We present a Scalable Windowed Axial-attention Neural Receiver (SWAN-Rx) for multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems. The receiver improves reception accuracy and reduces forward-pass computation while supporting frozen-weight reuse across bandwidth allocations and channel conditions. It combines a common resource-element representation of received observations and pilot context with alternating time-axis and local frequency-axis attention. Shared parameters and relative positional information permit soft detection across allocation lengths without inference-time parameter updates. At fixed subcarrier spacing, model configuration, and symbol count, bounded frequency neighborhoods give linear growth of forward-pass computation with bandwidth. Evaluations compatible with the 5G New Radio physical layer reuse weights trained on four physical resource blocks (PRBs) at up to 273 PRBs and on unseen channels. At the training bandwidth, SWAN-Rx achieves lower pre- and post-decoding bit error rates than the implemented three-step neural receiver (NRX) and DeepRx-MIMO baselines, with 16.9% and 39.7% fewer parameters and 56.3% and 35.7% fewer forward-pass operations, respectively. The same weights achieve lower pre-decoding error rates than these baselines throughout the evaluated bandwidth-extrapolation and 16-PRB channel-transfer tests, with an estimated 0.4-0.9 dB signal-to-noise ratio reduction relative to NRX at a pre-decoding error rate of 10−2 in the 16-PRB tests.
Figures & tables
Fig. 1: Resource-grid-to-sequence mapping. Each of the NsymNsc sequence elements contains the complete receive vector yt,k∈CNrx at one RE. Colors distinguish data REs and pilot REs assigned to the two streams. The mapping preserves the time-frequency coordinates of each RE. Grid dimensions are schematic.
Fig. 2: SWAN-Rx architecture. RE tokens combine received observations, pilot context, and role and time embeddings. Five pairs of time-axis and windowed frequency-axis attention layers process the grid, followed by the LLR head.
Fig. 3: Optional CE+SD extension. Intermediate CE features feed both the channel-estimation head and the SD stage, which produces bit LLRs.
Parameter
Setting
Transmit / receive antennas
2 / 4
Spatial streams / modulation
2 / 64-QAM
Carrier / subcarrier spacing
2.6 GHz / 30 kHz
Training PRBs / subcarriers
4 / 48 effective
OFDM symbols / pilot symbols
14 / 2
Training FFT size / CP length
64 / 6 samples
TABLE I: Default System and Training Configuration
Fig. 4: Reception on UMa at 4 PRBs: (a) pre-decoding BER; (b) post-decoding BER.
Fig. 5: Frozen-weight bandwidth extrapolation from 4 PRBs on UMa. Columns: 10 and 20 dB; (a)/(b): pre-decoding BER; (c)/(d): post-decoding BER. The dotted line marks the training bandwidth.
Fig. 6: Post-decoding BER under frozen-weight transfer from UMa to (a) TDL-B and (b) UMi at 4 PRBs.
Fig. 7: Frozen-weight transfer from 4-PRB UMa training to five unseen channels at 16 PRBs. Top: pre-decoding BER; bottom: post-decoding BER.
Fig. 8: Optimized inference at batch size 1 over 4–273 PRBs: latency (left) and peak allocated memory (right).
Receiver
Parameters
GFLOPs/sample
4 PRBs
16 PRBs
SWAN-Rx
507,916
0.726
2.906
NRX, 1 step
270,342
0.737
2.949
NRX, 2 steps
440,838
1.200
4.801
NRX, 3 steps
611,334
1.663
6.654
DeepRx-MIMO
842,588
1.130
4.519
TABLE II: Parameter counts and forward-pass computation. NRX depth variants are separately trained.
PRBs
Receiver
Samples/s
Memory (MiB)
4
SWAN-Rx †
6028.6
59.9
4
NRX, 3 steps
3076.3
79.3
4
DeepRx-MIMO †
4819.8
82.5
273
SWAN-Rx †
61.5
1165.4
273
NRX, 3 steps
27.3
10777.7
273
DeepRx-MIMO †
56.2
4730.8
TABLE III: Throughput and peak allocated memory at batch size 32. † indicates more decoded bit errors than the original implementation in at least one validation condition.
Fig. 9: Frequency-attention range comparison on UMa at 16 PRBs with weights trained at 4 PRBs. Left: pre-decoding BER; right: post-decoding BER.
Fig. 10: Post-decoding BER of one-, two-, and three-step NRX and SWAN-Rx on UMa: 4 PRBs (left); 16 PRBs (right).
Fig. 11: Post-decoding BER across 4 pilot layouts on UMa at 20 dB: 4 PRBs (left); 16 PRBs (right).
Fig. 12: SWAN-Rx attention on UMa at 15 dB, 4 PRBs, and pilot symbols (2,11) . Left: time attention; right: frequency attention. White dashed lines mark pilot symbols; color scales are shared within each column.
Fig. 13: Detection BER of SWAN-Rx and CE4+SD6 on UMa at 4 PRBs. Both use QK normalization and SwiGLU. Left: pre-decoding BER; right: post-decoding BER.
Fig. 14: Channel-estimation NMSE on data REs at 10 dB: (a) UMa; (b) UMi. Each curve averages 4 pilot layouts. LMMSE uses independently calibrated covariance matched to each channel and bandwidth; CE4+SD6 uses frozen 4-PRB UMa-trained weights.
We propose a cross-attention Transformer for joint decoding of uplink OFDM signals received by multiple coordinated access points. A shared per-receiver encoder learns the time-frequency structure of each grid, and a token-wise cross-attention module fuses the receivers to produce soft log-likelihood ratios for a standard channel decoder without explicit channel estimates. Trained with a bit-metric objective, the model adapts its fusion to per-receiver reliability and remains robust under degraded links, strong frequency selectivity, and sparse pilots. Over realistic Wi-Fi channels, it outperforms classical pipelines and strong neural baselines, often matching or surpassing a local perfect-CSI reference while remaining compact and computationally efficient on commodity hardware, making it suitable for next-generation coordinated Wi-Fi receivers.
Reliable channel estimation and symbol detection in 5G New Radio (NR), as well as emerging 6G radio systems, rely on demodulation reference signals (DMRS). The density and placement of these pilots on the orthogonal frequency-division multiplexing (OFDM) time--frequency resource grid affect both receiver reliability and spectral efficiency. Sparse DMRS layouts increase the time--frequency distance between pilot observations and data resource elements (REs), which can make pilot information harder to exploit with purely local processing. This paper proposes PARGAR (Pilot-Aware Graph Attention Receiver), a graph-attention neural receiver that represents the OFDM resource grid as a graph for bit detection. The graph combines local time--frequency edges with pilot-anchor links from data REs to selected pilot REs, and we further study a channel-scaled pilot-anchor rule based on approximate coherence time and coherence bandwidth. Using Sionna-generated OFDM simulations, we evaluate PARGAR against DeepRx, a convolutional neural network (CNN)-based receiver, across dense, two-symbol sparse, and single-symbol sparse DMRS layouts under channel-mismatch scenarios. Under the considered SISO uncoded setup, channel-scaled PARGAR explicitly injects pilot-aware inductive bias and yields lower BER in selected sparse-pilot scenarios while remaining competitive under dense pilots.
Sajedeh Norouzi, Mostafa Rahmani Ghourtani, Eric Samikwa +1
Institute of Computer Science, University of Bern, Bern, Switzerland. · 5G/6GIC, Institute for Communication Systems (ICS) University of Surrey, U.K.
This letter studies CSI denoising for MIMO--OFDM with variable NR resource block (RB) allocations. ReFLEX is a length-generalizable Transformer whose frequency attention uses a relative-frequency position bias (RFPB) generated from subcarrier offsets. A single checkpoint handles unseen RB lengths and can be applied to sparse DM-RS observations in the tested RB5/RB10 PUSCH setup without retraining. In a 3GPPTR38.901 UMa NLOS channel, ReFLEX achieves about −9.6dB NMSE on unseen RB lengths. In NR PUSCH/UL-SCH simulations, ReFLEX denoising followed by time-frequency interpolation reduces the 10% BLER threshold by about 2--3dB.
Zhibin Zhang, Robert Potekhin, Ziwei Wan +2
Moscow Institute of Physics and Technology (State University), Moscow, Russia · Yangtze Delta Region Academy, Beijing Institute of Technology (Jiaxing), Jiaxing 314019, China · School of Interdisciplinary Science, Beijing Institute of Technology, Beijing 100081, China