{"id":"fff837ad-959d-4aad-9a63-6fe2bc99b3df","arxiv_id":"2506.20248","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A DeepRx-based neural receiver outperforms classical iterative receivers when data and pilot symbols are superimposed in 6G uplink multi-user OFDM.","lead":"This paper applies the DeepRx neural receiver to superimposed demodulation reference signals in 6G uplink multi-user OFDM, comparing it against a classical iterative receiver. It reports that DeepRx generally achieves higher throughput with superimposed pilots, and that superimposed pilots can outperform orthogonal pilots at high signal-to-noise ratios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Section V-B2 admits a 4-UE QPSK scenario where the classical SI-DMRS receiver beats DeepRx, contradicting the abstract's 'consistently outperforms' claim.","rationale":"The reader's conditional verdict already flags several limitations; my stress-test isolates the one that is decisive against the strongest claim. The admitted 4-UE QPSK exception is internal evidence, not an external dispute, so it does not depend on reproducing the simulator. It directly contradicts 'consistently outperforms' and 'all scenarios'. The fix is straightforward—qualify the claim—so this does not warrant rejection of the underlying method; it requires revision. I therefore leave the reader's CONDITIONAL verdict unchanged. The OCC assumption is also a real weakness, but it is less decisive because DeepRx is trained on the actual channel and may learn to cope with residual pilot interference; the self-reported exception is unambiguous and load-bearing.","tokens_in":17947,"tokens_out":9744,"duration_ms":104327,"concrete_test":"Run the 4-user QPSK configuration from Table I (4x(1x16), uplink UMa, Sionna) with the Table II SI-DMRS settings (E_itr=0.24, E_DeepRx=0.14, window sizes [(12,14),(6,14)] for both), and plot throughput vs. SNR for DeepRx-SI and iterative-SI over the -12 to -2 dB range. If the iterative curve exceeds DeepRx at any SNR, revise the abstract to 'in most tested scenarios' and add a caveat. If DeepRx is never below, correct the Section V-B2 sentence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the paper's own counterexample to its headline claim. The abstract and Section I claim DeepRx 'consistently outperforms' classical receivers in all SI-DMRS scenarios. Section V-B2 reports, for MU-MIMO, that 'the only scenario where the conventional system achieves higher performance with SI DMRS is the case with QPSK and 4 UEs' (Figure 8, top). The previous sentence says the conventional baseline 'falls short of the throughput provided by the ML-based DeepRx schemes, especially with SI DMRS', so the exception is a comparison to DeepRx, not to the orthogonal-DMRS conventional receiver. The authors explain it by QPSK robustness, which makes it a systematic regime rather than a seed-dependent fluctuation. Thus the abstract's 'consistently' is not supported by the paper's own results. The OCC separability issue in Section III is a separate concern; this direct exception alone requires the claim to be weakened or the scenario to be explicitly excluded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies superimposed DMRS (SI-DMRS) transmission for uplink OFDM in single- and multi-user MIMO. The authors propose an enhanced DeepRx CNN receiver that takes SI-DMRS signals and pilot configurations as input, uses an LS channel estimate and a Channel-estimation CNN, then performs equalization and demapping through Detector and Demapper CNNs trained end-to-end with a BCE loss. They also design a classical LS-based iterative receiver with sliding-window smoothing and LDPC decoding in the loop. The two receiver families are compared against orthogonal 5G DMRS and against genie-aided linear detection in Sionna simulations with UMa channels, 1/2/4 users, 1/2 transmit antennas, 4/16 receive antennas, and QPSK/16QAM/64QAM. The central claims are that DeepRx consistently outperforms the classical receivers with SI-DMRS and that DeepRx with SI-DMRS yields throughput gains over DeepRx with orthogonal DMRS at high SNR.","tokens_in":18167,"tokens_out":7831,"duration_ms":79441,"significance":"If the reported results hold, the paper would provide a useful data point for the 6G discussion on AI/ML-based receivers and superimposed reference signals, showing that a neural receiver can remove pilot overhead while retaining reliable channel estimation. The main strengths are the reproducible simulation setup built on Sionna, the detailed reporting of architecture and scenario parameters in Tables I-III, and the straightforward LS/iterative formulation in Section IV, which is internally consistent. The paper also reports a depth ablation (Figure 9). However, the empirical comparison currently overclaims relative to its own results, and the OCC-based orthogonality argument and tuning fairness need to be resolved before the central comparison can be considered established.","major_comments":[{"comment":"The abstract and introduction claim that DeepRx 'consistently outperforms' the classical receivers in all SI-DMRS scenarios, but Section V-B2 reports that 'the only scenario where the conventional system achieves higher performance with SI DMRS is the case with QPSK and 4 UEs' (Figure 8, top). Because the authors attribute this outcome to the robustness of QPSK, it appears to be a systematic regime rather than a statistical fluctuation. The headline claim should be weakened to 'in the majority of tested scenarios,' or the QPSK/4-UE scenario should be explicitly excluded from the stated scope.","section":"Abstract and Section I vs. Section V-B2"},{"comment":"The orthogonality argument for OCC-based SI DMRS is incomplete. The statement that 'the inner product of two OCC sequences is zero' ignores the channel response: the received contribution of layer ell at RE (i,j) is H_{i,j}^{(ell)} p_{i,j}^{(ell)}, so after correlation with p^{(k)} the residual inter-layer term is sum_{i,j} H_{i,j}^{(ell)} p_{i,j}^{(ell)} (p_{i,j}^{(k)})^*, which is not zero unless the channel is constant over the OCC spreading block. Since the OCCs are spread over the entire time-frequency grid and the channel is frequency-selective (UMa, 3.5 GHz, 72 subcarriers), the claimed 'nearly eliminated' inter-layer pilot interference is not justified. Please specify the OCC group size relative to the channel coherence block or numerically evaluate the residual cross-layer interference.","section":"Section III"},{"comment":"The comparison is asymmetric in tuning effort. The text states that 'to achieve optimal performance from the iterative receiver, we ... conducted an extensive search for the optimal SI DMRS power ratio E, number of iterations, and sliding window sizes,' while for DeepRx no equivalent search is reported; the DeepRx power ratios in Table II appear fixed per scenario, and the only DeepRx hyperparameter studied is depth (Section V-C). If the classical baseline is individually optimized and the ML receiver is not, the conclusion that 'DeepRx consistently outperforms' may partly reflect unequal optimization. Please state how the DeepRx E values were selected and provide a sensitivity analysis over E (and ideally over window-size inputs to the channel-estimation CNN) for DeepRx.","section":"Section V-A and Table II"},{"comment":"All throughput and BER curves are reported without error bars, random seeds, or the number of independent channel realizations used for evaluation. The paper states that training samples were generated on-the-fly, but this does not quantify evaluation uncertainty. Given that the central claim is a comparative performance statement, confidence intervals or multiple-seed results are needed to establish that the observed gaps are not due to finite-sample noise, particularly in the QPSK/4-UE case where the conventional receiver is reported to win.","section":"Section V and Figures 5-8"}],"minor_comments":[{"comment":"Several window-size tuples in Table II are missing closing brackets (e.g., '(2,14]' in the 64-QAM rows); please correct these typographical errors.","section":"Table II"},{"comment":"The definition of v_{i,j}^{(k,u-1)} in Eq. (9) is hard to parse because the same symbol is used for a sum over all users and then a second sum over k' not equal to k; consider introducing separate names for these two interference terms.","section":"Section IV-A2"},{"comment":"The explanation that the QPSK/4-UE exception is 'most likely due to the high robustness of QPSK' is speculative; please replace it with a quantitative argument, such as the ratio of the residual pilot-interference floor to the decision-region size.","section":"Section V-B2"},{"comment":"The observation that DeepRx with orthogonal DMRS can outperform a linear genie-aided receiver with perfect channel knowledge is surprising and would benefit from an explicit explanation of the nonlinear detection gain earlier in the section, not only in the final paragraph.","section":"Section V-B1"},{"comment":"The conclusion claims 'keeping inference complexity within practical bounds,' but no complexity measurements are reported; either add runtime or FLOPs comparisons or soften this statement.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"This paper is within the scope of eess.SP and the results are of interest for AI/ML physical-layer work. The main issues for the editor are that the abstract's 'consistently outperforms' claim is contradicted by the paper's own QPSK/4-UE exception in Section V-B2, and that the OCC orthogonality argument in Section III needs a coherence-aware justification. The authors are affiliated with Nokia and several prior DeepRx papers are self-cited; this is not disqualifying, but the novelty relative to [20] and [21] should be sharpened in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about pilot overhead in 6G uplink. The authors extend DeepRx to handle superimposed DMRS in multi-user MIMO, and compare it against a classically processed iterative receiver. The simulations cover a sensible range of MIMO sizes and modulations, and the throughput curves are clear. The headline is that DeepRx makes SI DMRS viable and beats the classical baseline.\n\nWhat's actually new: applying OCC-based SI DMRS to the multi-user uplink and adapting the DeepRx channel estimation block accordingly. The classical iterative receiver with LDPC in the loop is also a reasonable baseline.\n\nThe soft spots are real. The abstract and introduction claim DeepRx 'consistently outperforms' classical receivers with SI DMRS. Section V-B2 admits the one exception: 4 UEs with QPSK, where the conventional receiver wins. That's not a footnote; it's a systematic regime, not a seed anomaly. So the headline claim needs to be weakened.\n\nSecond, the OCC design: the paper says the inner product of two OCC sequences is zero and therefore inter-layer pilot interference is eliminated. That's only true if the channel is constant across the OCC-combining block. Here the OCCs are spread over the whole time-frequency grid, where the channel varies significantly. The LS estimator and smoothing might mitigate this, and the simulations are done with real channel variation, so the results may be fine, but the theoretical justification in Section III is incomplete. It should be rewritten.\n\nOne point in the paper's favor: the classical baseline is extensively tuned per scenario (power ratio, window sizes, iterations), while DeepRx uses one fixed depth and hyperparameter set. So if anything, the comparison is conservative for DeepRx. That's not a flaw, though the paper should state it more explicitly.\n\nNo code or data release, no error bars, which is standard for this venue but limits the ability to reproduce. The citation pattern looks fair, and the prior SI-pilot work is cited.\n\nWho's it for: people working on 6G physical layer design, specifically AI/ML receivers and reference signal overhead reduction. It's a useful data point, not a breakthrough. A serious referee should see it, with the expectation that the claims be tightened.\n\nMy recommendation: send it to review, but tell the authors to fix the internal contradiction and address the OCC/channel-selectivity point. The simulation study itself seems solid enough to deserve referee time.","headline":"A solid simulation comparison with a useful result, but the 'consistently outperforms' claim is contradicted by the paper's own 4-UE QPSK scenario, and the OCC orthogonality argument ignores channel selectivity.","tokens_in":18708,"tokens_out":5001,"would_cite":false,"duration_ms":48229,"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":"A convolutional receiver can estimate channels from superimposed pilots and beat orthogonal pilots at high SNR.","keywords":["superimposed DMRS","orthogonal cover codes","DeepRx","neural receiver","uplink MU-MIMO","OFDM","channel estimation","6G"],"falsifier":"Re-run the DeepRx-versus-classical comparison on a channel with a large delay spread, or with shorter OCC groups, so the channel coefficient changes inside each OCC group; if the SI DMRS throughput advantage over orthogonal DMRS disappears or reverses, the OCC separability assumption is the cause.","tokens_in":17782,"feed_emoji":"📡","tokens_out":6822,"duration_ms":69293,"temperature":0.7,"pith_summary":"This paper argues that a 6G uplink can send data on every time-frequency resource element by superimposing low-power reference signals on top of data, rather than reserving separate pilot resources, and that a learned convolutional receiver can make that scheme pay off. The authors configure the superimposed DMRS with orthogonal cover codes so the pilot components of different users remain separable, then compare a DeepRx neural receiver against a classical iterative receiver that cancels data-to-pilot interference. In simulated single-user and multi-user MIMO uplinks, DeepRx outperforms the classical receiver in every superimposed-DMRS scenario tested, and at high signal-to-noise ratio the superimposed scheme with DeepRx delivers higher throughput than orthogonal 5G-style DMRS. If the result carries into practice, the pilot overhead that currently grows with the number of users and layers could be largely eliminated.","feed_headline":"Neural receiver turns overlapped pilots into a 6G throughput win","feed_subtitle":"DeepRx recovers data from pilot-data overlaps, outperforming classical receivers and cutting pilot overhead in uplink OFDM.","key_machinery":"The carrying mechanism is a code-domain pilot plan combined with a hybrid classical/neural receiver. Each user or layer is assigned an orthogonal cover code (OCC) sequence spread over the entire time-frequency grid; because the inner product of two OCC sequences is zero, the superimposed pilot components are nominally orthogonal even though data symbols occupy the same resource elements. DeepRx starts from a least-squares channel estimate, refines it with a convolutional ResNet, equalizes through parallel LMMSE-type and maximum-ratio-combining branches, and maps the refined symbol estimates to bit log-likelihood ratios with a second ResNet, training the whole pipeline end-to-end with a binary cross-entropy loss.","core_discovery":"The paper's central claim is that a CNN-based receiver trained end-to-end can estimate the channel from superimposed DMRS and detect data reliably enough that the superposition becomes a net throughput win. DeepRx beats the classical iterative receiver in all tested SI DMRS configurations, and SI DMRS with DeepRx provides higher throughput than orthogonal DMRS at high SNR across single-stream, two-stream, two-user, and four-user uplink cases. The authors attribute the gain to DeepRx learning to treat residual data-to-pilot interference as structured information rather than noise, while the classical receiver's channel estimates degrade under higher-order modulation.","pith_inferences":["Because the paper's orthogonality argument uses the inner product of the cover codes without accounting for the channel response multiplying each OCC symbol, the gains should be re-tested in more frequency-selective channels or with shorter OCC groups; if the channel varies inside a group, inter-layer pilot interference could return.","The same superposition could be combined with per-link adaptation: choose orthogonal DMRS at low SNR, SI DMRS at high SNR, and optimize the pilot-to-data power ratio per modulation order rather than fixing it per scenario.","The conclusion's observation that control channels occupy the first OFDM symbols suggests SI DMRS would be especially valuable when few symbols remain for data; a system-level study could quantify this benefit."],"forward_implications":["At high signal-to-noise ratio, SI DMRS with DeepRx yields higher throughput than orthogonal DMRS in the tested configurations, so removing dedicated pilot resources can translate directly into spectral efficiency gains when channel conditions are good.","DeepRx outperforms the classical iterative receiver in every SI DMRS scenario tested, indicating that learned receivers handle data-to-pilot interference better than hand-designed cancellation loops.","At low SNR, orthogonal DMRS remains the better choice even with DeepRx, so a practical system would likely switch between the two schemes based on operating point.","The classical iterative receiver with SI DMRS is competitive only under QPSK; with 16-QAM and 64-QAM its channel estimates are too imprecise, so the viability of superimposed pilots depends on having a sophisticated receiver.","A shallower DeepRx variant with fewer ResNet blocks reaches nearly the same throughput, so the reported gains do not require the largest model."],"supporting_citations":[{"why":"Defines the DeepRx architecture and the binary cross-entropy training loss that this paper adapts to superimposed DMRS.","marker":"[8]"},{"why":"Introduces superimposed pilot transmission with a neural receiver, the direct precursor this work extends to multi-user uplink.","marker":"[11]"},{"why":"Proposes an ML-driven interference-cancellation receiver for superimposed pilots in MIMO-OFDM, the closest classical/ML baseline family.","marker":"[20]"},{"why":"Presents an interference-cancellation neural receiver for superimposed pilots in multi-layer transmission, another nearby approach.","marker":"[21]"},{"why":"Shows spectral-efficiency gains from pilotless spatial multiplexing, which motivates the DMRS-free upper bound used here.","marker":"[14]"},{"why":"Provides the link-level simulation environment used for all numerical results.","marker":"[25]"}],"fun_headline_variants":["AI receiver beats classical for 6G overlapped pilots","Neural network receiver wins over classical in 6G uplink","DeepRx outperforms conventional receivers with superimposed DMRS","CNN receiver edges classical for pilot-data overlap in 6G","AI turns overlapped pilots into throughput advantage for 6G"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes the channel is flat enough across each OCC group that the orthogonal cover codes keep the superimposed pilots separable; if the channel response varies within a group, inter-layer pilot interference returns and the reported gains may shrink.","fun_headline_variants_meta":{"raw":{"variants":["AI receiver beats classical for 6G overlapped pilots","Neural network receiver wins over classical in 6G uplink","DeepRx outperforms conventional receivers with superimposed DMRS","CNN receiver edges classical for pilot-data overlap in 6G","AI turns overlapped pilots into throughput advantage for 6G"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000314,"raw_usage":{"total_tokens":1755,"prompt_tokens":892,"completion_tokens":863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":779}},"tokens_in":508,"tokens_out":863,"duration_ms":9044,"temperature":1.0,"reasoning_tokens":779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:53:26.563222+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the DeepRx-versus-classical comparison on a channel with a large delay spread, or with shorter OCC groups, so the channel coefficient changes inside each OCC group; if the SI DMRS throughput advantage over orthogonal DMRS disappears or reverses, the OCC separability assumption is the cause.","supporting_citations":[{"cited_title":"DeepRx: Fully convolu- tional deep learning receiver,","cited_arxiv_id":null,"evidence_quote":"Defines the DeepRx architecture and the binary cross-entropy training loss that this paper adapts to superimposed DMRS."},{"cited_title":"End-to-End Learning for OFDM: From Neural Receivers to Pilotless Communication,","cited_arxiv_id":null,"evidence_quote":"Introduces superimposed pilot transmission with a neural receiver, the direct precursor this work extends to multi-user uplink."},{"cited_title":"AI-Driven Iterative Receiver for Superimposed Pilot Schemes in MIMO-OFDM Systems,","cited_arxiv_id":null,"evidence_quote":"Proposes an ML-driven interference-cancellation receiver for superimposed pilots in MIMO-OFDM, the closest classical/ML baseline family."},{"cited_title":"Interference cancellation based neural receiver for superimposed pilot in multi-layer transmission,","cited_arxiv_id":null,"evidence_quote":"Presents an interference-cancellation neural receiver for superimposed pilots in multi-layer transmission, another nearby approach."},{"cited_title":"Deep learning-based pilotless spatial multiplexing,","cited_arxiv_id":null,"evidence_quote":"Shows spectral-efficiency gains from pilotless spatial multiplexing, which motivates the DMRS-free upper bound used here."},{"cited_title":"Sionna: An Open-Source Library for Next-Generation Physical Layer Research,","cited_arxiv_id":null,"evidence_quote":"Provides the link-level simulation environment used for all numerical results."}],"review_version":1}