{"id":"69a0403c-a968-40f0-8c7e-7a4010b595f6","arxiv_id":"2504.18008","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"TGDT combines graph attention networks and temporal convolutions to predict multiple corridor-level and intersection-level traffic measures at once from minimal inputs.","lead":"This paper introduces TGDT, a machine-learning model that estimates queue length, waiting time, and travel time for traffic along an urban corridor using only a small set of common traffic inputs. It is built from graph attention networks combined with temporal convolutions and trained on 50,000 hours of SUMO-simulated traffic for a Florida arterial.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline error claims (4 vehicles queue length, 1.5 vehicles volume) are contradicted by Table 1, which reports MAEs of 21.15 and 5.28 vehicles, so the central accuracy claim is not supported by the paper's own results.","rationale":"The reader's weakest assumption was simulation-to-reality transferability, which is a legitimate deployment concern. However, the more immediate load-bearing problem is that the headline accuracy numbers are contradicted by the paper's own Table 1. A central claim that is internally inconsistent cannot be accepted in its current form, because the reader cannot tell whether the model achieves MAE of about 21 vehicles or a maximum error of 4 vehicles for queue length. This is not a matter of external consensus or missing real-world data alone; it is a failure of the reported evidence to support the stated conclusion. The reader did note this contradiction in the rationale, so there is partial agreement, but the reader's formal weakest_assumption did not identify it as the primary issue. A REJECT verdict is appropriate for the current manuscript because the central quantitative claim is contradicted by the paper's own table. The concern is addressable: the authors could correct the claims, clarify units or metrics, or release reproducible evaluation scripts showing where 4 and 1.5 come from. Until then, the paper overstates its main result.","tokens_in":11683,"tokens_out":2920,"duration_ms":31800,"concrete_test":"Using the released repository and the held-out test split, recompute the maximum absolute error per 5-minute interval for maximum queue length and intervening volume, and also recompute the total MAE/RMSE for these MOEs. If the maximum absolute error does not approach 4 and 1.5 vehicles while the total MAE reproduces roughly 21.15 and 5.28, the Abstract and Section VI claims are unsupported. If the claimed values correspond to a different metric, subset, or unit, the authors should specify the exact mapping; otherwise the claims should be corrected to match the table.","verdict_should_be":"REJECT","load_bearing_attack":"The central quantitative claims in the Abstract and Section VI are internally inconsistent with Table 1. The paper states that TGDT achieves, at every 5-minute interval, a maximum error of 4 vehicles for queue length and 1.5 vehicles for intervening volume, alongside 24 seconds for travel time and 100 seconds for waiting time. However, Table 1's total rows for TGDT (w = 5 min) report Maximum Queue Length MAE = 21.152, MSE = 1798.9, RMSE = 42.414, and Intervening Traffic Volume MAE = 5.2843, MSE = 107.24, RMSE = 10.356. Even the TGDT-short variant reports MAE = 18.921 for queue length and 17.196 for volume. No footnote, equation, or test-set definition in the paper maps these table values to the claimed 4 and 1.5 vehicle errors. If the claim refers to a different metric, subset, or normalization, that mapping is absent. Thus, the paper's most prominent accuracy statement fails on its own evidence, independent of the additional and also unresolved question of whether SUMO-simulated SR 436 dynamics transfer to real-world corridor behavior. The simulation-reality gap matters for deployment, but the Table 1 contradiction means the synthetic-domain headline claim is not established as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TGDT, a modular deep learning digital twin for urban traffic corridors, combining graph attention networks, temporal convolutions, and transposed convolutions to estimate corridor-level travel time and intersection/phase-level maximum queue length and waiting time. The model is trained and evaluated on 50,000 hours of SUMO-simulated traffic for a Florida SR 436 corridor, using inferred OD matrices and real signal/behavior parameters. The authors report low errors on simulated test data, robustness under varying cycle lengths, traffic volumes, and green-time allocations, and a scalability claim of evaluating 1,000 scenarios in under a minute. The paper also describes sequential optimization over modular losses and provides a public code link.","tokens_in":12004,"tokens_out":3757,"duration_ms":38897,"significance":"If the reported accuracy and scalability hold, a modular GNN/TCN surrogate for corridor-level MOE estimation would be practically useful for real-time traffic signal optimization. The use of a large SUMO dataset, a direction-aware graph representation, and a modular architecture with sequential optimization are reasonable design choices, and the code release is a strength. However, the paper's headline accuracy claims are contradicted by its own Table 1, and the evaluation lacks real-world MOE ground truth, independent baselines, and repeated-run statistics. As presented, the central quantitative claims are not established, although the framework itself may be salvageable with additional analysis.","major_comments":[{"comment":"The headline accuracy claims are contradicted by Table 1. The Abstract and Section VI state maximum errors of 4 vehicles for queue length and 1.5 vehicles for intervening volume at every 5-minute interval, but Table 1's total rows for TGDT (w = 5 min) report MAE = 21.152 for Maximum Queue Length and MAE = 5.2843 for Intervening Traffic Volume. No footnote or test-set definition maps the table values to the claimed 4 and 1.5 vehicle errors, so the paper's most prominent quantitative claim is unsupported by its own evidence.","section":"Abstract and Section VI"},{"comment":"The evaluation uses only synthetic labels generated by SUMO; no real-world MOE measurements are used as ground truth. Since the digital twin is intended for a real corridor, the transferability claim requires at least a discussion of the simulation-to-reality gap or a validation against field data. Without it, the reported errors and the deployment claims in Section VI remain untested.","section":"Section III-B and Section IV"},{"comment":"The Abstract claims TGDT 'outperforms state-of-the-art baselines,' but in Table 1 the GAT-AE baseline achieves MAE = 1.1815 for Intervening Traffic Volume while TGDT achieves MAE = 5.2843, so for this MOE the proposed model is substantially worse. Moreover, all numerical baselines are the authors' own prior models, and no independent baselines or error bars from repeated runs are provided, which weakens the claimed state-of-the-art comparison.","section":"Table 1, Intervening Traffic Volume"},{"comment":"The corridor description is inconsistent: the Introduction says a 9-intersection, 10-mile corridor, while Section III-C specifies a graph with |V| = 8 and |E| = 16 and Section VI says the target arterial is 8 miles long with 8 signalized intersections. These numbers must be reconciled because the graph size and corridor length directly affect the scalability and error claims.","section":"Section I vs. Section III-C and Section VI"},{"comment":"The text says MAPE shows 'approximately 30 seconds for travel time and 100 seconds for waiting time,' but MAPE values in Table 1 are percentages (0.0248 and 0.0551). The absolute error figures appear to be MAE, not MAPE, and no corresponding MAE column is provided for travel and waiting time. Clarify which metric is being reported and add the missing table entries.","section":"Section IV, after Table 1"}],"minor_comments":[{"comment":"The reported RMSE of 1739.7 equals the reported MSE of 1739.7 in the same row; since RMSE should be the square root of MSE, this appears to be a typo.","section":"Table 1, Maximum Queue Length, TGDT-Short, Cycle Length High"},{"comment":"The text describes four modules (Minf, Mtt, Mql, Mwt), but the Figure 2 caption says the framework consists of three main modules; this should be corrected for consistency.","section":"Section II and Figure 2 caption"},{"comment":"References [7] and [8] share the same arXiv identifier and appear to be duplicate listings; the intended GCRNN citation should be distinguished from DCRNN.","section":"References [7] and [8]"}],"recommendation":"major_revision","confidential_remarks":"The discrepancy between the Abstract/Section VI headline numbers and Table 1 is severe and should be the first thing the authors are asked to resolve. Even if that is a reporting issue, the lack of independent baselines and real-world validation makes the current evidence insufficient for acceptance. The citation list is dominated by the authors' own prior work, which compounds the comparison concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things you should know about arXiv:2504.18008. First, the core contribution is a modular deep-learning surrogate: a GAT-based imputation module feeds a GAT-based travel-time regressor, and temporal CNNs produce per-phase queue length and waiting time outputs, trained sequentially. That integration appears new and is practically sensible. Second, the paper's headline accuracy claims do not hold up against its own Table 1. The abstract and conclusion say TGDT achieves a maximum error of 4 and 1.5 vehicles for queue length and volume at every 5-minute interval. Table 1 reports MAE of 21.152 vehicles for maximum queue length and 5.2843 vehicles for intervening volume under the same 5-minute setting. No definition, footnote, or test-set note reconciles those numbers. This is the first thing to fix before any other evaluation discussion.\n\nWhat the paper does well: the architecture is cleanly decomposed, the sequential optimization is a pragmatic answer to gradient interference in multi-output models, and the authors generated a large SUMO dataset (50,000 hours) with OD matrices inferred from real detectors and GPS. They also report distributional metrics (Hellinger, EMD) and show robustness under cycle-length, volume, and green-split variations. The GitHub link means the code is inspectable. That counts.\n\nSoft spots beyond the contradiction: the numerical baselines are all the authors' own prior models (MTDT, GAT-AE, FDGNN), so \"outperforms state-of-the-art\" is really \"outperforms our previous versions.\" There are no error bars or repeated runs. All evaluation is synthetic; no real-world MOE measurements are used, so transferability remains an assumption. Also minor: the intro says 9 intersections, elsewhere 8; the corridor is both 8 and 10 miles in different places. These are sloppy details but not fatal.\n\nWho this is for: traffic engineers and applied ML people who want a fast surrogate for corridor signal optimization. The architecture is worth reading, and the code could be reused. As submitted, the evaluation is not trustworthy as reported. A serious referee should see it, because the problem is important and the method is concrete, but the authors must resolve the accuracy-claim discrepancy, add independent baselines or clearly label the comparison, and ideally include at least one real-world validity check before acceptance.\n\nRecommendation: send to peer review with a major-revision decision.","headline":"Modular GAT/TCN corridor surrogate is a practical engineering integration, but the paper's headline error claims are contradicted by its own Table 1.","tokens_in":12522,"tokens_out":3064,"would_cite":false,"duration_ms":28743,"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":"TGDT is a temporal graph-based digital twin that simultaneously estimates corridor travel time, intersection-level queue length and waiting time, and intervening traffic volumes from minimal inputs, reporting travel-time error around 24…","keywords":["digital twin","urban traffic corridor","graph attention network","temporal convolutional network","travel time estimation","queue length","waiting time","traffic signal optimization"],"falsifier":"Run TGDT on the same corridor with ground-truth queue lengths, waiting times, and travel times collected from video detection, in-road sensors, or probe vehicles over multiple days, and check whether the 5-minute interval estimates stay within the reported bounds (travel time within 24 seconds, waiting time within 100 seconds, queue length within 4 vehicles, volume within 1.5 vehicles); if the real-world errors exceed these bounds systematically, the claim of serving as a real-time digital twin is falsified.","tokens_in":11470,"feed_emoji":"🚦","tokens_out":8848,"duration_ms":78648,"temperature":0.7,"pith_summary":"TGDT is a deep-learning digital twin for signalized urban corridors that estimates several traffic measures at once: bidirectional corridor travel time, per-phase maximum queue length and waiting time at each intersection, and intervening traffic volumes. The paper claims that a modular network combining graph attention layers with temporal convolutions, trained only on interval-aggregated inflows, signal timing plans, and driving behavior parameters, achieves travel-time errors around 24 seconds, waiting-time errors around 100 seconds, and maximum queue-length and volume errors of 4 and 1.5 vehicles on 5-minute intervals. Because the model runs fully parallelized and evaluates over a thousand scenarios in under a minute, the authors argue it can replace slow micro-simulation for real-time signal optimization and corridor assessment. Validation is done on 50,000 hours of synthetic logs from a calibrated micro-simulator of an eight-intersection arterial corridor, benchmarked against earlier graph-based and multi-task digital twins.","feed_headline":"Digital twin estimates corridor travel time within 24 seconds","feed_subtitle":"One model also outputs queue length and waiting time for every intersection and phase.","key_machinery":"The load-bearing mechanism is a two-level graph representation of the corridor: nodes are intersections, edges are directional road segments, and each scenario becomes a directed acyclic graph with node features holding 5-minute inflow volumes and edge features holding distances, turning counts, densities, and driving-behavior parameters. A static graph feeds an inflow-imputation module built on self-attention and graph attention layers; a dynamic graph, whose edge features evolve over ten time steps, feeds the travel-time module; the learned spatiotemporal representation is then upsampled by transposed convolutions and passed through a CNN encoder to produce queue-length and waiting-time series. Intermediate fusion merges node and edge embeddings, and three separate optimizers update the modules sequentially, which the paper says avoids gradient interference and makes the architecture extensible to more intersections and measures.","core_discovery":"The central discovery is that corridor-level and intersection-level measures of effectiveness can be estimated concurrently by one architecture: a graph attention network captures spatial dependencies among intersections and directional road segments, temporal convolutional layers turn those representations into per-phase time series, and a sequential optimization scheme trains each module (inflow imputation, travel time, queue length, waiting time) in a fixed order so that higher-level estimates inform lower-level ones. On an eight-intersection arterial corridor, TGDT reports mean absolute percentage errors around 2.5% for travel time, a Hellinger distance of 0.08 and a normalized Earth mover's distance of 0.04 for travel-time distributions, waiting-time MAPE near 5.5%, and maximum errors of 4 and 1.5 vehicles for queue length and directional volume per 5-minute interval. The paper positions this as an interpretable, scalable surrogate for micro-simulation that needs only commonly collected inputs.","pith_inferences":["If the simulator calibration faithfully reproduces the real corridor, the same training pipeline could be retargeted to a new corridor by re-running simulation and retraining; the paper does not demonstrate zero-shot transfer to unseen corridors.","The sequential optimization imposes an implicit causal chain from inflow to travel time to queue and waiting estimates, which could be exploited for error diagnosis: a failure at the queue-length stage can be traced back to the travel-time or inflow module.","A natural next test is to compare TGDT's outputs against real-world detector and probe data on the same corridor; if the sim-to-real gap is small, the framework becomes a practical optimization engine, and if not, the reported error bounds are only valid in the simulated world.","The same architecture could be extended to predict derived measures like delay, emissions, or fuel consumption by adding output heads, since those quantities are correlated with the already-estimated queue and travel-time series."],"forward_implications":["If the reported accuracy holds, traffic agencies can screen thousands of signal-timing plans in seconds, making real-time adaptive signal control feasible without running micro-simulations.","The modular sequential design means adding a new measure of effectiveness or a new intersection should only require retraining or fine-tuning the corresponding module, not the whole network.","The minimal input set of interval inflows, signal plans, turning ratios, distances, and behavior parameters lowers the data barrier for corridors that lack dense sensing.","Corridor-level travel time and intersection-level queue and waiting estimates are produced in lockstep, so a single model can support both route-level and signal-level decisions.","Because it runs in about a minute for 1,000 scenarios, TGDT can serve as the inner-loop surrogate for optimization algorithms that search over signal offsets, cycle lengths, and green splits."],"supporting_citations":[{"why":"Supplies the microscopic traffic simulator used to generate the 50,000 hours of training and evaluation logs.","marker":"[10]"},{"why":"Provides the self-attention mechanism used in the inflow-imputation module.","marker":"[16]"},{"why":"The prior multi-task digital twin that TGDT extends and outperforms on queue-length and waiting-time estimation.","marker":"[19]"},{"why":"The graph autoencoder baseline for intervening-traffic-volume imputation that TGDT is compared against.","marker":"[20]"},{"why":"The graph attention network baseline for travel-time distribution prediction that TGDT is compared against.","marker":"[21]"}],"fun_headline_variants":["Temporal graph twin estimates corridor travel time in seconds","One graph model outputs travel time, queues, and waits per intersection","Digital twin hits 2.5% MAPE for corridor travel time","Graph attention twin simulates 1,000 traffic scenarios in seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the calibrated micro-simulator reproduces real traffic dynamics closely enough that a model trained on its logs estimates real-world queue lengths, waiting times, and travel times accurately; the paper does not compare predictions against real-world measurements.","fun_headline_variants_meta":{"raw":{"variants":["Temporal graph twin estimates corridor travel time in seconds","One graph model outputs travel time, queues, and waits per intersection","Digital twin hits 2.5% MAPE for corridor travel time","Graph attention twin simulates 1,000 traffic scenarios in seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2564,"prompt_tokens":931,"completion_tokens":1633,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":1569}},"tokens_in":547,"tokens_out":1633,"duration_ms":11876,"temperature":1.0,"reasoning_tokens":1569,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:27:42.213109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TGDT on the same corridor with ground-truth queue lengths, waiting times, and travel times collected from video detection, in-road sensors, or probe vehicles over multiple days, and check whether the 5-minute interval estimates stay within the reported bounds (travel time within 24 seconds, waiting time within 100 seconds, queue length within 4 vehicles, volume within 1.5 vehicles); if the real-world errors exceed these bounds systematically, the claim of serving as a real-time digital twin is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the microscopic traffic simulator used to generate the 50,000 hours of training and evaluation logs."},{"cited_title":"Yousefzadeh, R","cited_arxiv_id":null,"evidence_quote":"The prior multi-task digital twin that TGDT extends and outperforms on queue-length and waiting-time estimation."},{"cited_title":"Yousefzadeh, R","cited_arxiv_id":null,"evidence_quote":"The graph autoencoder baseline for intervening-traffic-volume imputation that TGDT is compared against."},{"cited_title":"Dynamic Graph Attention Networks for Travel Time Distribution Prediction in Urban Arterial Roads","cited_arxiv_id":"2412.11095","evidence_quote":"The graph attention network baseline for travel-time distribution prediction that TGDT is compared against."}],"review_version":1}