{"id":"7be0c794-45c8-4f57-b13a-e61c4f25ea43","arxiv_id":"2607.04795","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Shift-MoE, a DJSCC encoder-decoder with channel-grouped shifts and gated MLP experts, lowers NMSE of multi-user PASS CSI feedback versus CsiNet, MLP, and Transformer baselines under noisy uplink conditions.","lead":"This paper designs a neural CSI feedback scheme for multi-user pinching-antenna systems that uses local shift operations and a mixture-of-experts network to compress and send channel matrices over a noisy uplink. It matters because pinching antennas need accurate CSI in FDD networks, and fixed compressors struggle with their grid structure and user-to-user differences.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Idealized perfect feedback CSI is a real soft spot, but the more load-bearing gap is that gains may be generic DJSCC/MoE rather than PASS-specific.","rationale":"The paper is a clean, self-contained IEEE-letter engineering contribution: explicit PASS geometry and multipath model (§II-A), clear Shift-MoE blocks (channel-grouped one-step shifts + gated MLP-MoE, Eqs. 13–18), and consistent NMSE gains across K, η, and (M,N) (Figs. 3–4). The reader’s weakest assumption is correctly identified and is acknowledged by the authors. I only partially agree because that idealization is shared by the baselines under the same DJSCC pipeline; therefore it does not uniquely invalidate the relative ranking, but it does leave absolute robustness unproven. The more decisive open question for the strongest claim is whether the gains are PASS-specific or would appear for any structured multi-user CSI under the same DJSCC+MoE recipe. A single imperfect-ˆh_fb experiment (or a non-PASS control) would settle both the reader’s concern and the attribution gap. No code/data release and purely synthetic evaluation keep the verdict CONDITIONAL rather than ACCEPT; nothing in the manuscript warrants REJECT. Verdict remains CONDITIONAL, confidence still high on the limited claim as stated.","tokens_in":9194,"tokens_out":730,"duration_ms":6182,"concrete_test":"Retrain/evaluate Shift-MoE and the strongest baseline (ShiftViT or Transformer) under the same imperfect feedback CSI model: replace ˆh_fb_k = h_fb_k by a noisy estimate with estimation SNR 0–10 dB below the data SNR, and report NMSE vs. SNR for K=5, M=16, N=8, η=1/2. If the relative gap collapses below ~1–2 dB or reverses at medium SNR, the headline robustness claim weakens; if the gap holds, the reader’s concern is largely mitigated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that Shift-MoE consistently outperforms CsiNet/MLP/Transformer/ShiftViT-w/o-MoE on multi-user PASS CSI feedback (Abstract; §IV, Figs. 3–4). The reader correctly flags the idealized equalizer ˆh_fb_k = h_fb_k (§II-B after Eq. 10), which the paper itself lists as future work (§V). That idealization is a genuine soft spot: residual feedback-channel estimation error is never injected, so reported NMSE curves are optimistic upper bounds on robustness. However, a more load-bearing concern for the strongest claim is attribution. All methods appear trained/evaluated under the same perfect-equalization DJSCC setup; the baselines are generic CSI-feedback nets (CsiNet, TransNet-style Transformer) not re-designed for the waveguide–PA grid, and no ablation isolates whether the gains come from (i) end-to-end DJSCC vs. separate source coding, (ii) shift vs. attention on any 2-D CSI, or (iii) MoE under any multi-user heterogeneity. Without a non-PASS (e.g., standard massive-MIMO) control or imperfect-ˆh_fb curves, it remains unclear that the architecture is necessary for PASS structure rather than a generally stronger DJSCC backbone.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"This letter studies CSI feedback for multi-user pinching-antenna systems (PASS) under FDD, where reciprocity is unavailable and the uplink feedback link is noisy. The authors formulate an end-to-end DJSCC problem that maps each user’s complex M×N PA-level CSI matrix to a compact latent code, transmits it over a quasi-static Rician feedback channel, and reconstructs the CSI at the BS. The proposed Shift-MoE encoder/decoder uses channel-grouped one-step shifts to capture waveguide–PA grid correlations without global attention, and a gated MLP mixture-of-experts module to adapt to heterogeneous multi-user CSI statistics. Numerical results on synthetic geometric multipath data report consistent NMSE gains over CsiNet, MLP, Transformer, and ShiftViT-without-MoE baselines across SNR, user count, compression ratio, and antenna configurations, with expert gating weights shown to vary across users.","tokens_in":9614,"tokens_out":1032,"duration_ms":7525,"significance":"PASS is an emerging reconfigurable architecture; CSI feedback under FDD is a practical bottleneck that has received little dedicated treatment. The paper’s contribution is a concrete, low-complexity architecture that couples structure-aware shift interactions with input-adaptive MoE under a DJSCC objective, together with systematic ablations (shift vs. MLP; MoE vs. ShiftViT) and scaling curves over K, η, and (M,N). If the gains hold under more realistic feedback-channel estimation and non-PASS controls, the design would be a useful template for grid-structured CSI feedback. Strengths include a clearly stated system model and training objective, explicit complexity comparison to MHSA, and multi-parameter numerical validation rather than a single operating point.","major_comments":[{"comment":"§II-B after Eq. (10): the simulations set ˆh_fb_k = h_fb_k (idealized perfect feedback CSI at the equalizer). Residual feedback-channel estimation error is never injected, so the reported NMSE curves are optimistic upper bounds on robustness. The paper itself lists imperfect feedback-channel knowledge as future work (§V). For a DJSCC claim whose central selling point is robustness over a noisy uplink, at least one imperfect-ˆh_fb curve (or a short sensitivity study) is load-bearing and should be added or the claim should be explicitly scoped to perfect equalization.","section":null},{"comment":"§IV, Figs. 3–4 and the Abstract claim: all methods appear trained/evaluated under the same perfect-equalization DJSCC setup, and the baselines (CsiNet, TransNet-style Transformer, MLP) are generic CSI-feedback nets not re-designed for the waveguide–PA grid. There is no non-PASS (e.g., standard massive-MIMO) control and no separate source-coding-plus-channel-coding baseline. Without these, it remains unclear whether the gains are PASS-structure-specific or simply those of a generally stronger DJSCC backbone (shift + MoE). A short control experiment or a clearer attribution discussion is needed to support the PASS-specific framing of the strongest claim.","section":null}],"minor_comments":[{"comment":"Table I and §III: free hyperparameters (dm, N1, N2, ndiv, α1, α2) are listed but not ablated; a brief sensitivity note would strengthen reproducibility.","section":null},{"comment":"§II-A, Eq. (6)–(8): the waveguide radiation vector and the definition of Hk are clear, but the relationship between the effective channel gk(Λ) and the matrix fed back could be restated once for readers less familiar with PASS.","section":null},{"comment":"Fig. 3 legend and text: “Transformer-based schemes [7], [15]” should name the exact variants used (e.g., TransNet) so the comparison is reproducible.","section":null},{"comment":"Minor notation: compression ratio η ≜ L/(2MN) is defined after Eq. (9); ensure L vs. Lc is used consistently in the text and figures.","section":null},{"comment":"Fig. 5: expert gating weights for five users are informative; stating the SNR and (M,N,η) at which they were collected would help interpretation.","section":null}],"recommendation":"major_revision","confidential_remarks":"Fit for a letter-length venue is reasonable if the two major points (imperfect feedback CSI and attribution/control) are addressed. The idealized equalizer is already flagged by the authors as future work; requiring at least a sensitivity curve is proportionate rather than scope-expanding. Novelty relative to generic DJSCC+MoE is the main editorial risk if the PASS-specific claim is left untested."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean IEEE-letter style piece that puts a Shift-MoE encoder-decoder into end-to-end DJSCC for multi-user PASS CSI feedback under FDD. What is actually new is the setting plus the concrete architecture: channel-grouped one-step shifts for the waveguide-PA grid plus gated MLP experts for user heterogeneity, trained jointly over a noisy uplink. They are among the first to treat this combination systematically.\n\nThe paper does the basics well. System model and PA-level channel (LoS + multipath) are stated clearly. Architecture equations for the shifts, gating, and expert aggregation are explicit. They ablate ShiftViT without MoE against MLP, CsiNet, and Transformer baselines, show consistent NMSE gains across SNR, user count, compression ratio, and (M,N), and give a simple complexity argument that shifts remove the quadratic attention term. Fig. 5’s per-user expert weights support the heterogeneity claim. Math and citation pattern look solid for the genre; free parameters are listed in Table I.\n\nSoft spots are real but proportionate. All data are synthetic draws from the same geometric model used for training, so generalization is untested. The equalizer assumes perfect instantaneous knowledge of the feedback fading coefficient (they flag imperfect knowledge as future work). Gains could partly be generic stronger DJSCC rather than PASS-specific structure; there is no non-PASS control experiment. No code or data release. None of these sink the central empirical claim under the stated assumptions.\n\nThis is for people already working on PASS, reconfigurable antennas, or learning-based CSI feedback who need a practical DJSCC baseline for multi-user FDD sims. It deserves a serious referee at a letters venue. I would not bring it to reading group unless we are covering PASS that week, and I would not cite it myself in the next year, but I would accept it for peer review.","headline":"Competent letter-length engineering paper: Shift-MoE DJSCC for multi-user PASS CSI feedback works in synthetic sims, novelty is the application, perfect feedback CSI is the main soft spot.","tokens_in":10173,"tokens_out":494,"would_cite":false,"duration_ms":9465,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A shift-plus-experts neural codec reconstructs multi-user pinching-antenna CSI over noisy FDD feedback better than standard learning baselines.","keywords":["pinching-antenna systems","CSI feedback","deep joint source-channel coding","mixture-of-experts","shift operation","FDD multi-user","NMSE"],"falsifier":"Retrain and re-evaluate the same Shift-MoE pipeline when the equalizer uses only a noisy pilot-based estimate of the feedback fading coefficient; if the NMSE advantage over CsiNet, Transformer, and ShiftViT disappears or reverses, the central claim fails under realistic feedback CSI.","tokens_in":10111,"feed_emoji":"📡","tokens_out":620,"duration_ms":7434,"temperature":0.7,"pith_summary":"Pinching-antenna systems can reshape wireless propagation cheaply, but in FDD networks the base station still needs accurate downlink CSI from each user over a noisy uplink feedback link. That CSI lives on a waveguide-by-antenna grid with clear local structure, yet the statistics differ strongly from user to user, so a single fixed compression map fails to generalize. The authors design an end-to-end deep joint source-channel codec called Shift-MoE: channel-grouped one-step shifts capture the grid correlations without quadratic attention, while a gated mixture of MLP experts adapts the mapping to each user's CSI distribution. Joint training over a Rician feedback channel yields lower normalized reconstruction error than CsiNet, plain MLP, Transformer, and shift-only variants across user counts, antenna sizes, and compression ratios. The result matters because it shows how structure-aware, user-adaptive feedback can unlock the promised multi-user gains of pinching antennas under realistic FDD constraints.","feed_headline":"Shift-MoE beats standard CSI feedback for pinching antennas","feed_subtitle":"Local shifts plus expert routing cut reconstruction error over noisy FDD links across users and rates","key_machinery":"Shift-MoE: deterministic one-step shifts of channel groups along the waveguide and pinching-antenna axes, followed by a token-wise gated mixture of MLP experts that routes each CSI realization to a weighted combination of specialists.","core_discovery":"An end-to-end DJSCC architecture that combines channel-grouped one-step feature shifts with a gated MLP mixture-of-experts module reconstructs multi-user PASS CSI more accurately over a noisy uplink than representative learning-based CSI feedback schemes, and the gains hold under changes in user number, waveguide-PA size, and feedback rate.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Shift-MoE cuts NMSE for multi-user pinching-antenna CSI feedback","Local shifts plus gated experts improve PASS CSI over noisy uplinks","Shift-MoE tops learning baselines on multi-user PASS CSI reconstruction","One-step shifts and MoE adapt CSI feedback across users and rates","Shift-MoE reconstructs heterogeneous PASS CSI better than fixed mappings"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The base station is given perfect knowledge of the instantaneous uplink feedback fading coefficient when it equalizes the received symbols, so residual feedback-channel estimation error never appears in the reported results.","fun_headline_variants_meta":{"raw":{"variants":["Shift-MoE cuts NMSE for multi-user pinching-antenna CSI feedback","Local shifts plus gated experts improve PASS CSI over noisy uplinks","Shift-MoE tops learning baselines on multi-user PASS CSI reconstruction","One-step shifts and MoE adapt CSI feedback across users and rates","Shift-MoE reconstructs heterogeneous PASS CSI better than fixed mappings"]},"model":"grok-4.5","effort":"low","cost_usd":0.004458,"raw_usage":{"total_tokens":1251,"prompt_tokens":714,"num_sources_used":0,"completion_tokens":98,"cost_in_usd_ticks":44580000,"prompt_tokens_details":{"text_tokens":714,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":439,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":714,"tokens_out":98,"duration_ms":4071,"temperature":1.0,"reasoning_tokens":439,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T13:24:24.220505+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain and re-evaluate the same Shift-MoE pipeline when the equalizer uses only a noisy pilot-based estimate of the feedback fading coefficient; if the NMSE advantage over CsiNet, Transformer, and ShiftViT disappears or reverses, the central claim fails under realistic feedback CSI.","supporting_citations":[],"review_version":1}