{"id":"abe45b9b-7548-4179-869c-97cb3d4c41f6","arxiv_id":"2412.13607","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A graph-free MLP-Mixer with independent patch-wise MLP masked pretraining matches or beats complex spatiotemporal models on large-scale traffic forecasting at a fraction of the compute.","lead":"PreMixer is an all-MLP traffic forecasting framework that combines an MLP-Mixer predictor with a patch-wise MLP pre-training model and spatio-temporal positional encoding, tested on the LargeST datasets with up to 8,600 sensors. It reports accuracy competitive with graph- and transformer-based models while training and inferring much faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not tested against NexuSQN, the MLP-Mixer from which the SpatialMixer is explicitly adapted; without this baseline the pre-training contribution cannot be isolated.","rationale":"The reader correctly identifies the pre-training ablation as statistically fragile. I agree that the 0.37/0.33 MAE differences in Table IV have no error bars and may be noise. However, a more foundational issue precedes that concern: the paper's forecasting model is explicitly built on the pattern-aware mixing from NexuSQN, yet NexuSQN is not included as a baseline. Without that comparison, the central claim of state-of-the-art accuracy is untestable, because the observed performance could be entirely due to borrowed components. This also subsumes the pre-training question: only by comparing PreMixer without pre-training against original NexuSQN can one determine whether the pre-training module adds any real value or merely compensates for a weaker reimplementation. The missing comparisons are addressable, so a conditional verdict remains appropriate; the paper should be required to add the closest MLP-Mixer and pretraining baselines and to report variance before acceptance.","tokens_in":23207,"tokens_out":4130,"duration_ms":37741,"concrete_test":"Run the official NexuSQN implementation on the same LargeST 2019 data with the same 6:2:2 split and report horizon-3/6/12 MAE/RMSE/MAPE on SD, GBA, GLA, and CA. Also run the paper's 'w/o Pre-training' variant under identical seeds and hardware. If PreMixer without pretraining does not beat NexuSQN on the majority of datasets, or if the gap between PreMixer and its own w/o Pre-training variant is within run-to-run noise (e.g., bootstrap confidence intervals overlap), then the SOTA and pre-training claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B explicitly adapts the pattern-aware mixing from NexuSQN [37] and fuses STPE with learnable node embedding, which is effectively NexuSQN's STNE. Yet Table II lists no MLP-Mixer baselines: STID, ST-MLP, RPMixer, and NexuSQN are absent, as are the pretraining baselines STEP and STD-MAE. Therefore the central claim that PreMixer outperforms state-of-the-art methods is unsupported against its closest architectural relatives. In particular, the ablation 'w/o Pre-training' (Table IV) only compares PreMixer against a degraded version of its own architecture; it does not compare against the original NexuSQN. If NexuSQN alone matches or beats PreMixer, the observed accuracy is attributable to borrowed components, and the pretraining gain (MAE +0.37 SD, +0.33 GBA, no error bars) may be noise or compensatory. The missing comparison is thus load-bearing for both the accuracy and the novelty claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PreMixer, an all-MLP framework for large-scale traffic forecasting that combines a patch-wise MLP autoencoder pre-training stage with a downstream MLP-Mixer predictor. The pre-training uses masked reconstruction and complementary contrastive learning on long-term history, while the downstream model uses spatio-temporal positional encoding and learnable node embeddings to avoid graph dependence. Experiments on four LargeST sub-datasets (SD, GBA, GLA, CA) report competitive MAE/RMSE/MAPE with substantially lower training and inference time than graph-based baselines.","tokens_in":23437,"tokens_out":6281,"duration_ms":60739,"significance":"If the empirical claims hold, PreMixer offers a graph-free, efficient alternative for large-scale traffic forecasting, with pre-training enhancing downstream accuracy at low computational cost. The efficiency gains reported in Table V are striking (e.g., about 4 hours total training on CA versus 84-207 hours for some baselines), and the patch-wise independent encoding idea is simple and potentially transferable. However, the significance is tempered by the absence of the most closely related MLP-Mixer and pre-training baselines, by the lack of statistical validation for the small pre-training gains, and by the overstatement that PreMixer outperforms all state-of-the-art methods.","major_comments":[{"comment":"The claim that 'PreMixer outperforms state-of-the-art methods' (Section I, contribution 4) is not supported by the reported numbers: on SD the average MAE of D2STGNN (17.38) and DGCRN (17.65) are both lower than PreMixer's 18.02, and on GBA TSMixer's average MAE (19.58) is lower than PreMixer's 20.06. PreMixer is best only on the two largest datasets, GLA and CA. This overstatement should be corrected, for example by describing the results as 'competitive' or 'best on the two largest datasets'.","section":"§IV-B, Table II"},{"comment":"The paper does not compare against any MLP-Mixer-based traffic baselines (NexuSQN [37], STID [28], ST-MLP [36], RPMixer [35]) or pre-training baselines (STEP [27], STD-MAE [31]), despite citing them in the literature review. Section III-B explicitly adapts the pattern-aware mixing operation from NexuSQN and fuses spatio-temporal positional encoding with learnable node embeddings, which closely resembles NexuSQN's spatio-temporal node embedding. As a result, the ablation 'w/o Pre-training' (Table IV) isolates the pre-training contribution only relative to PreMixer itself, not relative to the architecture it is built upon; a comparison with NexuSQN is needed to support both the novelty and accuracy claims.","section":"§IV-A2 and Table II"},{"comment":"The downstream improvement from pre-training is reported as a MAE decrease of 0.37 on SD and 0.33 on GBA, with no error bars or significance tests. Given the typical variance across random seeds in traffic forecasting experiments, these differences may be within run-to-run noise. The authors should report mean and standard deviation over multiple runs and provide a statistical test, at least for the ablations of the two key components (pre-training and contrastive learning), to substantiate the central claim that pre-training helps.","section":"§IV-D, Table IV"},{"comment":"The efficiency analysis omits the cost of the pre-training stage. Table V reports per-epoch training and inference times for the forecasting model, but the total training time should also include the pre-training phase on long-term historical data; otherwise the claim of 'without incurring significant time consumption' is not fully substantiated. Additionally, the TSMixer RMSE values on GLA and CA (e.g., 224.82 at horizon 3 on GLA, 106.28 at horizon 3 on CA) are orders of magnitude larger than the corresponding MAE values (20.38 and 18.40), which suggests a bug or misconfiguration in the baseline; this undermines the fairness of the comparison on those datasets.","section":"§IV-E, Table V"},{"comment":"The contrastive learning formulation is unclear and likely misindexed: the text 'Assume z(i,n)1 = z(i+Tp,n)1 and z(i+2Tp,n)1 be the two views' is confusing, and Eq. (4) sums over 2Tp items while the original patched input has Tp patches. Please rewrite the two-view generation and the positive-pair indexing precisely, and clarify why both masked and unmasked patches are reconstructed in the reconstruction loss.","section":"§III-B1, Eqs. (4)-(5)"}],"minor_comments":[{"comment":"The manuscript contains numerous typos and grammatical errors, such as 'sptiotemporal', 'rode networks', 'designes', 'imporvement', 'trining', and 'conducted'. A careful language revision is needed.","section":"Throughout"},{"comment":"The spatio-temporal positional encoding formula uses both C and D inconsistently; the dimensions of Upos should be defined clearly and the indexing in Eq. (7) should be consistent with the stated shape R^{T x N x C}.","section":"§III-B, Eq. (7)"},{"comment":"Reference [53] is a paper on translating math formula images, not an appropriate source for sinusoidal positional encoding; the original Transformer paper (Vaswani et al., 2017) should be cited instead.","section":"§II-B, reference [53]"},{"comment":"The baseline list omits the recent MLP-Mixer and pre-training baselines discussed in the literature review; at minimum, the authors should state why these methods were excluded, or add them to Table II.","section":"§IV-A2, baselines"},{"comment":"The sentence 'The PreMixer, equipped with auxiliary information, significantly outperforms its ablated versions' is an overstatement given the small differences in Table IV and the absence of significance tests; the wording should be softened.","section":"§IV-D"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a traffic forecasting journal and the efficiency results are potentially valuable. The main issues are overclaimed SOTA performance, missing comparison with the closest architectural relatives (especially NexuSQN), lack of statistical validation for the pre-training gain, and an efficiency comparison that omits pre-training cost. These are fixable within the scope of a revision, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a credible efficiency story with an interesting pretraining idea, but the accuracy claims are oversold and the experiment table omits the exact baselines needed to believe the central novelty.\n\nWhat's genuinely new: combining patch-wise MLP masked autoencoding with a complementary temporal contrastive loss, then feeding those representations into an MLP-Mixer predictor, is a combination I haven't seen in the traffic forecasting literature. The paper also does a service by running on LargeST's full CA dataset (8,600 sensors) and reporting runtime; the efficiency gains are striking and the plots make that point clearly. The ablations, while limited, at least show every component contributes something.\n\nThe soft spots are real and not minor. Table II is missing STID, ST-MLP, RPMixer, and NexuSQN. The last one is especially hard to excuse: the SpatialMixer is explicitly adapted from NexuSQN's pattern-aware mixing (including the STPE/node-embedding fusion that mirrors NexuSQN's STNE). Without that comparison, the pre-training contribution cannot be isolated from borrowed components. The ablation 'w/o Pre-training' only degrades PreMixer itself; it does not tell you whether the pre-training adds anything over NexuSQN alone. The ablation differences are also small—0.33 and 0.37 MAE—with no error bars or statistical test, so they could be run-to-run noise.\n\nThe SOTA claim is overstated. On SD, DGCRN and D2STGNN have lower average MAE; on GBA, TSMixer is better. The abstract says 'comparable state-of-the-art performance,' which is accurate, but the contribution list says 'outperforms state-of-the-art methods.' Those are different statements.\n\nNo code is shipped, which makes it hard to check the efficiency numbers, though the runtime table is internally consistent.\n\nThis is a solid candidate for peer review, but not for acceptance as-is. The missing MLP baselines are the load-bearing issue. A referee should ask for NexuSQN and ST-MLP runs, error bars on the ablations, and a rewrite of the accuracy claims. If the efficiency results hold up, this will be a useful paper for the large-scale traffic forecasting community.","headline":"Efficient all-MLP large-scale traffic forecaster with a novel pretraining scheme, but missing key baselines and overblown SOTA claims make the accuracy gains unproven.","tokens_in":23972,"tokens_out":2263,"would_cite":false,"duration_ms":20767,"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":"PreMixer: a graph-free all-MLP model with patch-wise MLP pre-training reaches state-of-the-art accuracy on large-scale traffic forecasting.","keywords":["traffic forecasting","MLP-Mixer","pre-training","masked autoencoding","contrastive learning","spatio-temporal positional encoding","large-scale datasets","computational efficiency"],"falsifier":"Run the full PreMixer and its no-pre-training variant multiple times with different random seeds on SD and GBA, then compare MAE distributions; if the 0.37 and 0.33 gaps fall within the run-to-run spread, the pre-training claim collapses. As a second check, pre-train PIEncoder on randomly permuted or time-shuffled patches and retrain the downstream predictor: if accuracy barely changes, the encoder is not contributing real temporal context.","tokens_in":23000,"feed_emoji":"🚦","tokens_out":7971,"duration_ms":83406,"temperature":0.7,"pith_summary":"This paper tries to establish that large-scale traffic forecasting can be done accurately and cheaply without graph neural networks or transformers. The proposed framework, PreMixer, couples an MLP-Mixer predictor with a lightweight MLP-based pre-training stage whose patch-wise encoder learns contextual representations from long-term history; it adds spatio-temporal positional encodings and learnable node embeddings in place of a predefined graph. On four real-world datasets with up to 8,600 sensors, the paper reports that PreMixer outperforms twelve baselines on most settings, including the largest dataset, while training roughly an order of magnitude faster than the fastest graph-based competitor. If true, the result matters because deployment-scale traffic forecasting currently struggles with the memory and time demands of graph and attention models.","feed_headline":"All-MLP traffic forecaster outruns graph networks at scale","feed_subtitle":"PreMixer matches state-of-the-art accuracy with a fraction of the training time on California's 8,600-sensor network.","key_machinery":"The load-bearing object is PIEncoder, a patch-wise MLP autoencoder. Long-term history (one week, 56 patches of length 12) is split into non-overlapping patches; each patch is embedded independently by a two-layer MLP with ReLU, then linearly reconstructed, under a loss that reconstructs both masked and unmasked patches (Equation 3) plus a temporal contrastive loss between two complementary-mask views (Equations 4–5). Patch independence means the encoder shares weights across time and nodes and therefore accepts any input length at forecasting time. The second mechanism is spatio-temporal positional encoding (STPE), which encodes time index and sensor index sinusoidally in separate halves of the feature dimension, together with a learnable node embedding; both are injected into the Mixer predictor to replace a predefined road-network graph.","core_discovery":"On its own terms, the paper's central discovery is that an all-MLP architecture with two auxiliary components—an MLP-based pre-training encoder and spatio-temporal positional encoding plus learnable node embeddings—can beat established spatiotemporal graph networks and Transformer baselines on large-scale traffic forecasting. The pre-training stage, PIEncoder, embeds each time-series patch independently through a shared two-layer MLP, reconstructs both masked and unmasked patches, and adds a temporal contrastive loss between two views generated by a complementary 50% mask. Because patches never exchange information during pre-training, the learned encoder can process arbitrary input lengths at the forecasting stage, where its representations are projected and added to the Mixer predictor's input. The paper reports the best average MAE/RMSE/MAPE on GBA, GLA, and CA among all compared methods, with the largest gains on the 8,600-sensor CA dataset, and shows that the pre-trained encoder transfers across regions with larger source regions helping more.","pith_inferences":["Because patch embeddings are independent and input-agnostic, the same pre-train-and-mix recipe could be dropped into other multivariate spatiotemporal forecasting problems—energy load, weather station networks, or city-wide demand—where weekly periodicity and sensor heterogeneity matter but no reliable graph exists.","The pre-training benefit is the least certain part of the claim: the only direct evidence is a two-dataset ablation with small MAE gaps and no error bars, so multi-seed runs are needed before the pre-training stage can be credited rather than the Mixer backbone or STPE.","The unusually high RMSE values reported for TSMixer on GLA and CA suggest that the plain MLP-Mixer baseline becomes numerically unstable at scale; if so, part of PreMixer's advantage may come from auxiliary embeddings that stabilize mixing, which would be worth isolating experimentally."],"forward_implications":["Large-scale deployment becomes practical: PreMixer trains and infers far faster than every baseline, and it is one of only a few models that runs on the largest 8,600-sensor dataset at all.","The pre-trained patch encoder can be reused across regions: transferring from the large CA dataset improves downstream results on SD, while transfers from smaller or differently distributed regions do not.","Each component contributes: removing pre-training, contrastive learning, context, or STPE degrades MAE on SD and GBA, so the full composition is what achieves the reported accuracy.","The architecture's success on the two largest datasets suggests that graph structure is not necessary for capturing spatial dependencies in large-scale traffic networks."],"supporting_citations":[{"why":"Supplies the LargeST benchmark datasets (SD, GBA, GLA, CA) and the 6:2:2 training/validation/test split used in all experiments.","marker":"[24]"},{"why":"Introduces masked pre-training to enhance downstream STGNNs for traffic; PreMixer's pre-train-then-finetune design positions itself relative to this approach.","marker":"[27]"},{"why":"Defines the TSMixer all-MLP architecture for time series, the main architectural baseline PreMixer extends and must outperform.","marker":"[34]"},{"why":"Provides the pattern-aware mixing formulation (NexuSQN) that PreMixer adopts for its structured SpatialMixer with spatio-temporal node embeddings.","marker":"[37]"},{"why":"Graph WaveNet is the strongest graph-based baseline that runs on all four datasets, so PreMixer's accuracy and efficiency comparisons depend on it.","marker":"[15]"},{"why":"STD-MAE is a transformer-based masked pre-training method for traffic; it represents the costly pre-training approach PreMixer claims to replace.","marker":"[31]"},{"why":"SimST argues GNNs are unnecessary for traffic forecasting, providing the conceptual basis for PreMixer's graph-free Mixer design.","marker":"[13]"},{"why":"The masked autoencoder (MAE) strategy from vision is the template for PIEncoder's masked reconstruction objective.","marker":"[29]"}],"fun_headline_variants":["MLP-only traffic model beats graph networks at scale","PreMixer: pre-trained MLPs outdo spatiotemporal graphs","Graph-free MLP forecaster scales to 8,600 sensors efficiently","Patch-wise MLP pre-training yields accurate large-scale forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central novelty—that pre-training on long-term patches is what gives PreMixer its edge—rests on an ablation on two datasets (SD and GBA) with no error bars or significance tests, so the small MAE gaps (0.37 and 0.33) could be run-to-run noise.","fun_headline_variants_meta":{"raw":{"variants":["MLP-only traffic model beats graph networks at scale","PreMixer: pre-trained MLPs outdo spatiotemporal graphs","Graph-free MLP forecaster scales to 8,600 sensors efficiently","Patch-wise MLP pre-training yields accurate large-scale forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1871,"prompt_tokens":1014,"completion_tokens":857,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":786}},"tokens_in":630,"tokens_out":857,"duration_ms":8560,"temperature":1.0,"reasoning_tokens":786,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:57:38.498976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full PreMixer and its no-pre-training variant multiple times with different random seeds on SD and GBA, then compare MAE distributions; if the 0.37 and 0.33 gaps fall within the run-to-run spread, the pre-training claim collapses. As a second check, pre-train PIEncoder on randomly permuted or time-shuffled patches and retrain the downstream predictor: if accuracy barely changes, the encoder is not contributing real temporal context.","supporting_citations":[{"cited_title":"Largest: A benchmark dataset for large- scale traffic forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the LargeST benchmark datasets (SD, GBA, GLA, CA) and the 6:2:2 training/validation/test split used in all experiments."},{"cited_title":"Contextualizing MLP-Mixers Spatiotemporally for Urban Data Forecast at Scale","cited_arxiv_id":"2307.01482","evidence_quote":"Provides the pattern-aware mixing formulation (NexuSQN) that PreMixer adopts for its structured SpatialMixer with spatio-temporal node embeddings."},{"cited_title":"Graph wavenet for deep spatial-temporal graph modeling,","cited_arxiv_id":null,"evidence_quote":"Graph WaveNet is the strongest graph-based baseline that runs on all four datasets, so PreMixer's accuracy and efficiency comparisons depend on it."},{"cited_title":"Masked au- toencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"The masked autoencoder (MAE) strategy from vision is the template for PIEncoder's masked reconstruction objective."}],"review_version":1}