Paper deep dive
Semantic CSI Feedback for Beam Selection: When Task-Aware Embeddings from Sparse Pilots Outperform Full-Bandwidth Reconstruction
Cristian J. Vaca-Rubio, Konstantinos Vandikas, Aneta Vulgarakis Feljan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/22/2026, 4:36:25 AM
Summary
This paper proposes a semantic communication framework for CSI feedback in FDD massive MIMO systems, where the UE transmits a task-aware semantic embedding optimized for beam selection rather than reconstructing the full channel. The study demonstrates that a compact embedding of 8 real values derived from only 43 sparse NR CSI-RS pilots in the angular-delay domain achieves higher beam prediction accuracy (up to 89%) than methods using full 512-subcarrier bandwidth or classical reconstruction-based approaches like CsiNet. Key findings include the superiority of the angular-delay domain for sparse pilots, the detrimental effect of linear interpolation, and the failure of signal-level reconstruction under extreme compression.
Entities (8)
Relation Signals (6)
Semantic Embedding → hasdimension → 8
confidence 98% · semantic embedding of just d=8 real values
Semantic CSI Feedback → achievesaccuracy → 89%
confidence 95% · AD+Semantic... achieves the highest beam prediction accuracy... 89.1%
Semantic CSI Feedback → outperforms → CsiNet
confidence 95% · Semantic feedback (Task-Aware) achieves 83.5–89% from the same 8-dimensional bottleneck. CsiNet achieves 2–6% accuracy... near-random
Angular-Delay Domain → isoptimalfor → Sparse Pilots
confidence 93% · AD wins decisively with sparse pilots... AD+Semantic (89.1%) surpasses every SF method
Linear Interpolation → degrades → Beam Prediction Accuracy
confidence 92% · Interpolation actively hurts... interpolation degrades accuracy across all evaluated methods
Spatial-Frequency Domain → isoptimalfor → Full CSI
confidence 90% · With full CSI, SF+Semantic leads at 83.5% because the spatial-frequency domain preserves all 512 frequency bins
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Classical CSI feedback in FDD massive MIMO transmits a compressed reconstruction of the channel, optimizing fidelity to the original signal regardless of the downstream task. We propose a semantic communication perspective: instead of reconstructing the channel, the UE transmits a learned \emph{semantic embedding} optimized end-to-end for beam selection at the gNB. Comparing reconstruction-oriented feedback (CsiNet) against task-aware semantic feedback across two input domains and three observation scenarios, we show that a semantic embedding of just $d=8$ real values from only 43 NR CSI-RS pilots in the angular-delay domain achieves the highest beam prediction accuracy, outperforming every method with access to the full 512-subcarrier channel. The key insight is that beam-relevant information is intrinsically low-dimensional: the semantic encoder learns to discard reconstruction-irrelevant structure and retain only a compact representation that is relevant to beam selection, realizing the core principle of semantic communication: transmit the intent, not the signal.
Tags
Links
- Source: https://arxiv.org/abs/2609.18368v2
- Canonical: https://arxiv.org/abs/2609.18368v2
Trouble viewing inline? Open PDF directly →
Full Text
22,746 characters extracted from source content.
Expand or collapse full text
Semantic CSI Feedback for Beam Selection: When Task-Aware Embeddings from Sparse Pilots Outperform Full-Bandwidth Reconstruction Cristian J. Vaca-Rubio Konstantinos Vandikas Aneta Vulgarakis Feljan Abstract Classical CSI feedback in FDD massive MIMO transmits a compressed reconstruction of the channel, optimizing fidelity to the original signal regardless of the downstream task. We propose a semantic communication perspective: instead of reconstructing the channel, the UE transmits a learned semantic embedding optimized end-to-end for beam selection at the gNB. Comparing reconstruction-oriented feedback (CsiNet) against task-aware semantic feedback across two input domains and three observation scenarios, we show that a semantic embedding of just d=8d=8 real values from only 43 NR CSI-RS pilots in the angular-delay domain achieves the highest beam prediction accuracy, outperforming every method with access to the full 512-subcarrier channel. The key insight is that beam-relevant information is intrinsically low-dimensional: the semantic encoder learns to discard reconstruction-irrelevant structure and retain only a compact representation that is relevant to beam selection, realizing the core principle of semantic communication: transmit the intent, not the signal. Index Terms: semantic communication, beam management, CSI feedback, deep learning, FDD massive MIMO, 5G NR †address: Ericsson Research, Stockholm, Sweden cristian.vaca.rubio, konstantinos.vandikas, aneta.vulgarakis@ericsson.com 1 Introduction Beam selection in 5G NR FDD systems presents a fundamental information asymmetry as the gNB needs to choose a transmit beam, but can only observe the downlink channel indirectly through UE reports. The standardized solution, exhaustive beam sweeping, scales linearly with codebook size and becomes prohibitive for large antenna arrays [1, 7, 5]. Deep learning offers a path forward to predict the optimal beam from compressed channel observations [2, 6, 4]. Existing learned feedback schemes such as CsiNet [15] and its transformer-based successors [10, 12] follow the classical communication paradigm which compresses the channel for faithful reconstruction at the receiver, then perform beam selection on the reconstructed CSI. This two-stage separation is suboptimal because the encoder is agnostic to the downstream task. Consequently, it preserves signal-level fidelity which may not be directly related to a given downstream task. Semantic and task-oriented communication [13, 8, 9, 11] challenges this separation by designing the encoder to transmit only the information relevant to the receiver’s intent. In our context, the “semantics” of a channel observation are not its subcarrier-level values, but the identity of the optimal beam. We instantiate this principle as semantic CSI feedback, where the UE encodes its pilot observations into a compact embedding z trained end-to-end for beam prediction at the gNB. The classification loss shapes z to encode beam identity rather than channel coefficients; a decoder retained during training grounds the embedding in physical channel structure but is discarded at inference. Our evaluation reveals three insights: (i) the optimal input domain depends on observation density, i.e., SF wins with full bandwidth, AD wins decisively with sparse pilots; (i) a task-aware embedding of just d=8d=8 reals from 43 raw pilots surpasses every full-bandwidth method, and interpolation actively hurts; (i) two-stage compress-then-classify fails catastrophically because the autoencoder discards beam-discriminative structure to minimize signal-level error. 2 System Model 2.1 Downlink Channel and Beam Codebook We consider a gNB equipped with an 8×88× 8 uniform planar array (UPA, with Naz=Nel=8N_az=N_el=8 and Nt=NazNel=64N_t=N_azN_el=64 antennas) serving single-antenna UEs over Nsc=512N_sc=512 OFDM subcarriers at 3.5 GHz3.5\,GHz. The downlink channel for UE k at subcarrier n is denoted by k[n]∈ℂNth_k[n] ^N_t. The beam codebook is constructed as a 2D oversampled DFT matrix whose steering vectors are defined as (μ,ν)=az(μ)⊗el(ν),a(μ,ν)=a_az(μ) _el(ν), (1) with elements [az(μ)]m=1Nazej2πmμ[a_az(μ)]_m= 1 N_aze^j2π mμ for m=0,…,Naz−1m=0,…,N_az-1 and [el(ν)]l=1Nelej2πlν[a_el(ν)]_l= 1 N_ele^j2π lν for l=0,…,Nel−1l=0,…,N_el-1. The spatial frequencies are sampled with an oversampling factor O=2O=2, such that μp=pO⋅Naz,νq=qO⋅Nel, _p= pO· N_az, _q= qO· N_el, (2) for p=0,…,2Naz−1p=0,…,2N_az-1 and q=0,…,2Nel−1q=0,…,2N_el-1. This yields a total of Nb=4NazNel=256N_b=4N_azN_el=256 candidate beams, collected into a codebook =bb=1NbW=\w_b\_b=1^N_b where each bw_b corresponds to a unique steering vector (μp,νq)a( _p, _q). The optimal beam index for UE k maximizes the average received power across subcarriers, defined as bk∗=argmaxb∈1,…,Nb1Nsc∑n=1Nsc|bHk[n]|2.b_k^*= _b∈\1,…,N_b\\; 1N_sc _n=1^N_sc|w_b^Hh_k[n]|^2. (3) After discarding beams that are never optimal for any UE in the dataset, Ncls=88N_cls=88 active beam classes remain. Channels are generated via DeepMIMO v4 [3] using the asu_campus_3p5 scenario at 3.5 GHz3.5\,GHz. 2.2 What the UE Observes In FDD, the UE measures downlink pilots but cannot send the raw channel back. The observed channel is corrupted by additive noise: ~k[n]=k[n]+[n],∼(0,σ2), h_k[n]=h_k[n]+n[n], (0,σ^2I), (4) where σ2σ^2 is set according to the per-sample SNR. We consider three observation scenarios: (i) Full CSI: all 512 subcarriers are observed (TDD-equivalent upper bound); (i) Raw pilots: one pilot every 12 subcarriers, yielding Np=43N_p=43 noisy measurements at positions np∈0,12,24,…,504n_p∈\0,12,24,…,504\; (i) Interpolated pilots: linear interpolation recovers estimates at all 512 subcarriers from the 43 pilot positions. The key question is whether filling in the gaps helps or whether the network is better served by receiving clean-but-sparse observations. 3 The Role of Input Domain Before any learning begins, the choice of representation determines what information survives sparsification. Spatial-Frequency (SF). Real/imaginary stacking of the channel: SF∈ℝ2×64×NscX_SF ^2× 64× N_sc. With 512 subcarriers, this is complete. With 43 pilots, most of the frequency axis is empty, leaving the network with a comb pattern missing 91.6% of its teeth. Angular-Delay (AD). A 2D transform (FFT along antennas, IFFT along frequency) maps the channel into the angle-delay domain. With full CSI, the IFFT output is truncated to Nτ=64N_τ=64 delay taps: AD∈ℝ2×64×64X_AD ^2× 64× 64. With raw pilots, the IFFT operates on the Np=43N_p=43 pilot observations directly, yielding AD∈ℝ2×64×43X_AD ^2× 64× 43. The multipath channel is sparse in this domain, with most energy concentrated in the first few taps corresponding to dominant propagation paths. Crucially, even from 43 pilot observations, the IFFT recovers a meaningful delay-domain representation because the channel has far fewer degrees of freedom than subcarriers. The AD transform thus acts as a free compressor that concentrates information without any learned parameters, serving as a physics-informed preprocessing step that reduces the burden on the downstream neural network. 4 Feedback Architectures Fig. 1 illustrates the semantic feedback pipeline. Let ∈ℝ2×Nt×NfX ^2× N_t× N_f denote the preprocessed input (SF or AD domain), where NfN_f depends on the observation scenario. We define a semantic encoder fθ:ℝ2×Nt×Nf→ℝdf_θ ^2× N_t× N_f\! ^d at the UE, a decoder gϕ:ℝd→ℝ2×Nt×Nfg_φ ^d\! ^2× N_t× N_f, and a semantic interpreter cψ:ℝd→ℝNclsc_ψ ^d\! ^N_cls, both at the gNB, with bottleneck dimension d=8d=8 reals. The UE computes the semantic embedding =fθ()z=f_θ(X) and transmits z over the uplink feedback channel. In the semantic communication framework [13], z represents the beam-relevant content of the observation rather than a compressed replica of the signal itself. Figure 1: Semantic CSI feedback pipeline. The UE extracts a semantic embedding =fθ()z=f_θ(X), transmitted over the feedback link. At the gNB, the semantic interpreter predicts the beam directly from z (b^=argmaxcψ() b= \,c_ψ(z)). The decoder (^=gϕ() X=g_φ(z)) serves as a structural regularizer during training, grounding the embedding in physical channel structure, but is not required at inference. 4.1 Direct Classification (Upper Bound) A convolutional network maps the full input directly to beam logits: b^=argmaxhω() b= \,h_ω(X), where hω:ℝ2×Nt×Nf→ℝNclsh_ω:R^2× N_t× N_f ^N_cls. No compression is applied; the gNB is assumed to have full access to X. This models TDD reciprocity or an unconstrained feedback link, serving as a simple approach to benchmark the solution. 4.2 CsiNet-Style Autoencoder: Signal-Level Feedback In the classical approach, the encoder is optimized for signal reconstruction, analogous to traditional source coding that preserves waveform fidelity without regard for the receiver’s task. We adopt the CsiNet compress-then-reconstruct paradigm [15] with architectural modifications, namely batch normalization and LeakyReLU activations for training stability, alongside adaptive average pooling in the encoder to handle variable input sizes across scenarios. The core principle, i.e., train an autoencoder for signal fidelity, then classify on the reconstruction, is preserved. The autoencoder (fθ,gϕ)(f_θ,g_φ) is trained to minimize reconstruction error: minθ,ϕ[‖gϕ(fθ())−‖22]. _θ,φ\;E [\|g_φ(f_θ(X))-X\|_2^2 ]. (5) Once converged, the parameters (θ,ϕ)(θ,φ) are frozen. A classifier cψc_ψ is then trained on the reconstruction: minψ[ℒCE(cψ(gϕ(fθ())),b∗)]. _ψ\;E [L_CE\! (c_ψ(g_φ(f_θ(X))),\,b^* ) ]. (6) Note that cψc_ψ receives ^=gϕ() X=g_φ(z), not z itself. The encoder has no incentive to preserve beam-discriminative features that happen to be orthogonal to the reconstruction objective. 4.3 Task-Aware: Semantic Feedback The task-aware architecture realizes the semantic communication principle [14] by optimizing the encoder not for signal fidelity, but for the compact representation needed by the receiver to determine the beam identity. Critically, the semantic interpreter cψc_ψ operates directly on the embedding z, not on a reconstructed signal. All parameters are trained jointly: minθ,ϕ,ψ[α‖gϕ(fθ())−‖22⏟ℒMSE+βℒCE(cψ(fθ()),b∗)⏟ℒCE], _θ,φ,ψ\;E [α \|g_φ(f_θ(X))-X\|_2^2_L_MSE+β L_CE(c_ψ(f_θ(X)),\,b^*)_L_CE ], (7) with α=0.1α=0.1 and β=1.0β=1.0. The asymmetric weighting (β≫αβ α) formalizes the semantic priority, in that beam prediction defines the relevant downstream task while reconstruction provides structural grounding. This formulation has two important consequences: (i) The semantic gradient directly shapes z. Since cψc_ψ acts on =fθ()z=f_θ(X), the ℒCEL_CE gradient propagates through the encoder without passing through the decoder. The encoder learns to allocate capacity in the embedding to beam-discriminative structure (the channel’s semantic content), even if this increases reconstruction error. In the language of semantic communication, the encoder extracts and creates a representation of the beam identity while it discards syntax (exact subcarrier values). (i) The decoder grounds semantics in physics. Without ℒMSEL_MSE, the pipeline reduces to a bottlenecked classifier (encoder→ ), which may overfit to spurious correlations, particularly under sparse, noisy observations. The reconstruction objective constrains z to remain a physically interpretable channel representation, preventing semantic collapse to task-specific shortcuts. This regularization is most pronounced under pilot-only observations, where pure classification overfits but semantic feedback improves accuracy (Table 1). The semantic embedding z is also a general-purpose representation, allowing the same compact message to serve multiple gNB tasks (beam selection, rank adaptation) without requiring task-specific re-encoding at the UE, thereby embodying the multi-purpose nature of semantic representations advocated in [13]. 4.4 Training Protocol All models are trained with Adam (lr=10−3=10^-3), batch size 128, for up to 50 epochs with early stopping (patience 15) and d=8d=8 bottleneck. Training SNR is drawn uniformly from [−5,30][-5,30] dB. Data is split 70/15/15 (train/val/test); evaluation is at fixed SNR points from −5-5 to 3030 dB and at infinite SNR. 5 Results 5.1 The Domain Reversal Table 1: Top-1 beam accuracy (%) points with d=8d=8. Scenario Method −5-5 55 1010 1515 3030 ∞ Full CSI SF+Direct 63.1 78.5 79.9 80.1 80.3 80.2 SF+CsiNet 3.9 5.2 5.1 5.2 4.5 4.5 SF+Semantic 61.0 81.0 83.2 83.2 83.6 83.5 AD+Direct 58.2 64.5 68.6 72.5 76.2 76.5 AD+CsiNet 3.5 3.9 4.8 4.7 5.0 5.0 AD+Semantic 58.8 63.0 68.1 71.9 75.8 75.9 Pilot Raw SF+Direct 50.2 69.8 70.6 71.3 71.7 71.6 SF+CsiNet 2.7 3.1 3.5 3.6 4.0 3.8 SF+Semantic 17.8 62.4 66.8 68.6 69.2 69.3 AD+Direct 66.3 83.5 85.8 85.2 85.6 85.5 AD+CsiNet 1.7 2.3 2.5 2.6 2.7 2.7 AD+Semantic 71.4 86.9 88.6 88.5 89.0 89.1 Pilot Interp SF+Direct 38.7 62.3 63.8 64.2 64.2 64.3 SF+CsiNet 2.3 3.2 3.1 3.9 3.6 3.5 SF+Semantic 11.2 48.7 60.1 61.8 63.4 63.5 AD+Direct 55.4 80.5 82.6 83.1 83.4 83.6 AD+CsiNet 2.6 2.7 2.6 2.7 2.0 2.0 AD+Semantic 49.8 80.2 83.2 84.2 84.2 84.2 Table 1 reveals the central finding. With full CSI, SF+Semantic leads at 83.5% because the spatial-frequency domain preserves all 512 frequency bins and the classifier exploits this richness. But with raw pilots, the ranking inverts, as AD+Semantic (89.1%) surpasses every SF method and even exceeds the best full-CSI result. SF with 43 pilots forms a comb signal, consisting of scattered frequency samples with no spatial continuity along the frequency axis. The network must learn to extract beam-relevant patterns from this fragmented observation. AD, by contrast, applies an IFFT that redistributes the pilot information into a compact delay profile. Because the physical channel has far fewer resolvable multipaths than subcarriers, the delay representation captures essentially all beam-relevant information from 43 pilots. Moreover, raw pilots outperform even full CSI in the AD domain because the pilot grid acts as a physics-matched dimensionality reduction. With 512 subcarriers, the IFFT yields a 64-tap delay profile where beam-relevant energy concentrates in the first few taps, i.e., the encoder must learn to ignore the remainder. With 43 pilots, the sparse sampling already discards this irrelevant fine-frequency structure, presenting the encoder with an input pre-aligned with the channel’s multipath sparsity. Fig. 2 shows this reversal persists across SNRs. Figure 2: Top-1 accuracy vs. SNR across three observation scenarios. The domain preference (SF vs. AD) reverses between full CSI and pilot-based observation. 5.2 Why Interpolation Hurts Conventional wisdom suggests that interpolating pilots back to full bandwidth should help because more inputs implies more information. The data, however, indicates otherwise, as interpolation degrades accuracy across all evaluated methods. Linear interpolation creates plausible-looking but physically incorrect frequency responses between pilot positions. The network cannot distinguish genuine channel structure from interpolation artifacts, and learns to rely on patterns that do not generalize. The “less is more” principle applies, where clean sparse data outperforms noisy dense data. The semantic encoder extracts beam-relevant structure directly from sparse pilots more effectively than any reconstruct-then-classify approach. 5.3 Signal-Level vs. Semantic Feedback Figure 3: Top-1 accuracy at infinite SNR (d=8d=8 bottleneck). Signal-level feedback (CsiNet) collapses to near-random in every scenario. Semantic feedback (Task-Aware) achieves 83.5–89% from the same 8-dimensional bottleneck. Fig. 3 exposes the failure of signal-level feedback under extreme compression. Both CsiNet and Semantic use identical encoder architectures compressing to the same d=8d=8 bottleneck; the only difference is what shapes z. CsiNet achieves 2–6% accuracy (near the 1/88≈1.1%1/88≈ 1.1\% random) because the 8 values that minimize MSE retain no beam-discriminative structure. The semantic formulation (7) resolves this issue by having cψc_ψ operate directly on z, allowing the ℒCEL_CE gradient to reshape the embedding to encode which beam rather than which channel, achieving 89% accuracy from the same 8 reals. This validates the core tenet of semantic communication: under extreme bandwidth constraints, jointly optimizing the encoder with the downstream task vastly outperforms optimizing for signal fidelity. 6 Conclusion We have demonstrated that semantic CSI feedback fundamentally outperforms classical signal-level reconstruction for beam selection. The design recipe is: (1) transform pilots to the angular-delay domain, exploiting multipath sparsity; (2) do not interpolate missing subcarriers; (3) compress to d=8d=8 reals and train end-to-end for beam prediction. The resulting 32-bit (4 bits per value) feedback payload is comparable to existing CQI/PMI reports, yet achieves 89% accuracy from a single CSI-RS transmission. Future work will address quantization-aware training and generalization across scenarios. Acknowledgment A Large Language Model (Claude, Anthropic) assisted with code development and text post-editing. All research design, experimental decisions, and scientific claims are the sole responsibility of the authors. Compliance with Ethical Standards This is a numerical simulation study for which no ethical approval was required. Conflicts of Interest The authors are employees of Ericsson Research. No external funding was received. The authors have no other relevant financial or nonfinancial interests to disclose. References [1] 3GPP (2023) NR; Physical layer procedures for data. Note: TS 38.214, v17.4.0 Cited by: §1. [2] A. Alkhateeb, S. Alex, P. Varkey, Y. Li, Q. Qu, and D. Tujkovic (2018) Deep learning coordinated beamforming for highly-mobile millimeter wave systems. IEEE access 6, p. 37328–37348. Cited by: §1. [3] A. Alkhateeb (2019) DeepMIMO: a generic deep learning dataset for millimeter wave and massive mimo applications. External Links: 1902.06435, Link Cited by: §2.1. [4] J. Chen, W. J. Hillery, K. R. Mestav, C. Nuzman, I. Saniee, and Y. Xing (2025) CSI compression for massive mimo: model-based or data-driven?. IEEE Wireless Communications 32 (1), p. 22–27. Cited by: §1. [5] R. M. Dreifuerst and R. W. Heath (2023) Massive mimo in 5g: how beamforming, codebooks, and feedback enable larger arrays. IEEE Communications Magazine 61 (12), p. 18–23. Cited by: §1. [6] R. M. Dreifuerst and R. W. Heath (2025) Neural codebook design for mimo network beam management. IEEE Transactions on Wireless Communications. Cited by: §1. [7] M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi (2019) Standalone and non-standalone beam management for 3gpp nr at mmwaves. IEEE Communications Magazine 57 (4), p. 123–129. Cited by: §1. [8] D. Gündüz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C. Chae (2022) Beyond transmitting bits: context, semantics, and task-oriented communications. IEEE Journal on Selected Areas in Communications 41 (1), p. 5–41. Cited by: §1. [9] D. Gündüz, M. A. Wigger, T. Tung, P. Zhang, and Y. Xiao (2024) Joint source–channel coding: fundamentals and recent progress in practical designs. Proceedings of the IEEE. Cited by: §1. [10] J. Guo, C. Wen, S. Jin, and G. Y. Li (2022) Overview of deep learning-based csi feedback in massive mimo systems. IEEE Transactions on Communications 70 (12), p. 8017–8045. Cited by: §1. [11] S. Guo, Y. Wang, N. Zhang, Z. Su, T. H. Luan, Z. Tian, and X. Shen (2024) A survey on semantic communication networks: architecture, security, and privacy. IEEE communications surveys & tutorials 27 (5), p. 2860–2894. Cited by: §1. [12] H. Ju, S. Jeong, S. Kim, B. Lee, and B. Shim (2024) Transformer-assisted parametric csi feedback for mmwave massive mimo systems. IEEE Transactions on Wireless Communications 23 (12), p. 18774–18787. Cited by: §1. [13] Q. Lan, D. Wen, Z. Zhang, Q. Zeng, X. Chen, P. Popovski, and K. Huang (2021) What is semantic communication? a view on conveying meaning in the era of machine intelligence. Journal of Communications and Information Networks 6 (4), p. 336–371. Cited by: §1, §4.3, §4. [14] M. B. Mashhadi and D. Gündüz (2021) Pruning the pilots: deep learning-based pilot design and channel estimation for mimo-ofdm systems. IEEE Transactions on Wireless Communications 20 (10), p. 6315–6328. Cited by: §4.3. [15] C. Wen, W. Shih, and S. Jin (2018) Deep learning for massive mimo csi feedback. IEEE Wireless Communications Letters 7 (5), p. 748–751. Cited by: §1, §4.2.