{"id":"28e2d855-5088-4546-87a0-5e7b159ab44c","arxiv_id":"2507.19513","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A dual-path network using sLSTM for temporal modeling and Conv3D for spatial features beats ConvLSTM-based STN on cellular traffic forecasting and on one unseen-region transfer test.","lead":"The paper replaces the ConvLSTM branch of an existing spatiotemporal network with a scalar LSTM (sLSTM) and adds transformer-based fusion to forecast 5G cellular traffic. On the Italia Telecom dataset, the new model reports lower errors than ConvLSTM baselines and transfers to an unseen region, but the comparison set is narrow and no code is released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Trentino generalization claim is unverifiable as reported: the paper never states whether unseen-region normalization uses Trentino's own per-cell statistics or Milan's, so the 30% improvement could be an artifact of scale normalization rather than learned transfer.","rationale":"The paper's central claim is that replacing ConvLSTM with sLSTM and adding transformer fusion improves both accuracy and cross-region generalization. The accuracy comparison in Table III is internally coherent, though headline percentages are inconsistent with the table (23% matches STN-TF rather than STN; 30% matches RMSE rather than MAE). These inconsistencies are fixable. The more consequential issue is generalization: Section IV-B2 states only 'Data are normalized per cell,' and Section IV-D never specifies whether Trentino is normalized with Trentino or Milan statistics. If Trentino statistics are used, the model is handed the true mean and scale of each unseen cell, which undermines the interpretation of Tables V and VI as evidence of transfer. I agree with the Reader's weakest assumption. I do not see an internal contradiction in the architecture or a reason to reject the Milan accuracy results, but the generalization claim cannot be accepted as-is. The appropriate verdict remains CONDITIONAL pending a precise normalization statement and a re-analysis under no-leakage normalization. I also noted that Table II lists identical train/val losses for different hyperparameter configurations, which is odd, but that concerns the tuning report rather than the main comparison and is secondary to the normalization issue.","tokens_in":14182,"tokens_out":8205,"duration_ms":94912,"concrete_test":"Request the exact Trentino preprocessing code or a precise statement: are the per-cell mean/std used to standardize Trentino computed from Trentino data, from spatially corresponding Milan cells, or from a global statistic? Then re-run the Trentino evaluation under two protocols: (A) standardize Trentino with Milan-derived statistics and inverse-transform with the same statistics; (B) standardize and inverse-transform with Trentino per-cell statistics. Include a persistence baseline (predict the last observed value) and a per-cell mean baseline. Report MAE/RMSE/R2 for all STN variants in both protocols. If the ~30% relative improvement over STN persists under protocol (A), the generalization claim is robust; if it shrinks or reverses, the headline result is an artifact of test-set normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim has two legs: better Milan accuracy and generalization to Trentino. The Milan leg is supported by Table III. The Trentino leg depends on how 'Data are normalized per cell' (Sec. IV-B2) is applied to an unseen region. If each Trentino cell is standardized with its own mean/std and predictions are inverse-transformed with those same statistics, the model is effectively told each cell's true mean and scale at test time. The evaluation then measures the ability to forecast normalized fluctuations around a given mean, not the ability to predict absolute traffic loads in a new region. A persistence or conditional-mean predictor could look good under this setup without transferring any learned magnitude information. Because Table V is the only evidence for the 30% generalization improvement, and the paper never states the normalization provenance for Trentino, the generalization leg of the central claim is not presently supported. This is a load-bearing ambiguity, not a stylistic omission.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STN-sLSTM-TF, a spatiotemporal traffic forecasting architecture that replaces the ConvLSTM temporal branch of the original STN with a scalar LSTM (sLSTM) and replaces linear fusion with transformer cross-attention. Experiments on the Italia Telecom dataset (Milan training, Trentino as an unseen test region) report that the proposed model achieves a 23% MAE reduction over a ConvLSTM-based baseline and a 30% improvement in generalization to the unseen region. The paper also presents ablation studies isolating the sLSTM branch and the transformer fusion, and an autoregressive multi-step evaluation.","tokens_in":14374,"tokens_out":6460,"duration_ms":70093,"significance":"If the claimed results hold, the paper would contribute a practical, lightweight alternative to ConvLSTM-based spatiotemporal forecasting, with a clear ablation demonstrating the contribution of each architectural component. The use of a large public dataset with a held-out geographic region for generalization testing is a valuable experimental design. However, the strength of the claims is currently limited by an unspecified normalization protocol for the unseen region, missing statistical uncertainty estimates, and an inconsistent baseline statement for the headline 23% figure. The central generalization claim is not verifiable as reported, and the accuracy claim lacks error bars.","major_comments":[{"comment":"Section IV-B2 states 'Data are normalized per cell' but never specifies whether the unseen Trentino region is normalized using Trentino's own per-cell statistics or statistics computed from Milan. This is load-bearing for the generalization claim in Section IV-D and Table V. If each Trentino cell is standardized with its own mean and standard deviation, the model is effectively provided with the target cell's marginal statistics at test time; a trivial mean-predictor would then achieve low absolute MAE/RMSE after inverse transformation, and the reported '30% improvement in model generalization' would largely measure scale invariance rather than learned transfer. Please state the exact normalization provenance for the Trentino evaluation and, if Trentino statistics were used, re-evaluate the generalization claim with a protocol that does not leak target-region statistics (e.g., using Milan statistics to transform Trentino, or reporting errors in normalized units).","section":"Section IV-C2, Table III"},{"comment":"The abstract and conclusion claim a '23% MAE reduction over ConvLSTM' (conclusion says 'over the original STN'). In Table III, the MAE reduction from STN-TF (7.2104) to STN-sLSTM-TF (5.5375) is 23.2%, while the reduction from the original STN (7.3917) is 25.1%; ConvLSTM itself is not evaluated as a standalone baseline. The 23% figure therefore appears to refer to the comparison with STN-TF, not with the original STN or with ConvLSTM. Please identify the exact baseline used for the headline number and correct the abstract and conclusion accordingly.","section":"Section IV-C2, Table III"},{"comment":"All test metrics are reported as single-run point estimates without variance or significance testing. In Table III, the advantage of STN-sLSTM-TF over STN-sLSTM is small (MAE 5.5375 vs 5.6319, SSIM 0.9921 vs 0.9912), and in Table V STN-sLSTM achieves lower MAE (1.6803 vs 2.1257) but higher RMSE (5.3016 vs 5.2649) than STN-sLSTM-TF. Without multiple seeds, confidence intervals, or a paired significance test, the claimed superiority of the full model over its closest variant is not statistically established. Please report runs over several seeds and include error bars or significance tests for the main comparisons.","section":"Sections IV-C2 and IV-D, Tables III and V"},{"comment":"The data split description is internally inconsistent: 'Models train on 1 million samples (70% of Milan, stride = 6) and test on 7 million samples (15% of Milan, stride = 1)' cannot both be true for the same dataset. If the total is 8.928 (million) samples, 70% is about 6.25 million and 15% is about 1.34 million, not 1 million and 7 million. Please clarify the actual number of training and test samples, the fraction they represent, and the stride definitions so that the evaluation protocol is reproducible.","section":"Section IV-B2"},{"comment":"The Milan quartile statistics in Table I are internally inconsistent: the 50th percentile is listed as 64.178 while the 25th percentile is 413.368, which is impossible for an ordered distribution. This appears to be a typo (perhaps 641.78 for the median). Since the table motivates the per-cell normalization and the regional disparity argument, the statistics must be corrected and verified.","section":"Table I"}],"minor_comments":[{"comment":"The neighborhood size notation is inconsistent: the text says 'local region of size (r + 1) × (r + 1)' but Eq. (5) writes entries from x^{(i-r/2, j-r/2)} to x^{(i+r/2, j+r/2)}. For r = 5, r/2 = 2.5, which is not an integer; please use a consistent indexing scheme (e.g., i-r to i+r with kernel size 2r+1) and specify the exact patch boundaries.","section":"Section III-B1 and Eq. (5)"},{"comment":"The column header 'Gap (Ovr)' is not defined in the caption or text; it appears to be the difference between validation loss and best loss (an overfitting indicator), but this should be stated explicitly.","section":"Table II"},{"comment":"The 'Number of samples 8.928' in Table I lacks units; if it represents millions of samples, please state that explicitly, as it is essential for interpreting the 1M/7M sample split described later.","section":"Section IV-B2"},{"comment":"The description of sLSTM is brief and does not provide the exact gating equations or memory-cell configuration; please either include the relevant sLSTM equations or give a precise pointer to the corresponding section in the xLSTM paper [18], so that the model is fully specified.","section":"Section III-B1"},{"comment":"The conclusion states 'Empirical results showed a 23% MAE reduction over the original STN' while the abstract says 'over ConvLSTM baselines'; these two statements are not consistent with each other or with Table III. Please unify the wording after clarifying the baseline.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate architecture extension with a clean ablation. The sLSTM swap helps consistently, and the computational cost analysis is honest. But the 30% generalization claim in the abstract is not supported by the reported protocol, because the paper never says whether the unseen Trentino region is normalized with Milan's statistics or its own. That is a load-bearing ambiguity, not a stylistic omission.\n\nWhat is actually new: STN-sLSTM-TF, an STN variant with an sLSTM temporal branch, a Conv3D spatial branch, and transformer cross-attention fusion. Each piece exists, but the combination is new and the ablation isolates the components cleanly. The Milan results are internally consistent: replacing ConvLSTM with sLSTM cuts MAE from 7.39 to 5.63, and the transformer fusion adds a smaller extra gain. Table IV shows the sLSTM variants degrade more gracefully through the autoregressive rollout. That part is credible.\n\nSoft spots, in proportion:\n\n1. The Trentino normalization question is the one that matters. Section IV-B2 says data are normalized per cell, and Section IV-D applies a Milan-trained model to Trentino, but never states whether Trentino cells are standardized with their own statistics or with Milan's. If each Trentino cell is rescaled by its own mean and std before evaluation, the model is given the true scale of every cell at test time, and the 30% RMSE gap mostly measures scale-invariant forecasting of normalized fluctuations, not transfer of learned magnitudes. The authors need to state this. If they used Trentino's own statistics, the generalization claim must be reframed.\n\n2. The headline numbers misreport the tables. The 23% MAE reduction matches the comparison to STN-TF (23.2%), not the original STN (25.1%), but the conclusion says 'over the original STN.' The 30% generalization improvement matches RMSE (31%), not MAE. This is fixable but sloppy.\n\n3. Test metrics are single-run point estimates, no seeds or error bars. The large STN-vs-sLSTM gap would survive, but the smaller gains from transformer fusion might not.\n\n4. The graph and transformer baselines cited in the intro (MT-GNN, Spacetimeformer) are not evaluated. The future-work section admits this, but it limits the strength of the comparative claims.\n\nOne more minor point: the ablation variants share hyperparameters tuned for the full model, so the component comparisons are slightly biased in favor of the full model. Not a big deal.\n\nBottom line: the Milan accuracy claim holds up as an ablation; the generalization claim is unverified as reported. This paper deserves a serious referee, but the revision needs to clarify the Trentino normalization and correct the headline numbers. I would send it to review, expecting the generalization claim to be either fixed or substantially softened.","headline":"A clean ablation showing sLSTM beats ConvLSTM inside an STN, but the headline generalization claim rests on an unspecified normalization choice for the unseen region.","tokens_in":14936,"tokens_out":3137,"would_cite":false,"duration_ms":31194,"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":"The paper claims that STN-sLSTM-TF, an STN variant with a scalar LSTM temporal branch and transformer cross-attention fusion, reduces cellular traffic forecast MAE by 23% over ConvLSTM on Milan data and transfers better to an unseen…","keywords":["cellular traffic forecasting","spatiotemporal networks","sLSTM","xLSTM","Conv3D","cross-attention fusion","model generalization","5G resource management"],"falsifier":"Open the evaluation pipeline and check whether Trentino normalization uses Milan-derived statistics or Trentino's own per-cell statistics; then rerun the Trentino test with Milan statistics applied to Trentino data and see whether STN-sLSTM's MAE advantage (1.68 versus 2.63) over STN persists. Additionally, compare gradient norms or convergence curves of sLSTM against ConvLSTM to test the gradient-stability claim directly.","tokens_in":13911,"feed_emoji":"📡","tokens_out":5977,"duration_ms":65995,"temperature":0.7,"pith_summary":"This paper tries to establish that a lightweight dual-branch network, STN-sLSTM-TF, forecasts cellular traffic more accurately than the established ConvLSTM-based STN architecture while transferring better to a region it never trained on. The authors replace the temporal ConvLSTM branch with a scalar LSTM (sLSTM), keep a three-layer Conv3D branch for spatial patterns, and fuse the two streams with transformer cross-attention. On the Italia Telecom Milan data they report a 23% reduction in mean absolute error over ConvLSTM and a 30% improvement in generalization when the Milan-trained model is applied directly to Trentino. The motivation is practical: 5G operators need accurate and cheap forecasts for resource allocation, and a simpler temporal unit plus attention fusion is claimed to deliver both.","feed_headline":"sLSTM spatiotemporal net cuts cellular forecast error by 23%","feed_subtitle":"STN-sLSTM-TF beats ConvLSTM baselines on Milan traffic and transfers to unseen Trentino.","key_machinery":"The central machinery is the scalar LSTM (sLSTM), a recurrent cell from the xLSTM family whose input and forget gates are scalar and exponentially gated, backed by multiple memory cells; it replaces ConvLSTM as the temporal branch. A three-layer Conv3D module supplies the spatial branch, and a transformer fusion layer computes cross-attention with spatial outputs as queries and temporal outputs as keys and values, replacing the original STN's linear fusion. The sLSTM carries temporal dependencies, the Conv3D captures local spatiotemporal patches, and the fusion layer adaptively combines the two streams before an MLP decodes the forecast.","core_discovery":"On the paper's own reporting, replacing ConvLSTM with sLSTM in the temporal branch and switching the fusion from a linear layer to transformer cross-attention yields the best overall spatiotemporal forecaster among the tested variants. STN-sLSTM-TF reaches a single-step MAE of 5.5375 on Milan versus 7.3917 for the original STN, a 23% reduction, and on the unseen Trentino generalization test both sLSTM variants beat all baselines, with STN-sLSTM-TF achieving an R2 score of 0.9580. The authors attribute the gains to sLSTM's exponential gating and scalar state updates, which they say stabilize gradients and improve convergence, and to cross-attention fusion that lets spatial features query temporal context, which they say helps track high-magnitude peaks and preserve spatial structure over longer horizons.","pith_inferences":["The reported Trentino gains may partly reflect per-cell normalization: since the paper says data are normalized per cell but does not state whether Trentino normalization uses Milan statistics or Trentino's own, the MAE and RMSE on Trentino could measure shape transfer rather than absolute traffic-scale transfer; a fair transfer test would normalize Trentino with Milan-derived parameters.","The comparison set is limited to the STN family, plain LSTM, and xLSTM; the 23% MAE reduction is relative to those baselines, and the paper itself lists graph-based, transformer, and diffusion-hybrid spatiotemporal models as future benchmarks.","The gradient-stability claim is supported by training-loss numbers but not by direct gradient measurements; comparing gradient norms or convergence curves of sLSTM versus ConvLSTM would test it explicitly.","The grid-level latency of STN-sLSTM-TF (326 ms) suggests that real-world deployment would need horizontal scaling or pruning, an operational consequence the paper notes but does not quantify."],"forward_implications":["If the central claim is correct, swapping ConvLSTM for sLSTM in a dual-path spatiotemporal network yields consistently lower one-step and multi-step forecast errors on the Italia Telecom dataset.","The reported 30% generalization improvement implies that a model trained on one urban region (Milan) can be applied directly to a different region (Trentino) without retraining, with MAE around 1.68 to 2.13 versus 2.63 for the original STN.","The sLSTM variants preserve SSIM above 0.94 across six autoregressive steps on unseen Trentino, meaning spatial structure is still captured at a 60-minute horizon.","Transformer fusion adds modest accuracy gains and better RMSE and R2 in high-traffic areas, but at a cost of roughly 4x the multiply-accumulate operations and about 10x the grid-level inference latency of the original STN.","Cell-level inference time stays below 10 ms across all variants, which the paper argues keeps the model within near-real-time control-loop requirements even if grid-level latency is high."],"supporting_citations":[{"why":"Defines the baseline STN architecture whose ConvLSTM temporal branch and linear fusion this paper modifies and compares against.","marker":"[15]"},{"why":"Supplies the sLSTM component, including exponential gating and scalar state updates, that replaces ConvLSTM in the temporal branch.","marker":"[18]"},{"why":"Provides the Italia Telecom dataset with Milan for training/testing and Trentino as the unseen generalization target.","marker":"[39]"},{"why":"Introduces the ConvLSTM model that serves as the temporal backbone of the baseline and as the general spatiotemporal forecasting reference.","marker":"[14]"},{"why":"Supplies the transformer cross-attention formulation used in the fusion layer.","marker":"[34]"}],"fun_headline_variants":["sLSTM traffic forecaster: 23% less error, 30% better transfer","Scalar LSTM boosts cellular forecasting: 23% MAE drop, 30% generalization gain","xLSTM-based net beats ConvLSTM: 23% MAE, 30% transfer","STN-sLSTM-TF: 23% lower MAE, 30% better generalization","sLSTM beats ConvLSTM on cellular traffic forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The generalization claim rests on how per-cell normalization is applied to Trentino: if each Trentino cell is normalized with its own mean and standard deviation rather than with statistics from Milan, the reported MAE and RMSE improvements on unseen data largely measure shape-transfer, not scale-transfer.","fun_headline_variants_meta":{"raw":{"variants":["sLSTM traffic forecaster: 23% less error, 30% better transfer","Scalar LSTM boosts cellular forecasting: 23% MAE drop, 30% generalization gain","xLSTM-based net beats ConvLSTM: 23% MAE, 30% transfer","STN-sLSTM-TF: 23% lower MAE, 30% better generalization","sLSTM beats ConvLSTM on cellular traffic forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001021,"raw_usage":{"total_tokens":4273,"prompt_tokens":876,"completion_tokens":3397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":3282}},"tokens_in":492,"tokens_out":3397,"duration_ms":27047,"temperature":1.0,"reasoning_tokens":3282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:21:46.340608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Open the evaluation pipeline and check whether Trentino normalization uses Milan-derived statistics or Trentino's own per-cell statistics; then rerun the Trentino test with Milan statistics applied to Trentino data and see whether STN-sLSTM's MAE advantage (1.68 versus 2.63) over STN persists. Additionally, compare gradient norms or convergence curves of sLSTM against ConvLSTM to test the gradient-stability claim directly.","supporting_citations":[{"cited_title":"Long-term mobile traffic forecasting using deep spatio-temporal neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines the baseline STN architecture whose ConvLSTM temporal branch and linear fusion this paper modifies and compares against."},{"cited_title":"Convolutional lstm network: a machine learning approach for precipitation nowcasting,","cited_arxiv_id":null,"evidence_quote":"Introduces the ConvLSTM model that serves as the temporal backbone of the baseline and as the general spatiotemporal forecasting reference."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer cross-attention formulation used in the fusion layer."}],"review_version":1}