier head θ. Both the backbone parameters Θ and the classifier head θ are initialized from a pre-trained model, while only the SS parameters are newly introduced and meta-learned. The model is then trained through episodic training across multiple base-station (BS) environments. 1) Model Architecture with Scale-Shift Adapters: Let f (· ; Θ) denote the pre-trained backbone (convolutional feature extractor) with frozen weights Θ. At each intermediate feature tensor z, the SS operation applies a channel-wise affine transformation SS(z ; ϕγ , ϕβ ) = ϕγ ⊙ z + ϕβ ,
Scale-Shift γ, β
(13)
where ⊙ denotes element-wise multiplication and ϕγ ← 1, ϕβ ← 0 at initialization. Because Θ is never modified, the SS mechanism preserves the pre-trained feature representations while requiring far fewer learnable parameters than full finetuning. The layer-wise Scale-Shift architecture is depicted in Fig. 1.
Frozen Layers
Adaptable Layers
Fig. 1. CNN architecture of the proposed MTL-BA framework. The frozen backbone f (·; Θ) is fixed, while the Scale-Shift adapters (ϕγ , ϕβ ) and the classifier head are the only updated parameters.
2) Episodic Training: Let B = {b1 , . . . , bK } be the set of K source BS environments used for meta-training. Each epoch consists of Tep episodes. At each training episode a task Tk is sampled from environment bk ∈ B, and its data s are partitioned into a support set Sk = {(xu , i∗u )}N u=1 and ∗ Nq a query set Qk = {(xu , iu )}u=1 . Following the sequential update protocol of [9], tasks within an epoch are processed one at a time: the inner-loop update is performed on Sk , and the outer-loop update is applied immediately using only the query loss of that single task, rather than averaging query losses over a meta-batch as in MAML [10]. In both loops, the loss is computed as the cross-entropy between the predicted and true beam distributions. Let q̂u = softmax h(xu ; Θ, Φ, θ) , where h denotes the composite mapping through the frozen backbone, SS adapters, and classifier head, and let q̂u,i∗u denote the predicted probability assigned to the true optimal beam i∗u . The loss over a data set D is then defined as |D|
1 X LD (θ, Φ; Θ) = − log q̂u,i∗u , |D| u=1
(14)
which yields the support loss LSk and the query loss LQk
when evaluated on Sk and Qk , respectively. 3) Inner-Loop (Support) Update: Given the current classifier head θ, the inner-loop update adapts θ to task Tk via Gin gradient steps on the support set, while Θ and Φ remain fixed: (0)
θk = θ, (i) (i−1) θk = θk − β ∇θ(i−1) LSk k
(15) (i−1) θk , Φ; Θ ,
(16)
where i = 1, . . . , Gin and β is the inner learning rate. The (G ) classifier state after the Gin inner steps is denoted by θk in . 4) Outer-Loop (Query) Update: After the inner-loop adaptation, the SS parameters Φ and the classifier head θ are jointly updated on the query set of the same task Tk : Φ Φ (G ) ← − α ∇{Φ,θ} LQk θk in , Φ; Θ , (17) θ θ where α is the outer learning rate. This single-task outer update is applied sequentially for every task sampled in the current (G ) epoch. Crucially, θk in carries over to the next episode as the new initial classifier state, so that the classifier accumulates knowledge across environments without re-initialization. This sequential, carry-over update rule contrasts with MAML [10], where the classifier is re-initialized to the meta-parameter at every task and the outer update uses the averaged query loss over a meta-batch. 5) Target Environment Adaptation (MTL-BA): After metatraining, adaptation to a new target environment proceeds by freezing Θ and fine-tuning only Φ and θ on a small labeled ad adaptation set Dad = {(xu , i∗u )}N u=1 collected from the target environment: (j+1) (j) Φ Φ ← − η ∇{Φ,θ} LDad Φ(j) , θ (j) ; Θ , θ θ (18) j = 1, . . . , Gad , where η is the adaptation learning rate and Gad is the number of adaptation gradient steps. Because Θ is frozen, the number of parameters to optimize is far smaller than in standard finetuning, making (18) efficient even with |Dad | as small as a single sample. The complete meta-training procedure is summarized in Algorithm 1. Comparison of MTL-BA and MAML: Both algorithms follow an episodic meta-learning framework and adapt with a small number of target samples, but they differ in several important respects: (i) MTL-BA freezes Θ and adapts only the lightweight SS parameters and head, reducing adaptation cost; MAML updates all weights, giving greater flexibility but requiring second-order gradient computation. (ii) MTL-BA is warm-started from a pre-trained backbone, yielding faster meta-convergence; MAML is trained from random initialization. (iii) MTL-BA performs sequential, per-task outer updates and carries the classifier head across tasks, enabling progressive knowledge accumulation; in contrast, MAML performs a batch-averaged outer update and adapts each task starting from a shared initialization.
Algorithm 1 MTL-BA Meta-Training for Beam Alignment Input: Pre-trained frozen backbone Θ; pre-trained classifier head θ; source environments B; inner/outer LR β, α; inner steps Gin ; meta-epochs E; episodes per epoch Tep . Output: Meta-trained SS parameters Φ∗ and classifier θ ∗ . Initialize ϕγ ← 1, ϕβ ← 0. for epoch e = 1, . . . , E do for episode t = 1, . . . , Tep do Sample task Tk from B; draw Sk , Qk . (0) Set θk ← θ. for i = 1, . . . , Gin do Compute LSk via (14). (i) Update θk via (16). end for Update Φ, θ via (17) using only LQk . (G ) θ ← θk in {carry-over: head persists across tasks} end for end for return Φ∗ ← Φ, θ ∗ ← θ.
IV. S IMULATION R ESULTS In this section, we present the simulation setup, dataset generation, training configuration of the proposed adaptive beam prediction framework, and a comprehensive performance evaluation of the proposed MTL-BA algorithm against different baselines. A. Simulation Setup To emulate realistic mmWave propagation, we adopt the DeepMIMO ray-tracing dataset [8], which provides sitespecific channel realizations across multiple BS environments. BS 3 through BS 13 of the O1 28 scenario serve as the source environments B = {b3 , . . . , b13 } for meta-training and validation, while the target environment btgt is drawn from the I3 60 indoor scenario, ensuring a non-trivial domain shift in both carrier frequency and propagation conditions between metatraining and deployment. The channel-generation parameters for the two scenarios are summarized in Table I. Thermal noise is added to the channel vectors during training, with a persubcarrier noise power of −119.1 dBm, computed assuming a noise floor of −174 dBm/Hz, a noise figure of 5 dB, and a processing gain of 10 dB. The narrow-beam codebook V is constructed as an O-DFT codebook of size NV = 128. The probing DFT codebook contains Mp = 32 wide sensing beams. The probing measurements xu and the corresponding optimal beam index i∗u are computed using (8) and (6), respectively, to form the labeled samples used for training, meta-training, and adaptation. All simulations are performed on a system equipped with NVIDIA A100 GPU. B. Deep Learning Model Architecture and Training Configuration The CNN backbone f (·; Θ) shared by all algorithms consists of two convolutional layers and a fully-connected layer,
TABLE I C HANNEL GENERATION PARAMETERS . Parameter
Training
Scenario Active BS Carrier frequency BS transmit power System Bandwidth Antennas (x, y, z) Antenna spacing OFDM subcarriers OFDM sampling factor OFDM limit Number of multipaths L
Deployment
O1 28 I3 60 BS 3 – BS 13 BS 2 28 GHz 60 GHz 30 dBm 0.5 GHz (1, 32, 1) λ/2 512 1 1 5
TABLE II N EURAL NETWORK ARCHITECTURE AND TRAINING HYPER - PARAMETERS . Parameter
MTL-BA
MAML
Conv 1 Conv 2 FC 1 FC 2 (head) Activation Loss function Episodes per epoch Tep Optimizer Pre-train epochs Inner LR β Outer LR α Adaptation LR η Inner steps Gin Meta-epochs E Support / query Ns /Nq
1 → 64 ch., kernel 3, pad 1 64 → 64 ch., kernel 3, pad 1 2048 → 256 256 → NV ReLU (after each layer) Cross-entropy (label smooth. ϵ = 0.05) 30 AdamW Adam 40 – 0.1 10−2 10−3 10−3 −3 10 10−4 5 5 200 500 20 / 50 20 / 50
followed by a classifier head θ; the detailed architecture is summarized in Table II. For MTL-BA, the SS adapters Φ are inserted after each convolutional layer and the first fullyconnected layer according to (13). The pre-trained backbone used to warm-start MTL-BA is obtained by training the CNN on the union of all source environments B for 40 epochs. C. Benchmark Methods and Evaluation Metrics To benchmark the performance of the proposed MTL-BA algorithm, we consider the following baselines: FT-LAST: A transfer learning baseline [3] in which the backbone Θ is pre-trained on B and only the last (fullyconnected) layer θ is fine-tuned on Dad . • FT-ALL: A transfer learning baseline that pre-trains on B and fine-tunes all parameters (Θ, θ) on Dad . • MAML: The model-agnostic meta-learning baseline [3], [10] trained from random initialization. •
For each operating point, we report the mean and standard deviation across 5 random seeds to ensure statistically meaningful comparisons. The number of adaptation samples |Dad | is varied from 100 to 1000, and the SNR is varied from −5 dB
Fig. 2. Top-1 and Top-3 accuracy versus SNR on target environment for (a) 100 and (b) 1000 adaptation samples.
to 35 dB. All baseline methods share the same CNN backbone architecture (Table II) to ensure a fair comparison. D. Performance Evaluation We evaluate the four algorithms on the target environment under varying numbers of adaptation samples and SNR levels. 1) Accuracy: Figure 2 reports the Top-1 and Top-3 beamprediction accuracy as a function of SNR for two adaptation sample sizes. MAML achieves the highest accuracy in most operating conditions, particularly at moderate SNR (15–25 dB), where the learned initialization provides a strong inductive bias for fast adaptation. At high SNR (35 dB), all methods converge to comparable performance with sufficient adaptation data, while at very low SNR (≤ 5 dB), the discriminative information in the probing measurements degrades significantly and the gap between methods narrows. MTL-BA achieves accuracy comparable to FT-ALL across all SNR levels, despite updating only the lightweight SS adapters and the classifier head, a substantially smaller parameter set than FT-ALL and MAML, which fine-tune the entire backbone. This confirms that the meta-learned SS parameters
updates only 33,664 parameters, comprising the SS adapter vectors and the classifier head, which is nearly identical to FTLAST (32,896) and approximately 17× fewer than FT-ALL and MAML, both of which update the full 570,048-parameter network. This parameter efficiency is a direct consequence of freezing the backbone Θ during adaptation. Furthermore, MTL-BA requires only 200 meta-training epochs, compared to 500 epochs for MAML, reducing the meta-training cost as well. Combined with the competitive accuracy and spectral efficiency reported above, these results demonstrate that MTL-BA offers a favorable accuracy–efficiency trade-off for resource-constrained deployment and training scenarios. V. C ONCLUSION
Fig. 3. Spectral Efficiency on target environment versus the number of adaptation samples for SNR ∈ {35, 25, 15} dB. TABLE III U PDATED PARAMETERS DURING ADAPTATION . MTL-BA BREAKDOWN : SS1 + SS2 + SS3 + HEAD . FT-ALL/MAML BREAKDOWN : C ONV 1 + C ONV 2 + FC 1 + HEAD . Method
No. of Parameters
Parameter Breakdown
FT-LAST FT-ALL MTL-BA MAML
32,896 570,048 33,664 570,048
256 × 128 + 128 256 + 12,352 + 524,544 + 32,896 128 + 128 + 512 + 32,896 256 + 12,352 + 524,544 + 32,896
provide an efficient adaptation mechanism, matching the performance of full fine-tuning at a fraction of the update cost. FT-LAST consistently lags behind, as adapting only the final layer is insufficient to overcome the distribution shift between the source and target environments. 2) Spectral Efficiency: Fig. 3 presents the spectral efficiency on the target environment as a function of the number of adaptation samples. At high SNR (35 dB), all methods approach the genie-aided optimal upper bound closely, with differences smaller than 0.4 bps/Hz, indicating that even imperfect beam prediction incurs negligible rate loss when the SNR is high. As the SNR decreases, the gap to the optimal widens for all methods, yet the relative ordering is preserved: MAML remains closest to the optimal, followed by MTL-BA and FT-ALL. Notably, despite the pronounced Top-1 accuracy gap between MAML and the remaining methods, the corresponding spectral efficiency gap is substantially smaller across all SNR levels. This is because the rate loss is governed by the beam gain of the selected beam, and near-optimal beams in the codebook yield similar array gains. MTL-BA achieves spectral efficiency on par with FT-ALL, further confirming that the SS adaptation mechanism effectively compensates for the distribution shift while updating far fewer parameters. 3) Complexity: Table III summarizes the number of parameters updated during adaptation for each method. MTL-BA
In this paper, we proposed MTL-BA, a framework that unifies transfer learning and meta-learning for adaptive beam alignment in mmWave systems. The key idea is to reuse a frozen pre-trained backbone for feature extraction while meta-learning only lightweight Scale-and-Shift adapters and a classifier head for fast adaptation to unseen environments. This combination reduces the number of adapted parameters by approximately 17× compared to MAML and full finetuning. Experimental results on DeepMIMO ray-tracing scenarios demonstrate that MTL-BA matches the accuracy and spectral efficiency of full fine-tuning across all SNR levels while approaching MAML’s accuracy at 60% lower metatraining cost. R EFERENCES [1] Y. Heng, J. Mo, and J. G. Andrews, “Learning site-specific probing beams for fast mmWave beam alignment,” IEEE Trans. Wireless Commun., vol. 21, no. 8, pp. 5785–5800, Aug. 2022. [2] N. Khan, A. Abdallah, A. Celik, A. M. Eltawil, and S. Coleri, “Explainable and Robust Millimeter Wave Beam Alignment for AI-Native 6G Networks,” in Proc. IEEE Int. Conf. Commun. (ICC), Montreal, QC, Canada, Jun. 2025, pp. 753–758. [3] Y. Yuan, G. Zheng, K.-K. Wong, B. Ottersten, and Z.-Q. Luo, “Transfer Learning and Meta Learning-Based Fast Downlink Beamforming Adaptation,” IEEE Trans. Wireless Commun., vol. 20, no. 3, pp. 1742–1755, Mar. 2021. [4] Z. Xu, S. Wang, and Y.-J. A. Zhang, “SAMBA: Scenario-adaptive meta-learning for mmWave beam alignment,” in Proc. IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur, Malaysia, Dec. 2023, pp. 1–6. [5] J. Zhang, Y. Yuan, G. Zheng, I. Krikidis, and K.-K. Wong, “Embedding model-based fast meta learning for downlink beamforming adaptation,” IEEE Trans. Wireless Commun., vol. 21, no. 1, pp. 149–162, Jan. 2022. [6] R. Yang, Z. Zhang, X. Zhang, C. Li, Y. Huang, and L. Yang, “Metalearning for beam prediction in a dual-band communication system,” IEEE Trans. Commun., vol. 71, no. 1, pp. 145–157, Jan. 2023. [7] Y. Mu, Q. Cui, Q. Li, X. Lyu, and X. Tao, “MPBA: Meta-predictive beam alignment for mmWave systems with environmental variability,” IEEE Trans. Commun., vol. 73, no. 11, pp. 12131–12145, Nov. 2025. [8] DeepMIMO, “DeepMIMO Dataset Generation,” https://deepmimo.net/, accessed: May 15, 2026. [9] Q. Sun, Y. Liu, T.-S. Chua, and B. Schiele, “Meta-transfer learning for few-shot learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2019, pp. 403–412. [10] C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in Proc. Int. Conf. Mach. Learn., 2017, pp. 1126–1135.