---
title: 'NEFT: Efficient Near-Field Transformer for XL-MIMO'
url: https://www.emergentmind.com/topics/near-field-efficient-feedback-transformer-neft
type: topic
---

# NEFT: Efficient Near-Field Transformer for XL-MIMO

Searching arXiv for the cited NEFT paper and closely related work to ground the article.
Near-Field Efficient Feedback Transformer (NEFT) is a unified Transformer framework for efficient near-field channel state information (CSI) feedback in extremely large-scale multiple-input multiple-output (XL-MIMO) systems. It was introduced for single-cell FDD downlink CSI compression and reconstruction in the near-field regime, where the plane-wave assumption breaks down and spherical-wave propagation induces distance-dependent amplitude and nonlinear phase across antennas [2509.12748]. NEFT is built on a hierarchical Vision Transformer backbone operating on a reshaped CSI tensor, and is extended with lightweight variants tailored to deployment constraints: NEFT-Compact uses multi-level knowledge distillation, while NEFT-Hybrid and NEFT-Edge target encoder- and edge-constrained scenarios through attention-free encoding and knowledge distillation [2509.12748]. In reported simulations, NEFT achieves a 15--21 dB improvement in normalized mean-squared error (NMSE) over representative baselines, and its lightweight variants reduce total FLOPs by 25--36% with negligible to modest accuracy loss [2509.12748].

## 1. Near-field XL-MIMO setting and CSI feedback formulation

NEFT is formulated for XL-MIMO systems operating in the electromagnetic near-field. The boundary between near-field and far-field is characterized by the Rayleigh distance
$$
R_R = \frac{2D^2}{\lambda},
$$
where $D$ is the array aperture and $\lambda$ is the wavelength [2509.12748]. As $D$ grows, users at typical cell distances can fall within the Fresnel region, where wavefronts are spherical rather than planar and the conventional far-field angular sparsity model is no longer appropriate [2509.12748].

The underlying communication model is a single-cell FDD downlink with a base station (BS) using a large ULA with $N_1$ antennas and a user equipment (UE) with $N_2$ antennas. The received signal is
$$
y = \mathbf{H}\mathbf{v}x + n,
$$
where $x$ is the transmitted symbol, $\mathbf{v} \in \mathbb{C}^{N_1\times 1}$ is the precoder, $\mathbf{H} \in \mathbb{C}^{N_2\times N_1}$ is the channel, and $n$ is noise [2509.12748]. For the narrowband near-field line-of-sight (LoS) geometric model used in the paper, the channel coefficient between the $n_1$-th BS antenna and $n_2$-th UE antenna is expressed as
$$
\mathbf{H}(n_1,n_2) = \frac{1}{r_{n_1,n_2}} \exp\!\left(-j\frac{2\pi}{\lambda} r_{n_1,n_2}\right),
$$
with
$$
r_{n_2,n_1} = \sqrt{(r\cos\theta - d_2\sin\phi)^2 + (r\sin\theta + d_2\cos\phi - d_1)^2},
$$
where $r$ is the distance from the first UE antenna to the first BS antenna, $\theta$ is the angle-of-departure, $\phi$ is the relative angle between UE and BS, and $d_1=n_1d$, $d_2=n_2d$ with inter-element spacing $d$ [2509.12748].

In the reported experiments, the BS ULA has $N_1=1024$ and the UE is single-antenna with $N_2=1$ [2509.12748]. The complex channel vector is reshaped into a $2\times 32\times 32$ tensor, where the two channels correspond to the real and imaginary parts, normalized to $[0,1]$ [2509.12748]. This tensor is the encoder input $\mathbf{H}_{in}$. The encoder produces a $K$-dimensional feedback codeword
$$
\mathbf{s} = \mathcal{E}(\mathbf{H}_{in}),
$$
and the decoder reconstructs
$$
\hat{\mathbf{H}} = \mathcal{D}(\mathbf{s}),
$$
with $L=2N_1N_2$ and compression ratio
$$
\gamma = \frac{K}{L} = \frac{K}{2N_1N_2}.
$$
The paper studies $\gamma \in \{16,32,64\}$ [2509.12748].

A central design choice is that NEFT does not rely on frequency-domain wideband tensors or polar-/angle--distance-domain sparsity. Instead, it is designed to learn near-field spherical-wave phase and amplitude structure directly from the reshaped spatial tensor [2509.12748]. This distinguishes it from other near-field CSI feedback approaches that first impose a hand-crafted sparse representation. A plausible implication is that NEFT prioritizes end-to-end representation learning over explicit physical reparameterization.

## 2. Hierarchical Vision Transformer architecture

NEFT is built on a hierarchical Vision Transformer encoder-decoder. The input is the $2\times 32\times 32$ CSI tensor [2509.12748]. In stage-0 patch embedding, the input is partitioned into $8\times 8$ non-overlapping patches of size $2\times 4\times 4$, and each patch is linearly projected to embedding dimension $C_1$, producing a feature map of size $C_1 \times 8 \times 8$ with sequence length $L_1=64$ [2509.12748].

The encoder stage-1 contains a Transformer block with multi-head global self-attention (MSA) and MLP [2509.12748]. Patch merging is then performed through a $2\times 2$ convolution, aggregating four neighboring tokens into one, reducing spatial resolution from $8\times 8$ to $4\times 4$ and doubling channels to $2C_1 \times 4 \times 4$, corresponding to sequence length $L_2=16$ [2509.12748]. The bottleneck tensor is flattened into the feedback codeword; its dimension is
$$
N_d = 2C_1 \times 4 \times 4,
$$
which equals $K$ under direct floating-point feedback [2509.12748].

The decoder mirrors the hierarchy through transposed-convolution upsampling, denoted as patch division. It maps $2C_1 \times 4 \times 4 \rightarrow C_1 \times 8 \times 8$ with kernel $k=2$ and stride $s=2$, and then $C_1 \times 8 \times 8 \rightarrow 2 \times 32 \times 32$ with $k=4$, $s=4$ [2509.12748]. Transformer blocks with global attention are applied at the corresponding decoder stages to reconstruct long-range dependencies and fine-grained local patterns [2509.12748]. Positional information is handled through a relative position bias $\mathbf{B}$ in attention; global attention does not rely on window shifting [2509.12748].

The self-attention computation is given for feature $\mathbf{X} \in \mathbb{R}^{L\times C}$ as
$$
\mathbf{Q} = \mathbf{X}\mathbf{W}^Q,\quad \mathbf{K} = \mathbf{X}\mathbf{W}^K,\quad \mathbf{V} = \mathbf{X}\mathbf{W}^V,
$$
$$
\mathrm{head}_i = \mathrm{softmax}\!\left(\frac{\mathbf{Q}_i\mathbf{K}_i^T}{\sqrt{d_k}} + \mathbf{B}\right)\mathbf{V}_i,
$$
$$
\mathrm{MSA} = \mathrm{Concat}(\mathrm{head}_1,\dots,\mathrm{head}_h)\mathbf{W}^O.
$$
For global attention and windowed attention, the paper gives
$$
\Omega(\mathrm{MSA}) = 4hwC^2 + 2(hw)^2 C,
$$
$$
\Omega(\mathrm{W\!-\!MSA}) = 4hwC^2 + 2M^2 hw C,
$$
where $h\times w$ is feature map size, $C$ is the number of channels, and $M$ is window size [2509.12748]. By operating at smaller sequence lengths with larger patches, NEFT retains global attention at feasible cost. With typical parameters $(C=40, M=4)$, the aggregated operations are reported as approximately $860{,}160$ for NEFT versus $5{,}013{,}504$ for W-MSA, corresponding to about $17\%$ of the windowed-attention cost [2509.12748].

This architectural choice should not be confused with attention decompositions such as FMMformer, which split attention into near-field banded and far-field low-rank components for linear-complexity sequence modeling [2108.02347]. NEFT uses global MSA directly, but makes it efficient by aggressively reducing token count through patching and hierarchy rather than by replacing the attention operator itself [2509.12748].

## 3. Lightweight variants and deployment-oriented design

NEFT was presented as a family rather than a single model. The principal variants are NEFT-Compact, NEFT-Hybrid, and NEFT-Edge, each corresponding to a different deployment constraint [2509.12748].

NEFT-Compact is a student Transformer trained by multi-level knowledge distillation from the full NEFT teacher [2509.12748]. The student reduces token and channel dimensions as well as MLP projection size, while the distillation process aligns three internal signals: reconstruction outputs (RA), attention maps (AA), and encoder codewords (CA) [2509.12748]. The associated losses are
$$
\mathcal{L}_{RA} = \|\hat{\mathbf{H}}_{\text{teacher}} - \hat{\mathbf{H}}_{\text{student}}\|_2^2,
$$
$$
\mathcal{L}_{AA} = \frac{1}{L_{\text{MSA}}} \sum_{l=1}^{L_{\text{MSA}}} \frac{1}{N_h^{(l)}} \sum_{i=1}^{N_h^{(l)}} \|\mathbf{A}_{\text{teacher}}^{(l,i)} - \mathbf{A}_{\text{student}}^{(l,i)}\|_2^2,
$$
$$
\mathcal{L}_{CA} = \frac{1}{N_d} \|\mathbf{z}_{\text{teacher}} - \mathbf{z}_{\text{student}}\|_2^2,
$$
with overall objective
$$
L = L_{\text{NMSE}} + \lambda_1 \mathcal{L}_{RA} + \lambda_2 \mathcal{L}_{AA} + \lambda_3 \mathcal{L}_{CA},
$$
where $\lambda_1=0.3$, $\lambda_2=2.0$, and $\lambda_3=2.0$ [2509.12748].

NEFT-Hybrid replaces the Transformer encoder with a lightweight CNN stack, while retaining a Transformer decoder [2509.12748]. The paper motivates this change by noting encoder-side local correlations, described as diagonal attention, and targets asymmetric deployment in which the UE encoder is the main computational bottleneck [2509.12748]. The encoder consists of a three-stage CNN with $3\times 3$ convolutions, batch normalization, and ReLU; each stage halves spatial resolution and doubles channels, producing a compact $4\times 4 \times C_0$ representation [2509.12748]. The per-layer convolutional complexity is given as
$$
\mathrm{FLOPs} = H_{\text{out}} W_{\text{out}} C_{\text{out}} K_h K_w C_{\text{in}}.
$$
The paper states that this reduces encoder FLOPs by up to approximately $50\%$ compared to a Transformer encoder, and reports encoder-side reductions up to $64\%$ across scenarios [2509.12748].

NEFT-Edge is a distilled student of NEFT-Hybrid with reduced dimensions while preserving the hybrid structure [2509.12748]. It is intended for IoT and edge deployments with tighter resource budgets [2509.12748].

At compression ratio $\gamma=16$, the paper reports the following experimental figures [2509.12748]:

| Model | Params / FLOPs | NMSE |
|---|---:|---:|
| NEFT | 551,740 params; 6.35M FLOPs | −31.14 dB |
| NEFT-Compact | 439,220 params; 4.70M FLOPs | −30.13 dB |
| NEFT-Hybrid | 427,841 params; 4.07M FLOPs | −26.12 dB |
| NEFT-Edge | 300,340 params; 2.61M FLOPs | −24.24 dB |

For the same setting, NEFT-Compact reduces parameters by $20.4\%$ and FLOPs by $25.98\%$ relative to NEFT, with NMSE degradation of about $1.01$ dB [2509.12748]. NEFT-Edge reduces parameters by $29.8\%$ relative to Hybrid, reduces total FLOPs by $35.9\%$, and reduces encoder FLOPs by $48.8\%$ [2509.12748]. Similar reductions are reported at $\gamma=32$ and $\gamma=64$, with Compact and Edge reducing total FLOPs by $25$--$36\%$ and Hybrid lowering encoder complexity by up to $64\%$ [2509.12748].

These variants formalize NEFT as a deployment spectrum rather than a fixed architecture. The full model targets maximal reconstruction performance, while Compact, Hybrid, and Edge progressively trade representational capacity for lower total or encoder-side complexity.

## 4. Training objectives, optimization, and feedback pipeline

The primary reconstruction metric is NMSE,
$$
\mathrm{NMSE} = \mathbb{E}\!\left[\frac{\|\hat{H} - H\|_2^2}{\|H\|_2^2}\right],
$$
reported in decibels as
$$
\mathrm{NMSE(dB)} = 10\log_{10}\!\left(\frac{\mathbb{E}[\|H-\hat{H}\|_F^2]}{\mathbb{E}[\|H\|_F^2]}\right).
$$
The paper also reports cosine similarity,
$$
\rho = \mathbb{E}\!\left[\frac{\langle \mathbf{H}, \hat{\mathbf{H}}\rangle_F}{\|\mathbf{H}\|_F \|\hat{\mathbf{H}}\|_F}\right].
$$
These metrics are used to evaluate the fidelity of CSI reconstruction after feedback compression [2509.12748].

The training dataset is generated using LoS near-field channels with Rayleigh distance $d_R$ and uniform sampling of $r \in [0.05 d_R, 0.5 d_R]$ and $\theta \in [0,2\pi]$ [2509.12748]. The dataset contains $100$k training samples, $10$k validation samples, and $10$k test samples [2509.12748]. Optimization uses AdamW for $200$ epochs with cosine-annealed learning rate
$$
\eta_t = \eta_{\min} + \frac{1}{2}(\eta_{\max}-\eta_{\min})\left[1 + \cos\!\left(\frac{\pi t}{T}\right)\right],
$$
where $\eta_{\max}=1e\!-\!4$, $\eta_{\min}=0$, and $T=200$ [2509.12748]. Early stopping is applied if NMSE fails to improve by at least $0.1$ dB over $20$ epochs [2509.12748]. During knowledge-distillation training, the learning rate is increased to $3e\!-\!4$ [2509.12748]. The implementation environment is PyTorch 2.6.0, CUDA 12.4, and an RTX 4080 GPU; the details section additionally lists an i7-13700K [2509.12748].

The feedback pipeline is explicit. The encoder input is the normalized $2\times 32\times 32$ real-imaginary tensor; patch embedding yields $L_1=64$ tokens, patch merging yields the $2C_1\times 4\times 4$ bottleneck, and flattening produces the feedback vector of length $K=N_d$ [2509.12748]. The BS decoder reshapes the received vector back to the bottleneck tensor and reconstructs the CSI through hierarchical upsampling and Transformer decoding [2509.12748]. The encoder resides at the UE and the decoder at the BS [2509.12748].

Quantization and entropy coding are not incorporated in the reported experiments; the paper trains and evaluates with continuous floating-point feedback vectors [2509.12748]. It states that uniform or nonuniform quantization and entropy coding may be added post-training in practical deployments, but these are treated as orthogonal add-ons rather than part of NEFT proper [2509.12748]. Residual or iterative refinement is likewise not used; reconstruction relies on the hierarchical ViT decoder [2509.12748].

## 5. Experimental evaluation and ablation results

The reported baselines are CsiNet, a CNN encoder-decoder, and ExtendNLNet, a CNN model with Non-Local blocks [2509.12748]. SwinCFNet is discussed in the paper but is not used as a baseline in the reported tables [2509.12748]. The evaluation is performed on the narrowband near-field LoS dataset under the BS and UE configuration described above [2509.12748].

Across all tested compression ratios, NEFT substantially outperforms the reported baselines in NMSE [2509.12748]. At $\gamma=16$, NEFT achieves $-31.14$ dB, compared with $-13.19$ dB for ExtendNLNet and $-10.06$ dB for CsiNet, corresponding to an improvement of approximately $18$--$21$ dB; the cosine similarity is reported as $99.94\%$ [2509.12748]. At $\gamma=32$, NEFT achieves $-28.67$ dB, versus $-11.19$ dB for ExtendNLNet and $-8.40$ dB for CsiNet [2509.12748]. At $\gamma=64$, NEFT achieves $-19.55$ dB, versus $-9.89$ dB and $-6.77$ dB, respectively [2509.12748].

NEFT-Hybrid achieves $-26.12$, $-24.29$, and $-17.82$ dB NMSE at $\gamma=16$, $32$, and $64$, respectively, while using the lowest or comparable FLOPs among the baselines and markedly reducing encoder complexity [2509.12748]. NEFT-Compact and NEFT-Edge reduce total FLOPs by $25$--$36\%$ with small NMSE degradation; the paper summarizes Compact as incurring about $1.0$--$1.6$ dB loss and Edge about $1.9$--$3.9$ dB loss depending on $\gamma$ [2509.12748].

The ablation results emphasize the role of multi-level knowledge distillation. At $\gamma=16$, removing KD or using reconstruction-only losses degrades Compact from $-30.13$ dB to $-28.98$ dB for Compact-onlyRecon and to $-28.33$ dB for Compact-w/o-KD [2509.12748]. For Edge, the corresponding degradation is from $-24.24$ dB to $-23.18$ dB for Edge-onlyRecon and to $-22.55$ dB for Edge-w/o-KD [2509.12748]. The paper also reports stable convergence in the loss curves, with the student approaching teacher loss while retaining a minor capacity gap [2509.12748].

Architectural ablations indicate that larger $4\times 4$ patches and a two-stage hierarchy reduce sequence length from $64$ to $16$ while preserving global attention, yielding approximately $17\%$ of W-MSA cost [2509.12748]. Relative position bias is reported to support spatial modeling without window shifting [2509.12748]. The paper does not directly study robustness to SNR mismatch, user mobility, or array scaling, and states that such robustness is not directly examined [2509.12748].

## 6. Interpretation, practical significance, and limitations

NEFT is explicitly positioned as a practical and scalable solution for near-field CSI feedback in XL-MIMO systems [2509.12748]. Its central contribution is to reconcile global Transformer-based modeling with the computational constraints of CSI feedback, especially at the UE encoder. The full NEFT model addresses accurate reconstruction, while Hybrid and Edge directly target asymmetric hardware scenarios in which encoder-side resources are much tighter than decoder-side resources [2509.12748].

Several practical limitations are stated. First, the training and evaluation are restricted to narrowband near-field LoS channels [2509.12748]. Performance under mixed LoS/NLoS conditions, wideband OFDM settings, or hybrid near-/far-field propagation is not validated in the paper [2509.12748]. Second, the feedback vector is continuous rather than quantized, so quantization bottlenecks and entropy coding effects are not part of the reported performance numbers [2509.12748]. Third, GPU training and inference are reported, but mobile or SoC latency and memory footprints are not measured [2509.12748].

The paper identifies several future directions, including cross-domain adaptation such as domain randomization, mixed-precision inference, online fine-tuning, and integration with variable-length bit-level feedback [2509.12748]. This suggests that the current NEFT formulation is best interpreted as a floating-point algorithmic core that could be integrated into a fuller communication stack rather than as a complete air-interface design.

A further point of clarification concerns naming. NEFT should not be conflated with other uses of “near-field” in the Transformer literature. For example, FMMformer uses a near-/far-field attention decomposition inspired by the fast multipole method for long-sequence modeling [2108.02347]. By contrast, NEFT’s “near-field” refers to electromagnetic propagation physics in XL-MIMO CSI, not to a banded/local attention decomposition [2509.12748]. The two works address different problem domains and employ different efficiency mechanisms.

## 7. Position within near-field CSI feedback research

Within near-field CSI feedback research, NEFT represents a Transformer-centered alternative to CNN-based feedback architectures such as CsiNet and ExtendNLNet, which serve as its direct baselines [2509.12748]. Its main empirical claim is that global dependency modeling on the reshaped spatial CSI tensor is especially effective for near-field spherical-wave channels, provided that the attention cost is controlled through aggressive patching and hierarchy [2509.12748].

The paper’s scope is distinct from approaches that explicitly exploit near-field sparsity in transformed domains. For example, vector-quantized deep joint source-channel coding for XL-MIMO near-field CSI feedback has been studied using a polar-delay transform, hierarchical CNN and Transformer backbones, and discrete latent transmission over an AWGN channel [2601.07584]. NEFT does not use polar-delay preprocessing, vector quantization, or DJSCC training; it instead learns directly from the spatial tensor and evaluates floating-point feedback [2509.12748]. This suggests a methodological contrast between representation-first and transform-first paradigms for near-field CSI compression.

In that context, NEFT can be understood as establishing that high-fidelity near-field CSI feedback is achievable with a hierarchical Vision Transformer even without explicit sparse-domain preprocessing, provided that token count is constrained and deployment-aware variants are available [2509.12748]. Its reported $15$--$21$ dB NMSE gains over the cited baselines and the availability of Compact, Hybrid, and Edge variants position it as a reference design for efficient near-field CSI feedback under diverse hardware constraints [2509.12748].

Source: https://www.emergentmind.com/topics/near-field-efficient-feedback-transformer-neft