{"id":"2f1a4dce-f74c-4f5b-96ed-4f4dab981c46","arxiv_id":"2506.17164","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Encoding messages before splitting codewords yields an RSMA variant with slightly better sum-rate and similar fairness to conventional RSMA, plus easier implementation.","lead":"This paper proposes CS-RSMA, a rate-splitting multiple access variant that encodes each user's message first, then splits the resulting codeword into common and private parts. It reports similar or better throughput than conventional RSMA while simplifying decoding, signaling, and retransmission.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CS-RSMA's reported SR gain over SIC-free RSMA is built into the optimization, not demonstrated by the architecture; a flexible conventional RSMA baseline should be tested.","rationale":"The reader's weakest assumption is that Eq. (25), the additive decomposition c_k I_c,k + I_non-SIC_p,k, may overcount the achievable rate. In good faith, this decomposition is plausible and likely correct: the CS-RSMA decoder in Eq. (30) is a product of marginal mismatched metrics for the common and private segments, and for independently generated codeword segments the GMI of a product metric adds. So the equation is not obviously an overestimate; it is more an omitted proof than a demonstrated error. The more decisive issue is the fairness of the comparison. The paper itself admits in Remark 3 that the only difference between the SIC-free conventional RSMA problem and the CS-RSMA problem is changing min to max. That means the SR gain is by construction, and a conventional RSMA variant that lets c_k = 0 users skip common decoding would attain exactly the same sum-rate objective as CS-RSMA. This directly undermines the abstract's comparative claim, while leaving the architectural and practical contributions intact. The proposed check would settle whether any genuine SR benefit remains under a fair baseline. Until then, the conditional verdict is appropriate, but for a different reason than the reader's primary concern.","tokens_in":19635,"tokens_out":34138,"duration_ms":358030,"concrete_test":"Re-run the SIC-free conventional RSMA sum-rate optimization with the common-rate constraint R_c <= min_{k in S} I_{c,k}, where S = {k : c_k > 0} is optimized jointly with c and P (i.e., users with c_k = 0 may skip common decoding), using the same channels, alphabets, complexity constraint delta, and precoder initializations as Section V. Compare ergodic SR with CS-RSMA in Fig. 3. If the gap narrows to zero or reverses, the claim 'better SR than conventional RSMA' should be replaced by 'no SR loss under a flexible baseline'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline SR advantage is substantially an artefact of an asymmetric baseline. Equations (23)-(24) impose min_{k'} I_{c,k'} on the common rate for conventional RSMA, and Remark 1 justifies this by asserting that conventional RSMA requires all users to decode the common stream. Proposition 2 and Remark 3 then show that the SIC-free conventional sum-rate problem (P2) differs from the CS-RSMA problem (P3) only in replacing this min with max_k I_{c,k}. Consequently, the SR improvement of CS-RSMA over SIC-free RSMA reported in Fig. 3 is a direct consequence of the formulation, not of codeword segmentation. If conventional RSMA is allowed to allocate the whole common stream to one user (c_k = 1 for the user with the largest I_{c,k}) and users with c_k = 0 are not required to decode the common stream, the sum-rate objective becomes max_k I_{c,k} + sum_k I_non-SIC_{p,k}, exactly the CS-RSMA objective in P3. The comparison therefore establishes the benefit of removing the all-users-decode-common constraint, not a performance benefit of the architecture itself. The practical complexity, signaling, and HARQ benefits are real and unaffected, but the abstract's 'better performance than conventional RSMA in SR' is stronger than what the simulations demonstrate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CS-RSMA, an RSMA architecture in which user messages are first encoded and the resulting codewords are then segmented into common and private parts, in contrast with conventional RSMA, which splits messages before encoding. The authors introduce a GMI-based analysis under mismatched decoding, define a decoding-complexity measure, formulate sum-rate (SR) and max-min fairness (MMF) precoder optimization problems for conventional RSMA and CS-RSMA, and evaluate them through numerical simulations and link-level simulations. They report that CS-RSMA slightly outperforms conventional RSMA in SR, performs similarly in MMF, and offers implementation benefits in encoding/decoding complexity, control signaling, and retransmission design.","tokens_in":19961,"tokens_out":13938,"duration_ms":147571,"significance":"If the rate expressions and comparisons were fully justified, the paper would make a useful contribution: CS-RSMA potentially preserves most of the rate benefits of RSMA while simplifying the physical-layer implementation, and the GMI framework is a relevant step toward evaluating RSMA under practical suboptimal receivers. The paper is also commendable for including link-level simulations and for explicitly defining a decoding-complexity constraint. The practical advantages discussed in Section VI-C are credible and largely independent of the contested rate analysis.","major_comments":[{"comment":"The achievable-rate expression for CS-RSMA, I_GMI,k = c_k I_c,k + I_non-SIC_p,k, is asserted in Remark 1 but is not derived from the mismatched-decoding metric. In the decoder of Eq. (30), for t in T_k the first product uses d(·) with the common symbol fixed and sums over private symbols, while the second product, taken over all t = 1,...,n, also includes t in T_k with the private symbol fixed and sums over common symbols; the same received symbol is thus counted twice. Even if the second product is meant to range only over t not in T_k, the sum of the two per-stream GMIs is not an immediate consequence of the GMI for the joint codeword, because the common and private streams are superposed rather than orthogonalized and the decoder is not a product of two independent per-stream decoders. Since Eqs. (23)-(25) are the objective functions for P2 and P3 and drive all SR/MMF numerical claims, this missing derivation is load-bearing. The authors should either prove the additive decomposition under the stated decoder or replace it with the correct joint GMI.","section":"Section III-B, Eq. (25) and Eq. (30)"},{"comment":"The claimed SR advantage of CS-RSMA over SIC-free conventional RSMA is to a large extent built into the formulation. In conventional SIC-free RSMA, a user with c_k = 0 does not need to decode the common message; the common codeword only has to be decodable by users with c_k > 0. The optimization P2 nonetheless imposes min_k I_c,k on the common rate for all users. If the common-rate constraint is instead applied only to the set {k : c_k > 0} and c is optimized, the optimal common-rate contribution becomes max_k I_c,k, yielding exactly the P3 objective of CS-RSMA. Thus the comparison in Fig. 3 establishes the benefit of relaxing the all-users-decode-common constraint rather than a benefit of codeword segmentation per se. Unless the authors justify why all users must decode the common stream even when they receive no common bits (for example, for SIC or system-level reasons), the baseline is suboptimal and the abstract's SR claim is stronger than what is demonstrated.","section":"Section IV-A, P2/P3 and Remark 1"},{"comment":"The convergence justification is not sufficient. Monotone increase and boundedness of the objective sequence do not imply convergence of the iterates to a stationary point for nonconvex subgradient ascent, and the presence of min/max operators and the logarithmic barrier further complicates the argument. Since the SR/MMF comparisons in Section V are obtained from these algorithms, the authors should either state the stationarity guarantees that can be established or present the results as heuristic with appropriate caveats.","section":"Section IV, Algorithms 1 and 3"}],"minor_comments":[{"comment":"The definition says a decoder of complexity-delta can compute exp(|·|^2) for at most n times for each received symbol; this should be delta times, not n times, since n denotes the block length.","section":"Section III-C, Definition 1"},{"comment":"The second product in Eq. (30) appears to range over all t = 1,...,n, which double-counts the observations in T_k; it likely should be restricted to t not in T_k. Even with that correction, the expression should be reconciled with the rate decomposition in Eq. (25).","section":"Eq. (30)"},{"comment":"The axis label 'GMI (bits/channel usee)' contains a typo: 'usee' should be 'use'.","section":"Figures 3, 4, 6, 7"},{"comment":"The stopping criterion contains a stray comma in 'Omega^{,v-1}_MMF'; this should be 'Omega^{v-1}_MMF'.","section":"Algorithm 3, line 10"},{"comment":"The separability of the decoding metrics for common and private streams is an assumption that underlies Eqs. (23)-(25); it should be stated more prominently and its implications for the GMI derivation should be discussed rather than relegated to a footnote.","section":"Section III-B, footnote 2"},{"comment":"The GMI formulas are imported from the companion preprint [20], which is cited as a same-day arXiv preprint. The authors should clarify the status of that work and, if space permits, include the key derivation or a more detailed citation so that the present paper is self-contained on this load-bearing point.","section":"Eqs. (9)-(15)"}],"recommendation":"major_revision","confidential_remarks":"The paper has two load-bearing issues: the rate decomposition in Eq. (25) is not derived from the mismatched decoder, and the SR comparison against SIC-free RSMA is weakened by a suboptimal baseline that imposes the min over all users. Both are fixable in a revision, and the practical implementation contributions (Section VI-C) are solid and worth preserving. The editor may also wish to verify the status of the companion preprint [20], since the GMI framework depends on it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the architecture is real and worth taking seriously, but the headline SR comparison is rigged in the paper's favor.\n\nWhat's new: CS-RSMA encodes each user's message into one codeword, then splits the codeword into common and private segments. Conventional RSMA splits the message first and encodes two codewords per user. This reordering is absent from the cited literature, and it has real practical consequences: no SIC, one codeword to decode per user, and HARQ that looks like ordinary MU-MIMO. The PHY transceiver designs and link-level simulations are a solid addition. Applying the FAGCI GMI from [20] to compare schemes under mismatched decoding is also a reasonable move, though it leans on a same-day companion preprint.\n\nThe soft spots, in order of severity. First, the SR gain over SIC-free RSMA in Fig. 3 is mostly by construction. Proposition 2 shows the two problems differ only in replacing min_k I_{c,k} with max_k I_{c,k}. CS-RSMA can put all common resources on the user with the best common rate; conventional SIC-free RSMA must make the common stream decodable by all. The paper acknowledges this in Remark 3 and even in Fig. 5, where the common-stream rate of CS-RSMA is higher for exactly that reason. So the comparison establishes the benefit of dropping the all-users-decode-common constraint, not the benefit of codeword segmentation per se. The MMF comparison, where the min is still present, shows near-identical performance, which is more credible.\n\nSecond, Eq. (25) asserts that the CS-RSMA rate is c_k I_{c,k} + I_{p,k}^{non-SIC}. That additive split is never derived from the mismatched decoder in Eq. (30). The decoder multiplies two separate metrics over disjoint time indices, and it is not obvious that the information extractable from the common segment and the private segment adds up that cleanly. If this overcounts, the SR advantage shrinks or vanishes. This needs to be proven or supported by simulation of actual codes. The LLS results show only a 0.05 dB gap at BER 1e-4, which is basically noise-level and doesn't settle the rate question.\n\nMinor: the subgradient ascent convergence argument (“objective is upper-bounded”) is not a real convergence proof, and the baseline for conventional RSMA could be more flexible. But these are secondary.\n\nBottom line: the architecture is a legitimate contribution with genuine implementation benefits. The paper deserves a serious referee, but the SR claims should be reframed or the baseline generalized to allow c_k=0 for some users, and Eq. (25) needs a derivation. As is, an editor should send it out, and the referee should ask for major revision.","headline":"A genuinely new RSMA architecture that removes SIC and reduces codeword count, but the reported SR gain over SIC-free RSMA is largely baked into the optimization baseline.","tokens_in":20453,"tokens_out":2878,"would_cite":true,"duration_ms":28061,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes CS-RSMA, a rate-splitting scheme that encodes each user once and then segments the codeword, and shows it improves on conventional RSMA in sum rate while cutting complexity, signaling, and retransmission overhead.","keywords":["rate-splitting multiple access","codeword segmentation","finite-alphabet signalling","mismatched decoding","generalized mutual information","SIC-free receiver","sum-rate optimization","max-min fairness"],"falsifier":"Run a fixed CS-RSMA configuration at a chosen SNR, vary the common-segment fraction $c_k$, and compare the measured LDPC throughput or BER with the prediction $c_k I_{c,k} + I_{p,k}^{\\mathrm{nonSIC}}$ computed from the paper's GMI formulas. If a receiver that jointly decodes the reassembled codeword achieves a rate materially different from that sum, or if the two contributions do not scale with $c_k$, the additive decomposition at the center of the sum-rate comparison is falsified.","tokens_in":19403,"feed_emoji":"📶","tokens_out":15174,"duration_ms":143273,"temperature":0.7,"pith_summary":"This paper introduces Codeword-Segmentation Rate-Splitting Multiple Access (CS-RSMA), an architecture that first encodes each user's message into one codeword and then cuts that codeword into a common segment and a private segment, rather than splitting the message before encoding as conventional RSMA does. It aims to show that this reordering preserves the known benefits of RSMA under realistic constraints: with finite-alphabet signals and receivers that treat undesired streams as Gaussian noise, CS-RSMA achieves higher sum rate than conventional RSMA with or without successive interference cancellation, and essentially the same max-min fairness, while using fewer codewords and less control signaling. If correct, the result matters because it gives network designers an RSMA implementation that is easier to build, easier to signal, and compatible with existing retransmission mechanisms, without paying a throughput penalty. The analysis is carried through generalized mutual information for mismatched decoders, precoder optimization for sum rate and fairness, and link-level simulations with LDPC codes.","feed_headline":"Codeword-segmented RSMA beats conventional RSMA in sum rate","feed_subtitle":"Encoding once, then slicing the codeword, lifts sum rate, holds fairness, cuts signaling overhead.","key_machinery":"The machinery is threefold. First, the CS-RSMA transceiver structure itself: one encoder per user, a stream splitter that divides each codeword into private and common segments, and a receiver that reassembles soft estimates of the segments before channel decoding. Second, the generalized mutual information (GMI) for the finite-alphabet Gaussian channel under interference (FAGCI), the achievable-rate metric for mismatched decoders that treat undesired private streams as Gaussian random variables; Eq. (10) is the approximation used in precoder optimization and Eq. (9) the exact rate used in evaluation. Third, the rate decomposition $I_{\\mathrm{GMI},k} = c_k I_{c,k} + I_{p,k}^{\\mathrm{nonSIC}}$: because the common stream no longer has to be decodable by every user, the sum-rate objective becomes $\\max_{k} I_{c,k} + \\sum_{k} I_{p,k}^{\\mathrm{nonSIC}}$ instead of $\\min_{k} I_{c,k} + \\sum_{k} I_{p,k}^{\\mathrm{nonSIC}}$. A decoding-complexity constraint ($\\delta$ exponentials per received symbol) keeps the comparison fair by fixing $\\delta = |\\mathcal{X}_c||\\mathcal{X}_{p,k}|$ across schemes.","core_discovery":"On its own terms, the central claim is that the order of encoding and splitting is what creates implementation cost in RSMA, not an essential ingredient of its gain. Conventional RSMA splits messages into common and private parts before encoding, so the common stream is a single message every user must decode and the common rate is capped by the weakest user, written $\\min_{k} I_{c,k}$. CS-RSMA encodes each user's message directly, segments the resulting codeword, and multiplexes the common segments, so user $k$ only decodes its own slice of the common stream. Its achievable rate is $I_{\\mathrm{GMI},k} = c_k I_{c,k} + I_{p,k}^{\\mathrm{nonSIC}}$ (Eq. (25)), where $c_k$ is the fraction of the common stream carrying user $k$'s segment, and the sum-rate optimization replaces the bottleneck $\\min_{k} I_{c,k}$ with $\\max_{k} I_{c,k}$. The paper supports this with GMI-based ergodic evaluations and link-level simulations: CS-RSMA sits slightly above conventional RSMA in sum rate, matches it in max-min fairness, and at a coded BER of $10^{-4}$ is about 0.05 dB better, while encoding, decoding, control signaling, and HARQ retransmission each simplify because one codeword per user replaces two streams.","pith_inferences":["Beyond the paper: the $\\max_k I_{c,k}$ structure implies a scheduling rule in which the transmitter decides on each channel realization which user's segment is carried by the common stream, which could harvest multiuser diversity beyond the paper's ergodic averages.","Beyond the paper: the additive decomposition in Eq. (25) is an assumption about separable decoding; a finite-blocklength experiment varying $c_k$ and measuring whether the common and private contributions add in proportion would show whether it is exact or a systematic overestimate.","Beyond the paper: because SIC cannot be applied to a common stream that is a multiplex of segments, any future CS-RSMA variant needing common-stream cancellation would have to cancel at the soft-symbol level, as the soft-SLIC receiver here does.","Beyond the paper: the same encode-then-segment principle should carry over to multi-layer RSMA, where the codeword-count saving grows with the number of common streams, though the paper does not derive the rate decomposition for that case."],"forward_implications":["The sum-rate optimum of CS-RSMA is at least as large as that of SIC-free conventional RSMA, since the objectives differ only by replacing $\\min_{k} I_{c,k}$ with $\\max_{k} I_{c,k}$; in the medium-SNR regime this preserves RSMA's advantage over SDMA.","Each user encodes and decodes one codeword rather than two streams, so encoding, decoding, CRC, interleaving, scrambling, and the associated control signaling are reduced relative to conventional RSMA.","Retransmission reduces to the standard MU-MIMO HARQ process, because the common stream carries no standalone message needing its own HARQ design.","Under finite-alphabet inputs and a fixed decoding-complexity budget, the GMI evaluations show CS-RSMA slightly ahead of conventional RSMA with and without SIC in ergodic sum rate, and effectively equal in max-min fairness.","Link-level simulations with 5G NR LDPC codes put CS-RSMA about 0.05 dB ahead of conventional RSMA at coded BER $10^{-4}$ under the same receiver implementation."],"supporting_citations":[{"why":"Supplies the GMI expression, its approximation, and the gradient for finite-alphabet input with Gaussian-treated interference, which is the paper's performance metric.","marker":"[20]"},{"why":"Provides the SIC-free receiver designs (joint de-mapping and soft SLIC) and low-complexity precoders that CS-RSMA adapts and compares against.","marker":"[5]"},{"why":"Gives the conventional SIC-free RSMA constellation-constrained optimization and baseline that CS-RSMA is measured against.","marker":"[7]"},{"why":"Establishes the RSMA common/private stream framework whose encoding order CS-RSMA reverses.","marker":"[3]"},{"why":"Defines the single-layer RSMA model used in the system setup and the multi-layer generalization mentioned for extension.","marker":"[9]"},{"why":"Documents the common-stream retransmission design problem that CS-RSMA eliminates by avoiding a standalone common codeword.","marker":"[10]"},{"why":"Provides the mismatched-decoding foundation showing that GMI is the achievable rate under the suboptimal decoders considered.","marker":"[19]"},{"why":"Supplies the 5G NR LDPC code used in the link-level simulations that confirm the theoretical ordering.","marker":"[26]"}],"fun_headline_variants":["CS-RSMA: encode first, slice codewords, win sum rate","Segmented codewords lift sum rate, simplify RSMA","Post-encode splitting boosts RSMA sum rate","One codeword per user, sliced, beats two streams"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a user's achievable rate under CS-RSMA is exactly $c_k I_{c,k} + I_{p,k}^{\\mathrm{nonSIC}}$, the sum of its weighted common-segment rate and private-stream rate, even though the decoder draws both from the same received signal; if these two contributions do not add, the reported sum-rate advantage of CS-RSMA shrinks or disappears.","fun_headline_variants_meta":{"raw":{"variants":["CS-RSMA: encode first, slice codewords, win sum rate","Segmented codewords lift sum rate, simplify RSMA","Post-encode splitting boosts RSMA sum rate","One codeword per user, sliced, beats two streams"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1528,"prompt_tokens":1102,"completion_tokens":426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":354}},"tokens_in":718,"tokens_out":426,"duration_ms":4179,"temperature":1.0,"reasoning_tokens":354,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:11:40.927378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a fixed CS-RSMA configuration at a chosen SNR, vary the common-segment fraction $c_k$, and compare the measured LDPC throughput or BER with the prediction $c_k I_{c,k} + I_{p,k}^{\\mathrm{nonSIC}}$ computed from the paper's GMI formulas. If a receiver that jointly decodes the reassembled codeword achieves a rate materially different from that sum, or if the two contributions do not scale with $c_k$, the additive decomposition at the center of the sum-rate comparison is falsified.","supporting_citations":[{"cited_title":"Rate-splitting multiple access: Finite constellations, receiver design, and SIC-free implementation,","cited_arxiv_id":null,"evidence_quote":"Provides the SIC-free receiver designs (joint de-mapping and soft SLIC) and low-complexity precoders that CS-RSMA adapts and compares against."},{"cited_title":"SIC-Free Rate-Splitting Multiple Access: Constellation-Constrained Optimization and Application to Large-Scale Systems","cited_arxiv_id":"2506.12668","evidence_quote":"Gives the conventional SIC-free RSMA constellation-constrained optimization and baseline that CS-RSMA is measured against."},{"cited_title":"Rate-splitting multiple access for downlink communication systems: bridging, generalizing, and outperforming SDMA and NOMA,","cited_arxiv_id":null,"evidence_quote":"Defines the single-layer RSMA model used in the system setup and the multi-layer generalization mentioned for extension."},{"cited_title":"Hybrid automatic repeat request for downlink rate-splitting multiple access,","cited_arxiv_id":null,"evidence_quote":"Documents the common-stream retransmission design problem that CS-RSMA eliminates by avoiding a standalone common codeword."},{"cited_title":"3rd Generation Partnership Project; Technical Specification Group Radio Access Network; NR; Multiplexing and channel coding (Release 17),","cited_arxiv_id":null,"evidence_quote":"Supplies the 5G NR LDPC code used in the link-level simulations that confirm the theoretical ordering."}],"review_version":2}