{"id":"30adc993-48fc-4dae-af3f-5cc5932d8d4e","arxiv_id":"2501.08871","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Graph neural networks operating on channel factor graphs and code Tanner graphs achieve near-optimal detection and joint detection and decoding on ISI channels, outperforming feasible classical baselines.","lead":"The authors build receivers from graph neural networks that detect symbols and decode LDPC codes together over channels with inter-symbol interference. In simulations, the GNN detectors match the optimal BCJR algorithm for BPSK and beat feasible baselines by about 6 dB for high-rate 16-QAM turbo detection.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The TDD/EXIT claims assume the GNN's total LLRs decompose as extrinsic plus a priori (Eq. 5), but the nonlinear GNN admits no such subtraction; the paper's own Fig.","rationale":"Read in good faith, the paper is a solid empirical study: the detection results (Fig. 4) show GNNs matching MAP for BPSK and beating LMMSE/EP at high rates; the JDD flooding results (Figs. 9-10) are consistent with end-to-end learning; the appendix shows stable training. The central unsupported link is the TDD rate/BER claim, which depends on the validity of extrinsic information exchange in a learned nonlinear component. The paper itself exposes the weakness in Sec. IV-C: the EXIT trajectory diverges after three turbo iterations, and the authors state that Eq. 5 does not hold for GNNs when a priori information dominates. That is a correctness risk internal to the argument, not a disagreement with consensus. The Gaussian-prior training mismatch noted by the reader is real but secondary; even with Gaussian a priori LLRs, a nonlinear detector cannot be decomposed by subtraction. A proper extrinsic GNN either needs to be trained to output extrinsic quantities directly or the TDD loop should use a different feedback rule. Without such a check, the headline 6.25 dB gain is not yet established as a genuine detection gain. The reader's conditional verdict is appropriate; this concern strengthens the conditions rather than overturning the paper. If the proposed test shows (i) and (iii) match, the concern is resolved and the TDD claim stands.","tokens_in":23856,"tokens_out":5954,"duration_ms":63920,"concrete_test":"Re-run the Fig. 13 TDD simulation (N=4608, rate 8/9, 16-QAM, Proakis-C) under three detector-output rules: (i) the paper's ℓE = ℓT − ℓA; (ii) direct use of ℓT without subtraction; (iii) a GNN retrained with an explicit extrinsic objective, e.g., minimizing BCE while including a gradient-reversal or adversarial penalty that makes the output statistically independent of ℓA. Compare BER versus Eb/N0 and the EXIT trajectory for (i) versus (iii). If the gap exceeds roughly 1 dB at the operating SNR, or if (i) violates the MAP-constrained TDD rate bound, the claimed gain depends on invalid extrinsic subtraction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest empirical claim is the ~6.25 dB TDD gain over EP (Fig. 13) and the EXIT-area rate predictions (Eq. 8, Fig. 12). These rest on the standard turbo assumption that a component's total LLR is the sum of its extrinsic and a priori parts, so the extrinsic LLRs fed to the SPA decoder can be computed by subtraction, ℓE = ℓT − ℓA (Eq. 5). The GNN detector is a nonlinear function of the a priori LLRs injected at the VN embedding (Sec. IV-B, Algorithm 3); there is no reason its output decomposes additively. The authors acknowledge this in Sec. IV-C: 'equation (5) does not hold for GNNs, if RSDD ≪ IA. In this case the a priori LLRs dominate inference and the loops lead to too significant amplification.' Fig. 12 shows the actual TDD trajectory diverging from the Gaussian-prior transfer characteristic after three outer iterations. If the detector's output contains residual a priori information, the turbo loop double-counts information, producing overconfident LLRs. The 6.25 dB gain over EP could then reflect this self-amplification rather than a genuine detection advantage, and it would not generalize to other codes, rates, or SNRs. The Gaussian-vs-non-Gaussian prior mismatch flagged by the reader is a symptom of this deeper issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes graph neural network (GNN) detectors built on Forney and Ungerboeck factor graphs for intersymbol-interference channels, together with channel-agnostic input embeddings, a joint detection and decoding architecture with shared variable nodes, and a flooding schedule for reduced latency. The detection results claim near-MAP performance for BPSK on the Proakis-C channel and gains over feasible classical and neural baselines for 16-QAM and 64-QAM. The joint detection and decoding results claim that a flooding GNN outperforms a turbo BCJR-SPA receiver at lower latency, and the turbo detection and decoding results claim a gain of more than 6 dB over the best feasible non-neural baseline in a high-rate 16-QAM scenario.","tokens_in":24169,"tokens_out":9185,"duration_ms":96657,"significance":"If the results hold, the paper makes a useful contribution by showing that GNN message passing on sparse factor graphs can bridge part of the gap between sum-product detection and optimal MAP detection, while providing flexibility for joint receiver optimization. The study is extensive, with comparisons to BCJR, SPA, N-SPA, CNN, FGNN, EP, and LMMSE baselines, an analysis of learned messages, operation-count estimates, and explicit robustness experiments under CSI uncertainty. I find no circularity in the detection evaluation: the GNN weights are trained parameters, and the rate/BER results are measured from trained models. The main weakness is the turbo detection and decoding claim, which rests on the assumption that the nonlinear GNN's total LLRs can be turned into extrinsic LLRs by subtracting the a priori LLRs; the paper itself acknowledges that this assumption fails in the relevant regime.","major_comments":[{"comment":"The TDD receiver computes the extrinsic LLRs fed to the SPA decoder as ℓE = ℓT − ℓA (Eq. (5)). This decomposition is exact only for components whose total LLR is additively separable in the a priori LLR; the GNN is a nonlinear function of the a priori LLRs injected at the VN embedding (Algorithm 3), so additivity does not follow. The paper itself concedes in Sec. IV-C that 'equation (5) does not hold for GNNs' when RSDD ≪ IA and that the loops lead to 'too significant amplification'; Fig. 12 further shows the turbo trajectory diverging after three outer iterations. Since the headline 6.25 dB gain over EP in Fig. 13 (and the abstract) and the RBMI_TDD curves in Fig. 11 are computed from these extrinsic messages, the gain may be an artifact of double-counted a priori information rather than a genuine detection advantage. Please re-evaluate the TDD scheme with a genuinely extrinsic GNN output (e.g., training against actual decoder feedback or using a separate a priori input with a loss that forces extrinsic behavior) and report the resulting BER and EXIT trajectories.","section":"Sec. IV-B, IV-C; Eq. (5); Figs. 12-13"},{"comment":"The Fig. 12 caption states that 'The GNN trajectory matches the prediction from the transfer characteristic,' but the body text states that the high-rate trajectory does not match precisely and that the low-rate case diverges after three turbo iterations because the a priori information from the GNN 'might be not extrinsic.' This contradiction, combined with the fact that Algorithm 3 trains the TDD detector only on Gaussian priors drawn from Eq. (6), means the EXIT-area rate predictions of Eq. (8) and Fig. 11 are not validated for the actual non-Gaussian decoder feedback. Please either provide matching trajectories under the actual turbo loop or explicitly label the Fig. 11 curves as approximate or optimistic bounds.","section":"Sec. IV-C; Fig. 12 and Algorithm 3"},{"comment":"The baseline description states that training of N-SPA does not converge for NIt > 7, yet the legends in Fig. 4 show N-SPA with NIt = 8. If the N-SPA curve is produced by an unconverged or differently configured model, it is not a fair baseline, and the claim that the GNN outperforms the other NN-based baselines is not fully established. Please correct the number of iterations or report the best converged N-SPA configuration together with its training status.","section":"Sec. III-E; Fig. 4"},{"comment":"The main empirical claims are presented without confidence intervals or multiple-seed results. The appendix demonstrates training-loss stability for JDD, but it does not quantify the variability of BER or BMI across random initializations and test realizations for the headline figures. Please add error bars or multiple-seed curves for the key comparisons (SDD rates, JDD BER, and TDD BER) or, alternatively, provide a reproducibility package (code and data) so that the comparisons can be independently verified.","section":"Secs. III-E and IV-C; Figs. 4, 9, 11, 13"}],"minor_comments":[{"comment":"There are several typos, e.g., 'the GNN is build upon' in the abstract; please copy-edit the manuscript.","section":"Abstract and Sec. I"},{"comment":"The y-axis of Fig. 4(b) is labeled 'test' in the manuscript; this appears to be a placeholder and should be RBMI_SDD.","section":"Fig. 4(b)"},{"comment":"The caption describes the plotted quantity as 'BER of the bits u' while the y-axis label reads 'BLER.' Please clarify which metric is shown and keep the two consistent.","section":"Fig. 9"},{"comment":"The derivation comparing optimal messages with SPA messages contains unbalanced parentheses and is hard to follow; please rewrite it in a clearer step-by-step form.","section":"Sec. III-F"},{"comment":"The algorithm sets LBCE but then updates with SGD(LMulti, θ); LMulti is not defined in Algorithm 3. Please align the notation with Eq. (12).","section":"Algorithm 3"},{"comment":"The phrase 'for QPSK (with memory L = 6)' should read 'with channel memory L = 6'; QPSK itself has no memory.","section":"Sec. III-E, QPSK setup"},{"comment":"References [16] and [44] refer to the same paper by Colavolpe and Germi; please merge or distinguish them appropriately.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The SDD and JDD material is solid and worth publishing once the extrinsic-LLR issue for TDD is resolved. The main risk is that the 6.25 dB TDD gain is a double-counting artifact; the authors themselves provide the admission needed to motivate a re-evaluation. For a simulation-heavy paper with no code release, adding multiple-seed results and confidence intervals would also materially strengthen the reproducibility of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a thorough, honest empirical study of GNN-based detection for ISI channels, and the detector-only results look credible. The new pieces are the application of the GNN message-passing framework from [24] to FFG/UFG detection, the three channel-agnostic embeddings, and the joint detection-decoding graph with a flooding schedule. The BPSK results matching MAP detection and the 16-QAM improvements over feasible baselines are supported by consistent rate curves, ablations, latency counts, and an operation-count table. That is real work, and the authors are careful about many details, including training stability and CSI robustness.\n\nBut the strongest headline—the 6.25 dB TDD gain over EP—needs a heavy caveat. The turbo extrinsic LLRs are computed with Eq. (5), ℓE = ℓT − ℓA. That subtraction is valid for APP detectors, but the GNN is a nonlinear function of the a priori LLRs; there is no reason its total output decomposes additively. The paper itself says in Sec. IV-C that Eq. (5) does not hold for GNNs when RS D D is small relative to IA, and Fig. 12 shows the actual TDD trajectory diverging from the Gaussian-prior transfer characteristic after three iterations. That is exactly the regime where the 6.25 dB gain appears. So the gain may be partly self-amplification from double-counted a priori information, not a robust detection advantage. The stress-test concern lands, and the authors’ own text supports it.\n\nOther soft spots are standard but worth naming: no confidence intervals or multiple-seed curves in the main figures, no released code or data, and the 16/64-QAM BCJR baseline is a 4/8-PAM surrogate. The surrogate is reasonable given complexity, but it should be labeled as a surrogate in the abstract and conclusion, not implicitly treated as the true MAP curve. These issues weaken reproducibility but not the detector-only contribution.\n\nWho is this for? Researchers working on ML-assisted receivers, GNN decoding, and turbo equalization. There are no theoretical proofs, but the empirical evidence is substantial enough to merit serious referee time. I would send it to peer review, and the main requests would be code/data release, confidence intervals, and a reframed TDD claim that separates detector-only gains from turbo-loop gains.","headline":"Solid detector-only GNN results, but the 6.25 dB TDD headline is shaky because the extrinsic-LLR subtraction in Eq. (5) does not hold for a nonlinear GNN—the authors admit this in Sec. IV-C.","tokens_in":24720,"tokens_out":2616,"would_cite":true,"duration_ms":30324,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph neural network whose learned message functions replace sum-product updates on channel factor graphs matches optimal MAP detection on the severe Proakis-C channel, and a joint detection-and-decoding variant with a flooding schedule…","keywords":["graph neural networks","joint detection and decoding","inter-symbol interference","factor graphs","turbo detection and decoding","BCJR equalization","LDPC codes","EXIT charts"],"falsifier":"Run the high-rate 16-QAM TDD setup of Fig. 13 but feed the GNN detector the actual extrinsic LLRs of the SPA decoder instead of the Gaussian priors of equation (6), and compare the EXIT trajectory to the transfer characteristic. If the trajectory diverges before the decoder converges, or if the BER gain over EP drops below 1 dB, the central generalization claim is falsified.","tokens_in":23643,"feed_emoji":"📡","tokens_out":7619,"duration_ms":69803,"temperature":0.7,"pith_summary":"Over channels that smear transmitted symbols into each other, the optimal detector is often too complex and standard message-passing equalization is unreliable because the underlying factor graphs contain short cycles. This paper claims that replacing sum-product messages with learned multilayer-perceptron messages on a graph neural network built from the channel's factor graph closes much of that gap: for BPSK it matches the optimal MAP detector on the severe Proakis-C channel, and for 16-QAM it beats all classical and neural baselines. The same GNN can be fused with an LDPC decoder by sharing variable nodes, giving a joint detection-and-decoding receiver whose parallel flooding schedule outperforms turbo BCJR-SPA at a fraction of the latency. In a high-rate 16-QAM turbo setup the paper reports a gain of about 6.25 dB over the best feasible non-neural baseline. If these results hold when the decoder's feedback is not Gaussian, they point toward a fully learned receiver that is optimized end-to-end rather than component by component.","feed_headline":"Neural receiver matches optimal detection, beats turbo equalizers","feed_subtitle":"Learned message passing on channel factor graphs fixes sum-product cycle failures and gains 6.25 dB in high-rate 16-QAM.","key_machinery":"The central object is a factor-graph GNN in which every variable node, factor node, and directed edge carries a learned d-dimensional state updated by a shared MLP; the edge attributes encode which tap of the CIR an edge corresponds to, and the messages are updated in a four-phase loop of factor-to-variable, variable, variable-to-factor, and factor updates. For joint detection and decoding, the detection factor graph and the code's Tanner graph are joined at the variable nodes, and a flooding schedule updates both factor-node sets in parallel. This machinery replaces SPA's fixed max-star computations with learned feature vectors, providing adaptive damping and the ability to encode inter-observation dependencies that SPA omits.","core_discovery":"The paper's central claim is that a graph neural network formed by attaching a small multilayer perceptron to every variable node, factor node, and directed edge of a Forney or Ungerboeck factor graph can learn message-passing rules that are substantially more robust than sum-product message passing on the same graph. Because the MLPs are shared across nodes of the same type and edge attributes encode which tap of the channel impulse response an edge corresponds to, the network respects the channel's shift-invariant structure and stays sparse, which is what lets it scale to block lengths and modulation orders where BCJR detection is computationally infeasible. When the same message-passing framework is applied to the Tanner graph of a 5G LDPC code and the two graphs share variable nodes, the combined network performs joint detection and decoding, and updating all factor nodes in parallel—the flooding schedule—improves both error rate and latency. For the severe Proakis-C channel, the paper reports that the GNN matches MAP performance for BPSK, outperforms all baselines for 16-QAM, and in the high-rate TDD scenario beats the best feasible classical baseline by about 6.25 dB.","pith_inferences":["Because the learned edge messages are not tied to the linear Gaussian likelihood structure, the same architecture may transfer to nonlinear ISI or faster-than-Nyquist signaling; the paper only evaluates linear AWGN ISI channels.","The 6.25 dB gain is measured against deliberately over-parameterized neural networks and feasible classical baselines; under hardware constraints such as pruning and fixed-point quantization the margin may shrink, a comparison the paper leaves for implementation-level work.","The Gaussian-prior training for TDD is the main fragility; a natural next experiment is to train with decoder-produced LLRs or add a cycle-loss penalty, which could remove the EXIT divergence observed after three turbo iterations."],"forward_implications":["GNN detection can serve as a near-MAP detector for severe ISI channels where BCJR is infeasible because the number of states grows exponentially with memory and modulation order.","The robustness to CSI uncertainty demonstrated for the GNN means receivers built this way can operate with imperfect channel estimates, where classical APP-based detectors degrade sharply.","The flooding schedule shows that exchanging information between detection and decoding in every iteration, rather than waiting for full decoder runs, improves both BER and latency for short block lengths.","In high-rate turbo detection and decoding, the GNN detector provides a gain of about 6.25 dB over the best feasible classical baseline, making learned detection attractive in the regime where BCJR is infeasible and LMMSE-based methods are suboptimal."],"supporting_citations":[{"why":"Provides the bipartite GNN message-passing framework with shared MLPs for nodes and edges that the paper adapts to detection and joint decoding.","marker":"[24]"},{"why":"Introduces the Forney factor graph of the ISI channel that serves as the base graph for the detection GNN.","marker":"[13]"},{"why":"Provides the Ungerboeck factorization used as the alternative sparse factor graph for detection.","marker":"[14]"},{"why":"Documents the suboptimality of sum-product detection on ISI factor graphs, the problem the GNN aims to fix.","marker":"[16]"},{"why":"Supplies the linear-complexity SPA on the UFG used as a baseline and as a complexity reference.","marker":"[17]"},{"why":"Defines the BCJR algorithm, the optimal MAP detector used as the performance ceiling and a TDD/JDD baseline.","marker":"[11]"},{"why":"Supplies the neural SPA baseline with trainable damping weights that the GNN outperforms and whose operation counts are compared.","marker":"[19]"},{"why":"Provides the CNN equalizer baseline used throughout the detection and TDD comparisons.","marker":"[6]"},{"why":"Gives the expectation-propagation equalizer, the best feasible non-neural baseline that the high-rate TDD GNN beats by about 6 dB.","marker":"[45]"}],"fun_headline_variants":["GNN receiver matches MAP, beats turbo equalizers","Graph neural network achieves robust joint detection and decoding","Deep GNN equalizer gains 6.25 dB in high-rate turbo decoding","GNN message passing on factor graphs beats classical ISI baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned message functions, trained on Gaussian-distributed a priori LLRs and a fixed Proakis-C training setup, remain accurate when the decoder feeds back real, non-Gaussian LLRs and when the channel changes.","fun_headline_variants_meta":{"raw":{"variants":["GNN receiver matches MAP, beats turbo equalizers","Graph neural network achieves robust joint detection and decoding","Deep GNN equalizer gains 6.25 dB in high-rate turbo decoding","GNN message passing on factor graphs beats classical ISI baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001102,"raw_usage":{"total_tokens":4666,"prompt_tokens":1083,"completion_tokens":3583,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":3513}},"tokens_in":699,"tokens_out":3583,"duration_ms":24898,"temperature":1.0,"reasoning_tokens":3513,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:15:44.161558+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the high-rate 16-QAM TDD setup of Fig. 13 but feed the GNN detector the actual extrinsic LLRs of the SPA decoder instead of the Gaussian priors of equation (6), and compare the EXIT trajectory to the transfer characteristic. If the trajectory diverges before the decoder converges, or if the BER gain over EP drops below 1 dB, the central generalization claim is falsified.","supporting_citations":[{"cited_title":"Graph Neural Networks for Channel Decoding,","cited_arxiv_id":null,"evidence_quote":"Provides the bipartite GNN message-passing framework with shared MLPs for nodes and edges that the paper adapts to detection and joint decoding."},{"cited_title":"Lower Bounds on Error Probability in the Presence of Large Intersymbol Interference,","cited_arxiv_id":null,"evidence_quote":"Introduces the Forney factor graph of the ISI channel that serves as the base graph for the detection GNN."},{"cited_title":"Adaptive Maximum-Likelihood Receiver for Carrier- Modulated Data-Transmission Systems,","cited_arxiv_id":null,"evidence_quote":"Provides the Ungerboeck factorization used as the alternative sparse factor graph for detection."},{"cited_title":"On the application of factor graphs and the sum-product algorithm to ISI channels,","cited_arxiv_id":null,"evidence_quote":"Documents the suboptimality of sum-product detection on ISI factor graphs, the problem the GNN aims to fix."},{"cited_title":"SISO Detection Over Linear Channels With Linear Complexity in the Number of Interferers,","cited_arxiv_id":null,"evidence_quote":"Supplies the linear-complexity SPA on the UFG used as a baseline and as a complexity reference."},{"cited_title":"Optimal decoding of linear codes for minimizing symbol error rate (corresp.),","cited_arxiv_id":null,"evidence_quote":"Defines the BCJR algorithm, the optimal MAP detector used as the performance ceiling and a TDD/JDD baseline."},{"cited_title":"Low-Complexity Near-Optimum Symbol Detection Based on Neural Enhancement of Factor Graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the neural SPA baseline with trainable damping weights that the GNN outperforms and whose operation counts are compared."},{"cited_title":"Extrinsic Neural Network Equalizer for Channels with High Inter-Symbol-Interference,","cited_arxiv_id":null,"evidence_quote":"Provides the CNN equalizer baseline used throughout the detection and TDD comparisons."},{"cited_title":"Block expectation propagation equalization for ISI channels,","cited_arxiv_id":null,"evidence_quote":"Gives the expectation-propagation equalizer, the best feasible non-neural baseline that the high-rate TDD GNN beats by about 6 dB."}],"review_version":1}