{"id":"b7a72910-0b15-4f24-b393-8f8155f18ab3","arxiv_id":"2507.20477","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Randomly permuting each user's JSCC features turns multi-user interference into roughly Gaussian noise, letting a single pretrained encoder and diffusion denoiser serve all users.","lead":"This paper proposes a multi-user wireless image transmission scheme in which each user's encoded features are randomly shuffled so that interference from other users looks like ordinary noise, which a pretrained diffusion model can clean up. It claims to match or beat existing semantic communication multiple-access methods without retraining user-specific models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's Gaussianization premise is assumed, not established: a random shuffle preserves the multiset of feature entries and cannot by itself produce the N(0,I) joint distribution on which diffusion denoising and step matching depend.","rationale":"The reader's weakest-assumption analysis and my independent reading converge on the same point: Lemma 1's Gaussianization premise is the single most load-bearing unsupported step. The framework is otherwise coherent—shuffle-based user separation, diffusion denoising, and semantic-aware beamforming form a plausible pipeline—and the paper deserves credit for the ablation study and for using pretrained point-to-point models without retraining. However, the mathematical foundation of the proposed approach is weaker than its empirical presentation suggests: random permutation is a combinatorial operation that preserves order statistics, so it cannot manufacture joint Gaussianity from arbitrary latent vectors; at most it can scramble spatial structure. The empirical plots and the remark in Section III-A are suggestive but do not establish the distributional identity used in Eq. (10) and Eq. (20). Because this concern is addressable by a concrete statistical test and a matched synthetic-noise ablation, it supports a conditional acceptance rather than rejection. Since the reader already recommended CONDITIONAL with the same core concern, my assessment leaves the verdict unchanged.","tokens_in":28017,"tokens_out":4038,"duration_ms":48892,"concrete_test":"On the same encoder and dataset used in Section III-A, compute z = C_k^{-1}(C_m(f_m)) for many independent image pairs and shuffles, for K=2 and K=8. Test multivariate normality against N(0,I) with a standard statistic (Mardia skewness/kurtosis or Henze-Zirkler) and plot the Q-Q of squared Mahalanobis distances against chi-square_{2N}, reporting the maximum and 99th-percentile deviations. Then run the full receiver twice at the same SINR: once with true shuffled interference and once with synthetic Gaussian interference of the same per-coordinate variance. If the Gaussianity test rejects at the 1% level, or if PSNR/LPIPS/CLIP differ by more than a small threshold (e.g., 0.5 dB or 0.01 LPIPS), Lemma 1's premise is unsupported and the diffusion model is operating outside its training distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Lemma 1's premise that C_k^{-1}(C_m(f)) ~ N(0,I) for m != k. This is not proven. Conditional on the latent vector f, a random permutation P preserves the multiset of entries of f, so the shuffled vector is supported on permutations of that fixed vector; unconditionally, z = C_k^{-1}(C_m(f)) is a mixture of permuted copies of the latent distribution, not a multivariate Gaussian unless f itself has very special invariance properties. The paper's support (Remark 1 and Fig. 2(a)) is an empirical marginal histogram, and Section III-A simultaneously reports that p(f) deviates from Gaussian. The proof of Lemma 1 then assumes each interferer is exactly Gaussian and uses additivity to conclude the sum is N(0,I) in Eq. (10); this is precisely the unproven point. Because Eq. (10), the step-matching rule (20), and the diffusion denoiser trained only on AWGN all presuppose an independent isotropic Gaussian effective noise, any residual non-Gaussianity—heavy tails, coordinate dependence, or variance mismatch—propagates directly into the receiver and cannot be repaired by the denoiser. The central multi-user claim is therefore conditional on an empirical Gaussianization that the paper never quantifies.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-user downlink semantic communication framework in which a single, point-to-point-trained JSCC encoder/decoder and a single diffusion denoiser support all users. Inter-user interference is handled by assigning each user a unique random shuffling pattern before complex symbol mapping; the authors argue that this makes interfering JSCC features statistically indistinguishable from Gaussian noise, so that diffusion denoising and step matching designed for AWGN channels can be reused without retraining. A semantic-aware beamforming problem is formulated through a generalized-logistic surrogate for reconstruction quality, and an extension for semantically correlated users groups users by CLIP similarity and applies a cooperative phase-aligned beamforming strategy. The paper reports simulations on COCO and VIMEO showing that the proposed method improves PSNR, LPIPS, and CLIP scores over ADJSCC, OMDMA, and DeepMA baselines, together with ablations supporting the diffusion denoiser and shuffle-based mapping.","tokens_in":28338,"tokens_out":5486,"duration_ms":64690,"significance":"If the central Gaussianization premise holds, the framework is practically significant: it removes the need for user-specific JSCC models and additional training for multi-user semantic communication, while also providing a privacy mechanism through user-specific shuffling patterns. The paper deserves credit for a clean system formulation, for reusing pretrained point-to-point models as an external artifact, and for evaluating the approach against standard multi-user baselines with visual examples. However, the significance is conditional: the theoretical core, Lemma 1, assumes rather than proves that shuffled interferers are independent isotropic Gaussian noise, and the correlated-data extension relies on an unquantified empirical phase condition. These issues must be addressed before the claimed generality can be accepted.","major_comments":[{"comment":"The proof of Lemma 1 assumes as its hypothesis the property C_k^{-1}(C_m(f)) ∼ N(0,I) for m≠k, but this property is asserted rather than derived. Conditional on a fixed latent vector f, a random permutation merely reorders the entries, so the shuffled vector is supported on permutations of that fixed vector; unconditionally, it is a mixture of permuted copies of the latent distribution, which is not jointly Gaussian unless p(f) already has very special invariance properties. The paper's own Fig. 2(a) reports that p(f) deviates from Gaussian, and the empirical justification in Remark 1 and Section III-A is based on marginal histograms and t-SNE projections, which do not establish joint Gaussianity or independence across coordinates. Since Eq. (10), the step-matching rule (20), and the use of a diffusion denoiser trained only on AWGN all depend on the interference-plus-noise being independent isotropic Gaussian, this is the load-bearing step of the paper. I ask the authors to either provide conditions under which the shuffled interference provably converges to N(0,I), or supply quantitative distributional evidence on the actual SA-1B features (e.g., covariance and independence tests, higher-order cumulants, or a Gaussianity test statistic), or explicitly reframe Lemma 1 as an empirical approximation with measured error bounds.","section":"§III-A, Remark 1 and Lemma 1 (Eq. (10))"},{"comment":"Even if the marginal distributions of the shuffled interference are approximately Gaussian, the step-matching rule (20) matches only the scalar ratio τ_k/sqrt(α_k^2+τ_k^2) against the diffusion schedule sqrt(β̄_t). The denoiser is trained on additive white Gaussian noise in a point-to-point setting, so any residual structure in the interference—coordinate dependence, heterogeneous variances across feature dimensions, heavy tails, or correlation between the real and imaginary components—survives the power normalization and is not repaired by the denoiser. The paper should quantify how close the actual interference-plus-noise input is to the Gaussian assumption used at training time, for example by feeding true Gaussian noise and shuffle-transformed interference into the same receiver at the same power and reporting the resulting output PSNR/LPIPS/CLIP, rather than only qualitative curves in Fig. 2(b).","section":"§III-B, Algorithm 1, Eq. (20)"},{"comment":"The beamforming objective in P1 is built on the generalized-logistic approximation S(γ)≈a+b/(c+γ^{-e}), but the hyperparameters a,b,c,d,e are never specified, fitted, or validated against the actual JSCC+diffusion model used in the simulations. The stated relation e=10d/ln10 is also dimensionally unclear if d is a slope in dB and e is used as an exponent on a linear SINR term. The tightness of the surrogate lower bound (26) and the convergence of Algorithm 2 depend on this model, yet the ablation in Fig. 9(a) validates only the final beamforming versus other beamformers, not the logistic approximation itself. Please provide the fitted parameter values and the data used to fit them, and include a sensitivity analysis of the reported gains with respect to these parameters.","section":"§III-C, Eq. (21)–(30)"},{"comment":"The cooperative phase-alignment step for two-user groups relies on the stated empirical condition ∠(h_i^H v̄_j)+∠(h_j^H v̄_i)≈0, which appears in the text without evidence or derivation. Since the correlated-data gains in Figs. 11 and 12 depend on this phase alignment, the paper should either justify this condition analytically, provide the measured distribution of the phase sum over the simulated channels, or relax the design so that it does not depend on an unquantified empirical observation.","section":"§IV-C, Algorithm 4"}],"minor_comments":[{"comment":"There are several typos and wording errors that should be corrected: \"tranceiver\" for transceiver, \"yeilding\" for yielding, \"distrupt\" for disrupt, and \"iteratively denies\" for \"iteratively denoises\" in the introduction.","section":"§II-A"},{"comment":"The noise notation is inconsistent: the text writes σn with n∼CN(0,1) and then says σ^2 is the noise power; please define σ explicitly as the noise amplitude or write the noise term as n∼CN(0,σ^2) to avoid ambiguity.","section":"§II-A.2, Eq. (4)"},{"comment":"The two displayed forms of the logistic approximation use γ[dB] and γ^{-e} in the same equation; please clarify the units of γ in each form and define d and e in one place without the dimensionally unclear relation e=10d/ln10.","section":"§III-C, Eq. (21)"},{"comment":"The objective in the exhaustive partition step reuses the index i in both the outer and inner sums; the inner sum should use a different index (e.g., j∈G_{l,i} and k∈G_{l,j}) to be mathematically well defined.","section":"§IV-B, Algorithm 3, line 14"},{"comment":"The parameter SN used to duplicate images and to define the video scenario is not formally defined before it appears in the description of the correlated-data experiments; please define it explicitly in Section V-A.","section":"§V-A"}],"recommendation":"major_revision","confidential_remarks":"The main issue is theoretical rather than experimental: the Gaussianization premise that makes the framework work is stated as an assumption inside Lemma 1. If the authors can supply a convincing distributional justification or an honest reframing with measured error bars, the paper could become a solid systems contribution. The heavy reliance on the authors' own pretrained models from [19] is acceptable because those models are treated as external artifacts, but the comparison with baselines trained on COCO should be checked for fairness regarding training data and model capacity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a real idea. Shuffle each user's JSCC features with a user-specific permutation, so that at any other user's receiver the interferers look like Gaussian noise, and clean them with a diffusion model trained only for point-to-point AWGN. If that works, it removes the need for user-specific or jointly trained JSCC models in multi-user semantic communication, which is the main practical obstacle for OMDMA-style schemes. The experiments support the claim reasonably well: with a shared encoder/decoder and the same diffusion denoiser, the proposed scheme matches or beats DeepMA (jointly trained) on PSNR/LPIPS/CLIP, and the ablation shows the shuffle and the denoiser each contribute.\n\nThe best part is the empirical care around the Gaussianization claim: they show that raw JSCC features are far from Gaussian and that interference is more damaging than AWGN, then show that after shuffling the empirical distribution is close to Gaussian and the damage matches AWGN. That is the right way to support a statistical premise that is hard to prove. The correlated-data extension with CLIP similarity and two-user cooperative beamforming is a nice bonus, even though the phase-alignment trick only works for pairs.\n\nThe soft spots are real but not fatal. Lemma 1 is stated as a conditional: \"Given a set of mapping functions that satisfies C^{-1}(Cm(f)) ~ N(0,I)\", and the proof then relies on that premise. The paper never quantifies how close the shuffled distribution actually is to Gaussian. t-SNE and one performance curve are suggestive, not a distributional test. There is a genuine tension: p(f) deviates from Gaussian, and a random permutation preserves the multiset of entries of each vector, so exact joint Gaussianity cannot hold for finite N. What likely saves the scheme is that the diffusion model only needs the effective noise to be roughly i.i.d. for step matching to land, and the empirical plots suggest it is good enough. But the paper should say this honestly instead of presenting the lemma as if derived. The other gap is benchmarking: they do not give the prior methods (ADJSCC, OMDMA, DeepMA) a diffusion denoiser too, so part of the gain may be the DM rather than the shuffle. No error bars, no code release, and the logistic curve parameters a-e are fitted without reporting values.\n\nWho is this for: researchers in semantic communication and multi-user MISO, who will read it as a practical framework with a plausible statistical engine. It deserves a serious referee, but that referee should push for a quantitative Gaussianity check (moments or KL), a DM-augmented baseline, and code. I would conditionally accept if those are addressed.","headline":"Genuinely useful framework for training-free multi-user semantic communication, but the central Gaussianization premise is asserted rather than derived; worth serious refereeing.","tokens_in":28802,"tokens_out":1951,"would_cite":false,"duration_ms":22175,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Shuffling each user's semantic features converts inter-user interference into Gaussian-like noise, so a single diffusion denoiser trained only for point-to-point links can serve all users without retraining.","keywords":["semantic communication","multi-user multiple access","joint source-channel coding","diffusion model denoising","shuffle-based orthogonalization","inter-user interference","semantic-aware beamforming","cooperative transmission"],"falsifier":"A direct test: take the actual shared JSCC encoder, extract a single feature vector $f$ from a typical image, shuffle it once, and run a normality check, for example a Kolmogorov-Smirnov or Shapiro-Wilk test on the entries, or a check of the off-diagonal covariance of the shuffled vector, repeated over many images to get a rejection rate. If most single shuffled vectors fail the test, the premise behind Lemma 1 fails in its simplest case, and the measurable consequence should appear with just two users (one interferer, no averaging over many): the reconstruction error of the diffusion-denoiser pipeline should visibly deviate from the curve obtained by injecting pure Gaussian noise of the same power. The same test applied to an encoder trained with an entropy or sparsity penalty would show whether the Gaussianization claim is a property of the shuffle or only of this particular encoder's feature statistics.","tokens_in":27821,"feed_emoji":"🔀","tokens_out":14043,"duration_ms":132919,"temperature":0.7,"pith_summary":"This paper tries to establish that the hardest part of multi-user semantic communication, structured interference between users' learned feature streams, can be neutralized without any user-specific models. Its device is a per-user random shuffle of the feature vector's positions before transmission: after the receiver undoes its own shuffle, every other user's signal has been rearranged and therefore looks statistically like Gaussian noise rather than like another image's semantics. That reduction lets the receiver treat interference as ordinary channel noise and strip it out with the same diffusion-model denoiser already used for point-to-point links, with the starting denoising step chosen by matching the measured interference-plus-noise power to a diffusion noise level. The paper also claims the shuffle acts as an implicit privacy key, because non-target users see only Gaussian-like streams, and it extends the design to users requesting semantically related images, where similarity-based grouping and cooperative beamforming harvest part of the interference as useful signal. The payoff, if the claim holds, is a multi-user system built from one shared encoder, decoder, and denoiser, each trained once on point-to-point channels, that matches or beats jointly trained multi-user baselines.","feed_headline":"A shuffle turns wireless interference into static","feed_subtitle":"Unique per-user shuffles let one shared model treat rivals' signals as noise, with zero retraining.","key_machinery":"The load-bearing object is the shuffle-based mapping $C_k$, which randomly permutes the $2N$ entries of the JSCC feature vector $f_k$ by a user-specific permutation $p_k$ and then pairs the reordered entries into $N$ complex symbols, $[C_k(f_k)]_i = [f_k]_{p_{k,i}} + j[f_k]_{p_{k,i}+N}$. Its work is to make every other user's contribution $C_k^{-1}(C_m(f_m))$ a shuffled rearrangement of structured features, which Lemma 1 treats as standard Gaussian noise, collapsing the received feature into $\\hat{f}_k = \\alpha_k f_k + \\tau_k n_r$, the exact form of a scaled clean signal plus Gaussian corruption. On top of that reduction sit two components: step matching, which sets the diffusion starting point by matching $\\tau_k/\\sqrt{\\alpha_k^2+\\tau_k^2}$ to the cumulative noise schedule $\\sqrt{\\bar{\\beta}_t}$ and then applies deterministic (DDIM-style) reverse sampling using a denoiser trained only on point-to-point AWGN channels, and semantic-aware beamforming, which replaces the unavailable closed-form objective with a fitted generalized-logistic curve of reconstruction quality and solves the resulting nonconvex problem by majorization-minimization with a quadratic transformation. For correlated data, a CLIP-based similarity metric groups users and a phase-aligned cooperative beamforming post-processing exploits the empirical condition $\\angle(h_i^H \\bar{v}_j) + \\angle(h_j^H \\bar{v}_i) \\approx 0$ to make paired users' interference constructive.","core_discovery":"The paper's central claim is that inter-user interference in multi-user semantic communication can be transformed, at the symbol-mapping stage, from structured semantic attack into Gaussian-like noise. Each user $k$ is assigned a unique random permutation $p_k$, and the mapping $C_k$ pairs the shuffled entries of the JSCC feature vector $f_k$ into complex channel symbols, so that the inverse mapping used at receiver $k$ returns $C_k^{-1}(C_m(f_m))$ for an interferer $m$, a shuffled version of that user's features. Invoking Lemma 1, which assumes such shuffled interference is distributed as $\\mathcal{N}(0, I)$, the received latent feature is shown to reduce to the clean form $\\hat{f}_k = \\alpha_k f_k + \\tau_k n_r$ with $n_r \\sim \\mathcal{N}(0, I)$, where $\\alpha_k$ is the desired signal gain and $\\tau_k$ aggregates interference and channel-noise power. Because this is precisely the clean-signal-plus-Gaussian form of an intermediate diffusion state, the receiver normalizes the signal, matches $\\tau_k/\\sqrt{\\alpha_k^2+\\tau_k^2}$ to the diffusion noise schedule to find a starting timestep, and runs deterministic reverse diffusion with the point-to-point denoiser. On the transmit side, beamforming is optimized against a generalized-logistic approximation of the semantic-fidelity curve rather than a closed-form rate expression. The paper reports that this recipe matches or exceeds jointly trained and user-specific baselines on PSNR, LPIPS, and CLIP scores, and that grouping semantically similar users with cooperative phase-aligned beamforming adds further gains when the transmitted images share content.","pith_inferences":["An implicit distinction the paper does not separate: Gaussianization is either a property of a single feature vector's entries (roughly i.i.d. standard Gaussian, so any permutation still looks Gaussian) or a central-limit effect over many interferers; the first fails for sparse, peaky, or spatially correlated encoder outputs, and the paper concedes in its conclusion that optimal shuffle patterns r","If the premise holds, the framework's natural regime is overloaded access, $K > N_t$, where user-specific OMDMA models become impractical; the ablation in the paper suggests the shuffle's benefit concentrates precisely where the user count exceeds the antenna count.","Because the trained models are user-agnostic, the shuffling pattern can be rotated per session or per image block without retraining, effectively turning the scheme into a low-cost physical-layer cipher in the semantic domain.","A cheap pre-deployment audit would test the premise directly: run a normality test on the entries of shuffled feature vectors from the actual encoder; rejection would predict where the diffusion denoiser begins to hurt rather than help, and would motivate the content-adaptive 'local shuffle' the paper leaves to future work."],"forward_implications":["A base station can add or remove users without any retraining: the shared JSCC encoder-decoder and diffusion denoiser are trained once in a point-to-point setting, and only the per-user permutation pattern changes.","Interference no longer needs a statistical model: after shuffling, all structured semantic interference is demoted to channel noise, so the whole point-to-point denoising chain applies unchanged.","Shuffling patterns function as implicit private keys: an eavesdropping or non-target user sees a stream statistically indistinguishable from Gaussian noise, so semantic content is not recoverable without the intended pattern.","Semantic quality under interference, measured by LPIPS and CLIP score, improves most exactly in the regime where interference dominates, high SNR with many users, because that is where demoting interference to noise changes the outcome.","For users requesting correlated content, grouping by CLIP similarity and pairing users for phase-aligned cooperative beamforming converts part of the interference into useful signal energy, producing the largest gains at low SNR."],"supporting_citations":[{"why":"Supplies the point-to-point JSCC encoder-decoder and diffusion denoiser that the proposed framework reuses without retraining, including the step-matching recipe.","marker":"[19]"},{"why":"The OMDMA baseline that motivates the work; its user-specific-model route to semantic orthogonality is what shuffle-based orthogonalization replaces.","marker":"[20]"},{"why":"DeepMA baseline; jointly trained multi-user deep multiple access whose retraining burden the proposed scheme removes.","marker":"[22]"},{"why":"ADJSCC baseline model applied to all users in the benchmark comparisons and used to instantiate OMDMA.","marker":"[9]"},{"why":"Provides the image corpus whose 20,000 extracted feature vectors ground the empirical distribution analysis that motivates the shuffle.","marker":"[24]"},{"why":"Supplies the deterministic reverse-sampling (DDIM) update used in the denoising algorithm's iterative loop.","marker":"[26]"},{"why":"Supplies the generalized-logistic approximation and MM lower bound that turn the semantic beamforming problem into a solvable surrogate.","marker":"[27]"},{"why":"Provides the equivalence that lets the power-constrained beamforming problem be solved by a power-normalized reformulation.","marker":"[28]"},{"why":"Provides the phase-rotation invariance property used to align phases within cooperative user pairs.","marker":"[29]"}],"fun_headline_variants":["Shuffle-based orthogonalization makes interference Gaussian noise","One universal JSCC model, unique shuffles per user","Random permutations convert multi-user interference to harmless static","Diffusion denoising with shuffle orthogonalization outperforms baselines","Shuffling patterns double as private keys in semantic channels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire scheme rests on the premise, stated as the hypothesis of Lemma 1 in Section III-B and supported only by the empirical plots of Section III-A, that after shuffling, one user's feature vector is statistically indistinguishable from standard Gaussian noise; but a random shuffle only rearranges the entries of a fixed vector, so a vector whose entries are not already roughly independent Gaussian samples cannot become Gaussian by shuffling.","fun_headline_variants_meta":{"raw":{"variants":["Shuffle-based orthogonalization makes interference Gaussian noise","One universal JSCC model, unique shuffles per user","Random permutations convert multi-user interference to harmless static","Diffusion denoising with shuffle orthogonalization outperforms baselines","Shuffling patterns double as private keys in semantic channels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2579,"prompt_tokens":1163,"completion_tokens":1416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":779,"completion_tokens_details":{"reasoning_tokens":1337}},"tokens_in":779,"tokens_out":1416,"duration_ms":11698,"temperature":1.0,"reasoning_tokens":1337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:43:28.162654+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: take the actual shared JSCC encoder, extract a single feature vector $f$ from a typical image, shuffle it once, and run a normality check, for example a Kolmogorov-Smirnov or Shapiro-Wilk test on the entries, or a check of the off-diagonal covariance of the shuffled vector, repeated over many images to get a rejection rate. If most single shuffled vectors fail the test, the premise behind Lemma 1 fails in its simplest case, and the measurable consequence should appear with just two users (one interferer, no averaging over many): the reconstruction error of the diffusion-denoiser pipeline should visibly deviate from the curve obtained by injecting pure Gaussian noise of the same power. The same test applied to an encoder trained with an entropy or sparsity penalty would show whether the Gaussianization claim is a property of the shuffle or only of this particular encoder's feature statistics.","supporting_citations":[{"cited_title":"Orthogonal model division multiple access,","cited_arxiv_id":null,"evidence_quote":"The OMDMA baseline that motivates the work; its user-specific-model route to semantic orthogonality is what shuffle-based orthogonalization replaces."},{"cited_title":"DeepMA: End-to-end Deep Multiple Access for Wireless Image Transmission in Semantic Communication","cited_arxiv_id":"2303.11543","evidence_quote":"DeepMA baseline; jointly trained multi-user deep multiple access whose retraining burden the proposed scheme removes."},{"cited_title":"Beamforming design for semantic-bit coexisting communication sys- tem,","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized-logistic approximation and MM lower bound that turn the semantic beamforming problem into a solvable surrogate."},{"cited_title":"Iterative algorithm induced deep-unfolding neural networks: Precoding design for multiuser MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Provides the equivalence that lets the power-constrained beamforming problem be solved by a power-normalized reformulation."}],"review_version":2}