{"id":"148c7cf2-0773-4320-8167-6ffaab886082","arxiv_id":"2412.05000","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CoDiffMob injects city-level flow statistics and hourly movement rhythms into a diffusion model's starting noise, improving reported trajectory and flow fidelity, though the evaluation is partly circular.","lead":"This paper proposes CoDiffMob, a diffusion model that replaces random i.i.d. noise with collaborative noise priors built from city-level flow statistics, for generating synthetic urban mobility trajectories. The authors report strong fidelity gains over baselines, but the evaluation is confounded because the same aggregate statistics used to build the priors are also the evaluation targets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central collective-flow claim is likely compromised by evaluation leakage: the noise prior is built from the same flow matrix F and population P that later define the CPC/MAPE targets, so the reported 32% advantage over baselines does not measure generative quality as claimed.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing flaw: the collaborative noise prior is constructed from the same F and P that define the evaluation targets, and the paper never establishes that F comes from a held-out or external source. My reading of the full text confirms this: Algorithm 1 explicitly requires F and P; Eq. 7 samples transitions from \\tilde{F}_{l_{i-1}}; Eq. 8 maps those transitions into noise; Eq. 9 uses the real temporal rhythm; and Eqs. 15-16 compare generated flows against the real flow matrix. The flow metrics are therefore measuring how faithfully the model can re-emit statistics that were injected as input, rather than whether the diffusion model learned collective mobility patterns. The ablation study in Table 2 actually strengthens this concern: removing noise fusion gives CPC 0.7165 versus 0.6842 for CoDiffMob on ISP, meaning the rule-based flow prior alone is responsible for the high flow similarity, while the full model trades some aggregate fidelity for better individual metrics. The individual metrics are a fairer test of generative quality, and CoDiffMob does show gains there, but those gains are not the basis for the advertised 'over 32%' improvement and they lack error bars. A corrected evaluation with held-out flows, or with baselines granted the same aggregate side information, would be needed to know whether the method contributes beyond a conditional downscaling procedure. Because the reader already rejected the paper on this basis, my independent stress test does not change the verdict; it reinforces it.","tokens_in":18016,"tokens_out":3312,"duration_ms":34775,"concrete_test":"Split each dataset temporally or by user: compute F, P, and P_t^move on a training portion (e.g., days 1-5) and evaluate CPC/MAPE against the aggregate flow matrix of a disjoint held-out portion (e.g., days 6-7), or against an independently collected OD matrix for the same city. If CoDiffMob's CPC/MAPE advantage over DiffTraj substantially shrinks or disappears under this split, the reported collective-flow gains are an artifact of using the evaluation target as input. Report results across at least 5 sampling seeds with standard deviations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise of the 32% claim is that CoDiffMob learns to reproduce collective OD flows. In Section 4.2 (Eq. 7) and Algorithm 1, the noise prior is constructed from the true collective flow matrix F and population P: home locations are initialized from P, next locations are sampled from \\tilde{F}_{l_{i-1}} during exploration, and the resulting location sequence is mapped to noise via Inverse-DDIM (Eq. 8). Rhythmic BatchNorm (Eq. 9) further shapes noise variance using P_t^move computed from the real dataset (Eq. 5). The evaluation in Eqs. 15-16 then computes CPC/MAPE between the generated aggregate flow F_Y and the same real F_X that was used to construct the prior. High flow similarity is therefore at least partly a reconstruction of input statistics, not evidence that the diffusion model captured collective dynamics. Baselines receive no such target statistics, so the comparison is unfair. The individual trajectory metrics (Radius, Distance, Duration, DailyLoc) are less affected by this leakage, but the headline improvement is dominated by the flow metrics, and the downstream utility claim inherits the same pipeline. As reported, the central claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoDiffMob, a diffusion model for urban mobility trajectory generation that constructs a 'collaborative noise prior' by combining an EPR-based rule-based sampler with the dataset's collective flow matrix F, fusing the resulting informative noise with i.i.d. Gaussian noise, and applying a rhythmic batch normalization based on the dataset's hourly movement probability. The authors claim that this prior improves both individual trajectory realism and collective flow reproduction, reporting an average improvement of over 32% over six baselines, plus privacy and downstream-utility experiments. The paper includes motivation based on observed correlations between initial noise and generated trajectories, and it releases code and data.","tokens_in":18274,"tokens_out":4127,"duration_ms":57683,"significance":"The idea of injecting collective mobility statistics into the noise space of a diffusion model is timely and potentially useful for synthetic mobility generation, especially for urban computing and privacy-preserving data sharing. The manuscript's strengths include a clear problem formulation, a concrete algorithmic specification (Algorithm 1), public code and data, and an explicit privacy evaluation. However, the central quantitative claim—the >32% improvement, driven mainly by flow metrics—is currently undermined by evaluation leakage: the same flow matrix F used to construct the noise prior is also the target of the CPC/MAPE metrics. The ablation study further shows that the rule-based flow sampler alone achieves better flow metrics than the full diffusion model, so the reported flow improvements do not demonstrate that the diffusion model learned collective dynamics. If the evaluation is repaired with held-out flow statistics, proper baselines, and uncertainty quantification, the contribution could be solid, but as presented the headline claim is not established.","major_comments":[{"comment":"The evaluation of collective-flow quality is circular. The collaborative noise prior is built from the real dataset's flow matrix F and population P (Eq. 7, Algorithm 1), and the CPC/MAPE metrics measure agreement between the generated aggregate flow and that same F (Eqs. 15-16). The reported >35% flow improvement over baselines therefore partly reflects direct injection of the target statistics into the sampling process rather than the diffusion model's learned representation. Please re-run the flow evaluation with a held-out temporal or spatial split to compute F, and give baselines access to the same flow statistics, or compare against a baseline that also samples transitions directly from F.","section":"§4.2, Algorithm 1, and §5.2, Eqs. (15)-(16)"},{"comment":"The ablation results undermine the claim that the diffusion model captures collective patterns. On ISP, the 'w/o Noise Fusion' variant—which relies only on the collaborative transition sampling process—achieves CPC=0.7165 and MAPE=0.5351, which are better than full CoDiffMob's CPC=0.6842 and MAPE=0.6361. Thus the rule-based flow prior alone reproduces the flow matrix more accurately than the full model, and the improvement over baselines is attributable to direct sampling from F rather than to the learned denoiser. Please report flow metrics for the pure rule-based sampler and for the diffusion model alone, and discuss the resulting trade-off explicitly.","section":"§5.5, Table 2"},{"comment":"Individual trajectory metrics (Radius, Distance, Duration, DailyLoc) are reported as single Kolmogorov-Smirnov statistics without confidence intervals, sample sizes, or statistical significance tests. Several apparent improvements over DiffTraj are small (e.g., ISP Distance 0.1203 vs. 0.1662; MME Duration 0.0796 vs. 0.0949). Without uncertainty quantification across multiple generation seeds, the claim of consistent superiority on individual-level distributions is not supported. Please provide error bars and significance tests, such as paired bootstrap or repeated KS tests.","section":"§5.2, Table 1"},{"comment":"The downstream utility experiment shows accuracy values on a 0-0.3 scale, and the differences between Real+DiffTraj and Real+CoDiffMob are small. No error bars or significance tests are reported. The conclusion that CoDiffMob 'leads to more significant improvements in downstream task performance' is therefore not supported by the evidence as presented. Please report mean and standard deviation over repeated runs and a statistical test of the difference.","section":"§5.4, Figure 6"}],"minor_comments":[{"comment":"The sentence 'In this paper, we propose CoDiffMob, a diffusion model for urban mobility generation with collaborative noise priors, we emphasize the critical role of noise' is a comma splice; consider '...collaborative noise priors. We emphasize...'.","section":"Abstract"},{"comment":"The notation for the noise prior is inconsistent: z_p and z_P are both used, and 'z_i.i.d.' is not typeset consistently. Please unify the subscript and define all symbols.","section":"§4.2.2, Eq. (8)"},{"comment":"Algorithm 1 calls π with additional arguments (n_ω, β1, β2, P, ρ, γ, F) that do not appear in the definition of π in Eq. (7). Please align the notation.","section":"Algorithm 1 and Eq. (7)"},{"comment":"The caption labels panel (c) as 'DiffTraj' and reports CPC=0.3992 for both (a) and (c); the text says 'the image on the right shows the result generated by the best baseline, Difftraj.' This appears to be a typo in the CPC value or the caption.","section":"Figure 3 and surrounding text"},{"comment":"In the definition of N_CC, 'min(FX,FY)' should be the element-wise minimum min(FX_ij, FY_ij) over matrix entries.","section":"Appendix B.2, Eq. (15)"},{"comment":"The statements 'average improvement of over 40%' and 'over 31%' for trajectory metrics are not derivable from Table 1 as presented; please specify the exact aggregation formula and the baseline used for each percentage.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The evaluation leakage is the central issue and must be fixed before the headline claims can be accepted. The manuscript is otherwise clearly written and the code release is a strength. I recommend major revision rather than rejection because the method is well specified and the fix—held-out flow evaluation, fair baselines, and significance testing—is within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take.\n\nWhat's actually new: transferring noise-prior ideas from video diffusion to mobility is a sensible move, and the co-design is concrete. Sample transition sequences with an EPR model conditioned on aggregate flow F, invert those sequences through DDIM to get informative noise, fuse with white noise, then rescale variance by the population's movement probability. That mechanism is not in DiffTraj, TrajGDM, or the video papers. The motivation analysis is also solid and worth a look: noise direction predicts movement direction at R2=0.955, and the temporal rhythm correlation is a nice observation. Code and data are promised.\n\nThe soft spot is the flow evaluation, and it is not minor. The prior is built from F and P of the training set (Eq. 7, Algorithm 1), and CPC/MAPE compare generated flows to the same F (Eqs. 15-16). Baselines are not given those target statistics. So the reported >35% flow improvement largely measures how well the method replays input statistics, not how well it learned collective dynamics. This is an unfair comparison, not a hidden internal contradiction, but it does undermine the headline claim. A corrected evaluation would use a held-out flow matrix as both prior and target, or give baselines the same flow prior, or at least report flow metrics when the prior is ablated. The individual trajectory metrics are less directly affected but have no error bars or significance tests, and the downstream gains in Figure 6 are small and shown without variance.\n\nFor a researcher in mobility generation or diffusion noise priors, this is still a useful read: the method is novel and clearly described, and the evaluation trap is instructive. But I would not take the 32% number at face value, and I would want the flow evaluation redone before trusting the collective-pattern claims.\n\nRecommendation: this deserves a serious referee, not a desk reject. If it came to me fresh, I would ask for major revision focused on the leakage, plus uncertainty quantification on the trajectory metrics.","headline":"Clever noise-prior construction, but the headline flow gains are partly baked in because the same flow matrix that builds the prior is also the evaluation target.","tokens_in":18825,"tokens_out":3127,"would_cite":false,"duration_ms":223561,"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 replacing i.i.d. noise with collaborative noise priors makes diffusion-based urban mobility generation match both individual trajectories and collective flows, with an average improvement of over 32% across six…","keywords":["urban mobility generation","diffusion models","collaborative noise priors","collective flows","synthetic trajectories","privacy-preserving data generation","trajectory generation"],"falsifier":"Build the flow matrix and population distribution for the noise prior from the first half of the observation week, evaluate CPC and MAPE on the second half, and compare with baselines trained the same way. If CoDiffMob's flow metrics drop to baseline levels, the claim that it learns collective mobility patterns would be refuted, since the apparent improvement would come from feeding the test-period statistics into the prior.","tokens_in":17803,"feed_emoji":"🏙️","tokens_out":7327,"duration_ms":73475,"temperature":0.7,"pith_summary":"CoDiffMob is trying to show that the noise a diffusion model starts from is the decisive variable for generating realistic urban mobility, not just the denoiser architecture. The paper argues that standard i.i.d. Gaussian noise ignores the spatiotemporal correlations and social interactions in human movement, so it builds collaborative noise priors that fuse random noise with rule-based transition sequences steered by aggregate flow data. If the central claim is right, synthetic trajectories could replace privacy-sensitive, expensive web-derived mobility data in downstream tasks while preserving both individual-level statistics and city-wide flow patterns. The paper reports that this approach outperforms six baselines by over 32% on average and that the generated data resists membership-inference attacks.","feed_headline":"Noise priors lift mobility generation by over 32 percent","feed_subtitle":"Fusing collective flow patterns into diffusion noise yields synthetic trajectories that match real movement and protect privacy.","key_machinery":"The central object is the collaborative noise prior, defined as $z_P = \\mathrm{RhythmicBN}(z_{\\mathrm{iid}} + z_F)$, where $z_F$ is obtained by mapping a flow-conditioned EPR transition sequence through inverse DDIM, and $z_{\\mathrm{iid}}$ is standard Gaussian noise. The EPR (exploration and preferential return) model supplies individual-level actions — stay, return home, revisit, explore — and the exploration step is redirected by a flow-based policy drawn from the collective origin-destination matrix, so population patterns enter at the sampling level. Rhythmic batch normalization then rescales the per-time-step variance of the noise batch to be proportional to the crowd's moving probability, injecting the city's daily rhythm. The mechanism does the argument's work because DDIM sampling is deterministic: once the prior is fixed, the trajectory is fixed, so embedding collective knowledge into the starting noise is what lets a single denoiser reproduce both individual preferences and aggregate flows.","core_discovery":"The paper's central claim is that the initial noise fed into a DDIM trajectory generator is not a neutral random canvas: it already encodes movement semantics, and therefore the choice of noise is a primary modeling decision for urban mobility generation. Using inverse DDIM on real trajectories, the paper reports that noise-space direction tracks real-world moving direction with R²=0.9552 and that per-time-step noise variance follows the crowd's daily movement probability. On this basis CoDiffMob constructs a collaborative noise prior by sampling location-transition sequences with an EPR-style rule-based model that is steered by the collective flow matrix and population distribution, mapping those sequences into noise space through inverse DDIM, adding white noise, and applying rhythmic batch normalization so that noise variance matches the movement rhythm. Starting denoising from this prior, the model is claimed to generate trajectories that match individual statistics and collective origin-destination flows simultaneously, with average improvement over 32% versus six baselines and with privacy preserved.","pith_inferences":["The paper computes the collective flow matrix from the same real dataset that supplies the evaluation flows; a cleaner test would build the prior from a separate time window or region and evaluate on held-out flows, to isolate how much of the reported improvement is prior injection versus learned generation.","The observed correlation between movement direction and noise-space direction suggests noise-prior editing could become a general steering mechanism for spatiotemporal sequence generation, letting practitioners impose direction or timing constraints by reshaping the starting noise without retraining.","A natural stress test is to feed CoDiffMob a deliberately wrong or corrupted flow matrix and measure whether generated flows move toward that wrong matrix; if they do, the prior dominates generation, which is both a design feature and a risk for misleading synthetic data.","The rhythmic batch-normalization trick points to a cheap way to inject external temporal signals such as weather or special events into any trajectory diffusion model by rescaling per-step noise variance instead of changing the network."],"forward_implications":["Synthetic trajectories generated this way can substitute for real trajectories in training mobility predictors; the paper reports that mixing CoDiffMob output with real data improves accuracy relative to mixing DiffTraj output or using real data alone.","The reported flow metrics (CPC above 0.6 and MAPE below 0.7 on both datasets) mean a single generative model can reproduce aggregate origin-destination structure, not just individual-level statistics.","Privacy checks indicate the model produces novel trajectories rather than memorized ones: over 90% of generated trajectories have no real counterpart with more than 40% overlap, and membership-inference success stays below 0.6.","The ablation results show that removing the noise prior degrades flow fidelity sharply (CPC drops from 0.68 to 0.34 on the ISP dataset), so the prior is the component doing the collective-pattern work."],"supporting_citations":[{"why":"Supplies the UNet-based diffusion denoiser architecture that CoDiffMob adapts and is the strongest baseline it must beat.","marker":"[66]"},{"why":"Defines DDIM, whose deterministic noise-to-trajectory mapping lets the paper invert generated sequences into informative noise.","marker":"[36]"},{"why":"Provides the TimeGeo/EPR rule-based mobility model that the collaborative transition sampling extends with flow information.","marker":"[14]"},{"why":"Provides the exploration and preferential return mechanism used for individual-level location transitions.","marker":"[35]"},{"why":"Defines the CPC flow-similarity metric used to measure collective-pattern fidelity against baselines.","marker":"[34]"},{"why":"Provides the Kolmogorov-Smirnov test used to compare per-trajectory metric distributions between real and synthetic data.","marker":"[24]"},{"why":"Provides the DDPM forward/reverse process that the paper positions as the i.i.d.-noise baseline and whose noise schedule underlies the denoiser.","marker":"[12]"},{"why":"Supplies the preconditioning and noise-level schedule used in training the diffusion denoiser.","marker":"[15]"}],"fun_headline_variants":["Collaborative noise priors lift mobility generation by 32%","Why noise matters: diffusion models for urban mobility","Privacy-preserving synthetic mobility with crafted noise","Noise priors capture crowd flow for realistic trajectories","CoDiffMob: fusing collective patterns into diffusion noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the collective flow matrix and population distribution used to build the noise prior come from the same real trajectories that are later used to compute the flow evaluation metrics; the paper never shows the prior is built from held-out data, so the reported flow improvements may partly reflect the evaluation statistics being fed into the generator.","fun_headline_variants_meta":{"raw":{"variants":["Collaborative noise priors lift mobility generation by 32%","Why noise matters: diffusion models for urban mobility","Privacy-preserving synthetic mobility with crafted noise","Noise priors capture crowd flow for realistic trajectories","CoDiffMob: fusing collective patterns into diffusion noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1651,"prompt_tokens":1021,"completion_tokens":630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":553}},"tokens_in":637,"tokens_out":630,"duration_ms":7402,"temperature":1.0,"reasoning_tokens":553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:59:50.349152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build the flow matrix and population distribution for the noise prior from the first half of the observation week, evaluate CPC and MAPE on the second half, and compare with baselines trained the same way. If CoDiffMob's flow metrics drop to baseline levels, the claim that it learns collective mobility patterns would be refuted, since the apparent improvement would come from feeding the test-period statistics into the prior.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the UNet-based diffusion denoiser architecture that CoDiffMob adapts and is the strongest baseline it must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TimeGeo/EPR rule-based mobility model that the collaborative transition sampling extends with flow information."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the exploration and preferential return mechanism used for individual-level location transitions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the CPC flow-similarity metric used to measure collective-pattern fidelity against baselines."}],"review_version":1}