{"id":"958973b4-331c-467f-8cdd-7abeacff9662","arxiv_id":"2502.01107","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"GTG combines graph networks, adversarial domain adaptation, and preference-based shortest path search to generate trajectories in cities without historical trajectory data, but its evaluation leaks the test origins and destinations into the generator.","lead":"The paper introduces a model that learns how people travel in one city and then generates realistic driving routes in another city using only the second city's road map. The authors report large gains over prior trajectory generators, but the evaluation gives the model the true start and end points of the test trips, which makes the comparison unfair.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Target test OD pairs are fed to the generator and then used to score the results, so the 'no trajectory data' claim and the macro-metric comparisons are not supported.","rationale":"The paper's strongest claim is empirical: GTG generalizes to new cities and significantly outperforms baselines. That claim rests on the evaluation protocol, and the protocol uses target test OD pairs. The problem statement in Eq. 2 explicitly promises a target city with no trajectory data, so if OD pairs are considered part of trajectory data, the setting itself is violated. If OD pairs are treated as externally supplied demand, the paper never says so and never gives the same demand information to most baselines. Either way, Table 1 cannot establish the advertised generalization advantage. The micro-metric comparison is particularly problematic because it matches each generated trajectory to the real trajectory with the same OD, so a model that simply returns the true shortest path between the supplied endpoints would score well on macro Distance/Radius metrics even if it had learned nothing about human preferences. The reader identified the same weakest assumption, and I agree that it is load-bearing. A separate issue, the degeneracy of the preference-learning loss in Eq. 33, is also concerning, but the OD leakage is sufficient on its own to reject the central claim. A controlled re-evaluation with source-only OD prediction would test whether the reported margins survive without leaking target trajectory information.","tokens_in":19839,"tokens_out":6353,"duration_ms":74705,"concrete_test":"Re-run the new-city experiments with GTG forced to use OD pairs predicted by a gravity model fitted only on source-city trajectories (no target ODs at all). If the Distance/Radius JSD values rise substantially (e.g., more than 2x) or approach the levels of unconditional baselines, the reported advantage is largely an artifact of test-OD leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GTG generates trajectories for a new city without any target trajectory data (Sec. 2.2, Eq. 2). But the method only generates a path between a supplied origin and destination (Sec. 3.4), and the evaluation protocol supplies the target test set's true OD pairs: the micro-metric definition in Sec. C.2 (Eq. 41) compares each generated trajectory with the real trajectory 'with the same travel demand,' and Sec. C.1 states that TS-TrajGen and GTG take both the starting and destination road segments from the travel demand as input. Since each real test trajectory contains its OD, using those ODs leaks target trajectory information into the generation process. This makes the 'no trajectory data' framing false, and it also biases the macro metrics: Distance and Radius JSD values in Table 1 are partly predetermined by the true OD distribution rather than by route-generation quality. Baselines such as RW, DE, SE, SG, SV, MS, DT, and VO are not given destinations (Sec. C.1), so the comparison is unbalanced. The claim that GTG 'significantly outperforms' existing models in generalization is therefore not supported by the reported experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GTG, a trajectory generation model intended to transfer from one city to another without using target-city trajectory data. The model combines Space Syntax topological features, a SAGAT encoder, disentangled adversarial domain adaptation for cross-city travel-cost prediction, and a shortest-path search with learned preference updates. The evaluation on Beijing, Xi'an, and Chengdu claims that GTG substantially outperforms existing trajectory generators in a new-city, no-fine-tuning setting, and that fine-tuning on small amounts of target data further improves the results.","tokens_in":20049,"tokens_out":7422,"duration_ms":78826,"significance":"If the claims were supported, the paper would make a useful contribution: learning invariant mobility patterns that transfer across cities could enable synthetic trajectory generation for cities without historical trajectory data, which would be valuable for downstream analytics and privacy-sensitive applications. The modeling pipeline is coherent, the code is promised publicly, and the paper includes experiments on three real cities with baselines and ablations. However, the reported experimental protocol is not a valid test of the stated generalization claim, because the generator is given the target test set's origin-destination pairs and most baselines are not given the same information. The core significance is therefore conditional on a re-designed evaluation that either generates OD demand from no target trajectory data or clearly reframes the contribution as conditional route generation.","major_comments":[{"comment":"The stated task is to generate a trajectory dataset for a target city with no trajectory data, but the method only produces a path after an origin and destination are supplied, and the experiments obtain those OD pairs from the target test set. Eq. (41) evaluates each generated trajectory against its real counterpart 'with the same travel demand,' and Section C.1 states that GTG and TS-TrajGen take the starting and destination road segments from the travel demand as input. This means the evaluation injects test-trajectory endpoint information into the generator, directly contradicting Eq. (2). It also makes the Distance and Radius macro metrics in Table 1 largely predetermined by the true OD distribution rather than by route-generation quality. Unless the method is extended to generate OD demand without any target trajectory data, the paper solves conditional route generation, not the trajectory-dataset generation problem stated in Eq. (2).","section":"Sec. 2.2, Eq. (2); Sec. 3.4; Sec. C.1; Sec. C.2, Eq. (41)"},{"comment":"The comparison in Table 1 is unbalanced because most baselines (RW, DE, SE, SG, SV, MS, DT, VO) do not receive destination information, while GTG does. Only TS-TrajGen is conditioned on the same travel demand as GTG, so the aggregate 'significantly outperforms' claim conflates a conditioning advantage with a generalization advantage. The paper should report conditioned and unconditioned baseline groups separately and compare GTG with baselines given identical inputs.","section":"Sec. C.1 and Table 1"},{"comment":"No repeated runs, confidence intervals, or significance tests are reported for the main results. Many JSD values in Table 1 are on the order of 1e-4 to 1e-3, and without variance estimates the claim of 'significant improvements' is not supported even for the comparison with TS-TrajGen.","section":"Sec. 4.2, Table 1"}],"minor_comments":[{"comment":"'Experiments on three datasets demonstrates' should be 'demonstrate'.","section":"Abstract"},{"comment":"The text says 'decoupling of z(s)i and z(s)i'; this should refer to the semantic latent variable z(s) and the domain latent variable z(d).","section":"Sec. 3.3"},{"comment":"The hyperparameter list has 'λd = 100and λg = 5' with a missing space; please also state the choice of the number of SAGAT layers and the cluster count K as hyperparameters in one place.","section":"Sec. C.3"},{"comment":"The LSTPM row for GTG2 under BJ reports MRR@3 = 0.60 while ACC = 0.78 and NDCG@3 = 0.87; this outlier should be checked for a typo.","section":"Table 2"},{"comment":"There are typos including 'prefernece' and 'paradiagm'; the typesetting of 'V AE' with a space should also be cleaned up.","section":"Sec. 5"},{"comment":"The main text says the MaxEnt IRL theoretical analysis is in the code repository, but Appendix A.2 contains the derivation; the main text should cite the appendix instead of an external repository. The constraint expression in Eq. (34) is also garbled and should be reformatted.","section":"Sec. 3.4 and Appendix A.2"}],"recommendation":"reject","confidential_remarks":"The OD leakage is the central problem: the method is evaluated with the target test set's true origin-destination pairs, and the headline comparison against baselines that do not receive destinations does not establish the claimed generalization ability. This is not a local presentation issue; it affects the main claim. I would encourage the authors to resubmit a version that either generates OD demand from no target trajectory data or explicitly defines the contribution as conditional route generation, with all baselines compared under the same conditioning and with repeated-run statistics. As submitted, the evidence cannot support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a reasonable recombination of existing pieces, and the authors ship code and run a broad set of experiments. But the main generalization claim isn't supported: the target test OD pairs are given to the generator, so the 'no trajectory data' framing is misleading and the macro metrics are partly forced.\n\nWhat's new: combining Space Syntax features with disentangled adversarial domain adaptation and a shortest-path preference search for cross-city trajectory generation. That's a genuine pipeline novelty. The paper also does the right thing by evaluating on three cities, including downstream task support, fine-tuning, and ablations. Code is public. The ablation showing cost prediction matters is informative.\n\nWhere it falls down: the evaluation protocol. Section C.1 says TS-TrajGen and GTG take both the starting and destination road segments from the travel demand as input. The micro-metric definition (Eq. 41) compares each generated trajectory with the real trajectory 'with the same travel demand.' So the true test-set OD pairs are fed into the generator, then used to score the output. That leaks target trajectory information and makes the 'generate trajectories for a new city with no trajectory data' claim false in an important way. It also stacks the deck against baselines that don't get destinations. The Distance and Radius JSD values are largely predetermined by the OD input, not by route quality.\n\nThe preference learning objective has a separate issue: Eq. 33 can be minimized by driving all preference weights to zero, since the shortest path cost is then zero and the real trajectory cost is zero. The authors mention MaxEnt IRL and point to the code for theory, but the appendix derivation is a single-sample Monte Carlo approximation without error control. That's not fatal by itself, but combined with no error bars or multiple seeds, the quantitative claims are weaker than the text suggests.\n\nBottom line: the core idea is worth taking seriously, but the paper as written doesn't demonstrate cross-city generalization. A serious referee should engage with it—there's enough substance and code to warrant a round of revision. My own verdict would be reject-and-resubmit, with the OD leakage as the primary issue.","headline":"The pipeline is a genuine recombination of known ideas with released code, but the evaluation leaks target test OD pairs into the generator, so the headline generalization claim doesn't hold.","tokens_in":20587,"tokens_out":1868,"would_cite":false,"duration_ms":18954,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A model trained on one city's trajectories can generate realistic routes in a new city with no local trajectory data.","keywords":["trajectory generation","cross-city generalization","urban mobility","Space Syntax","disentangled adversarial domain adaptation","travel cost prediction","shortest path search","preference learning"],"falsifier":"Run the new-city generation task without giving the generator the target city's origin–destination pairs (origins only, or no demand information) and measure the macro Jensen–Shannon divergences; if GTG's advantage over the baselines collapses, the claimed generalization depends on test-set travel demand rather than on learned invariant mobility patterns.","tokens_in":19602,"feed_emoji":"🚕","tokens_out":15464,"duration_ms":146830,"temperature":0.7,"pith_summary":"This paper tries to establish that mobility patterns learned in one city transfer to trajectory generation in another: a model trained on trajectories from a source city can produce realistic routes in a target city that has no historical trajectories, given only that city's road network and origin–destination demands. The authors' load-bearing insight is that two patterns are invariant across cities: people generally choose paths near the minimal travel cost, and a road's travel cost reflects its topological role in the local network. The GTG model implements this by extracting Space Syntax topological features, learning a city-invariant road representation through disentangled adversarial training, and generating each trajectory as a shortest path under learned preference weights. On Beijing, Xi'an, and Chengdu, the paper reports zero-shot generations that match real data more closely than knowledge-driven and data-driven baselines on both city-scale distributions and per-trip distances. If this holds, trajectory generation no longer requires the very trajectory data that privacy rules and commercial conflicts make hard to collect.","feed_headline":"Trains in one city, then writes another's routes with no local records","feed_subtitle":"By treating travel preferences as city-invariant, it beats per-city baselines on real Beijing, Xi'an, and Chengdu data.","key_machinery":"Three coupled pieces carry the argument. First, a topological feature extractor computes Space Syntax indices — total depth, integration, connectivity, and choice (betweenness) — for each road segment and aggregates them with a spatial-aware graph attention network, producing a road representation meant to be comparable across cities. Second, a disentangled domain-adaptation module splits each representation into a semantic code and a city code, using gradient-reversal adversarial training and an orthogonality loss so the semantic code retains cost-relevant information while becoming insensitive to city identity. Third, a preference-learning module assigns each road segment a combined weight of predicted travel time/speed and a learned hidden cost, then updates those weights so that shortest-path searches under them match real source trajectories; the paper notes this is equivalent to maximum-entropy inverse reinforcement learning. The generator is shortest-path search on the target road network using the transferred preference weights.","core_discovery":"The central claim is that cross-city trajectory generation works because the map from road topology to travel cost, and the preference for low-cost paths, are invariant across urban environments. The model represents each road segment by Space Syntax indices plus basic attributes, aggregates these through a spatial-aware graph attention network, and splits the result into a semantic code and a city-domain code via adversarial disentanglement. Travel costs are predicted from the semantic code as a learned combination of time, speed, and a hidden cost, and the model then learns preference weights by adjusting them until shortest paths between observed origin–destination pairs reproduce the source city's real trajectories. At test time the preference weights plus the target road network drive shortest-path generation; the paper reports that this procedure, with no training on target-city trajectories, beats every baseline on all macro and micro metrics in all six cross-city direction pairs.","pith_inferences":["Outside the paper's protocol, a stronger generalization test would withhold the target city's origin–destination pairs from the generator; the paper's micro metrics pair generated and real trips with the same demand, so the 'no trajectory data' claim currently leans on treating demand as non-trajectory information.","Not reported in the paper, a natural extension is to run the same model across transport modes (bike, ride-hail, pedestrian) to see whether the invariant preference pattern is a property of cities or of a particular mode.","A direct probe the paper does not report: train the cost predictor only on the source city, then measure how well its predicted target-road costs match target-road travel times and speeds; that would isolate the cost-invariance component of the gains."],"forward_implications":["A city with no collected trajectories can receive synthetic route data as soon as its road network is known, provided origin–destination demands are specified.","The reported macro results imply generated datasets reproduce city-scale patterns—trip distance, radius of gyration, and road-visit frequency—far better than random-walk, EPR, or deep generative baselines.","Synthetic GTG trajectories can pre-train next-location prediction models to within a few points of real-data training, so they can stand in for private or commercially restricted data.","The fine-tuning experiments show that as small amounts of target-city trajectories accumulate, all metrics improve further, making the method useful in a staged data-collection rollout."],"supporting_citations":[{"why":"Defines the four Space Syntax measures used to make road segments comparable across cities.","marker":"Hillier et al. 1976"},{"why":"Supplies the disentangled semantic/domain representation and adversarial training scheme.","marker":"Cai et al. 2019"},{"why":"Provides the subgraph-sampling algorithm that lets the graph aggregator train on large road networks.","marker":"Chiang et al. 2019"},{"why":"Supplies the GATv2 attention layer the spatial-aware aggregator is built on.","marker":"Brody, Alon, and Yahav 2022"},{"why":"Gives the maximum-entropy inverse-reinforcement-learning interpretation that motivates the preference-update loss.","marker":"Wulfmeier, Ondruska, and Posner 2015"},{"why":"Provides map matching that converts raw GPS traces into road-segment trajectories and travel-cost labels.","marker":"Yang and Gidofalvi 2018"},{"why":"Contributes the rank loss used to keep travel-cost prediction robust across city-specific biases.","marker":"Burges et al. 2005"},{"why":"Defines one of the knowledge-driven baselines (Density-EPR) that the method is compared against.","marker":"Pappalardo et al. 2015"}],"fun_headline_variants":["Zero-shot trajectory generation for new cities via invariant travel costs","Learn routes in one city, generate them in another with no local data","Cross-city trajectory model transfers via invariant travel-cost logic","No target-city training? GTG still generates routes that match real movement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that feeding the generator the target city's true origin–destination pairs, drawn from the same test trajectories used for evaluation, does not count as supplying target-city trajectory data, and that comparing against baselines that receive less (or none) of that demand information is still a fair test of generalization.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot trajectory generation for new cities via invariant travel costs","Learn routes in one city, generate them in another with no local data","Cross-city trajectory model transfers via invariant travel-cost logic","No target-city training? GTG still generates routes that match real movement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001011,"raw_usage":{"total_tokens":4260,"prompt_tokens":922,"completion_tokens":3338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":3266}},"tokens_in":538,"tokens_out":3338,"duration_ms":26876,"temperature":1.0,"reasoning_tokens":3266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:33:49.560092+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the new-city generation task without giving the generator the target city's origin–destination pairs (origins only, or no demand information) and measure the macro Jensen–Shannon divergences; if GTG's advantage over the baselines collapses, the claimed generalization depends on test-set travel demand rather than on learned invariant mobility patterns.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the four Space Syntax measures used to make road segments comparable across cities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the disentangled semantic/domain representation and adversarial training scheme."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the subgraph-sampling algorithm that lets the graph aggregator train on large road networks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the GATv2 attention layer the spatial-aware aggregator is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides map matching that converts raw GPS traces into road-segment trajectories and travel-cost labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the rank loss used to keep travel-cost prediction robust across city-specific biases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines one of the knowledge-driven baselines (Density-EPR) that the method is compared against."}],"review_version":1}