Pith. sign in

REVIEW 2 major objections 6 minor 56 references

EventOD claims that an LLM's ternary direction signals, scaled by two lightweight networks, can steer a frozen pretrained graph-diffusion OD generator to produce event-conditioned mobility flows without event-specific retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 09:46 UTC pith:KGRWEBXL

load-bearing objection EventOD is a credible, well-ablated method for adapting frozen OD generators to event conditions, but the paper's headline sparse-supervision claim is not actually tested. the 2 major comments →

arxiv 2607.22655 v1 pith:KGRWEBXL submitted 2026-06-26 cs.AI

EventOD: Event-Aware OD Flow Generation via LLM-Guided Semantic Modulation

Jie Zhao , Jie Feng , Can Rong , Zhihan Hou , Peng Lu , Yong Li This is my paper
classification cs.AI
keywords origin-destination flow generationevent-aware mobilitylarge language modelsgraph diffusion modelssemantic modulationdisaster responsehurricane mobilitypandemic mobility
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that event-aware origin–destination (OD) flow generation does not require retraining a generator for every disaster. Its proposed framework, EventOD, extracts per-region ternary control vectors from coarse event signals using a large language model, learns continuous magnitudes for those directions with two small MLPs (AlphaNet and BetaNet), and injects the modulated features into a frozen pretrained graph-diffusion OD generator via multiplicative input-level modulation. On hurricane and pandemic county-level benchmarks, EventOD reports consistent gains in reconstruction accuracy and distributional fidelity over strong baselines — for example, CPC 0.401 versus 0.315 for the fine-tuned diffusion baseline, with JSD-OD dropping from 0.270 to 0.199 — while keeping the generator parameters fixed. The paper also provides a local-stability bound showing that the input modulation perturbs each reverse diffusion step in a controlled way, and a retrieval-augmented fallback (AlphaKnow/BetaKnow) for sparse-supervision settings. If the claim holds, rapid event adaptation of existing mobility models becomes feasible with modest per-region cost and no architecture-specific retraining.

Core claim

The central claim is that disruptive events change mobility by shifting regional functions and demographics before those shifts are visible in OD flows, so adaptation can be performed at the feature level. EventOD operationalizes this by having an LLM emit, for each region, a POI control vector and a demographic control vector with entries in {−1,0,1} that encode the expected direction of change; AlphaNet and BetaNet then learn the magnitude of each shift; the adjusted features are computed as p̃ = p ⊙ exp(α ⊙ v_p) and d̃ = d ⊙ exp(β ⊙ v_d). These event-conditioned features condition a frozen pretrained graph-diffusion generator, and only the two lightweight networks are trained. The paper's

What carries the argument

The load-bearing mechanism is the log-space multiplicative modulation identity: p̃_i = p_i ⊙ exp(α_i ⊙ v_i^(p)) and d̃_i = d_i ⊙ exp(β_i ⊙ v_i^(d)). Here v are discrete LLM-derived ternary direction vectors, and α, β are tanh-bounded magnitudes learned by two small MLPs; the exp form keeps features non-negative and guarantees bounded perturbation. The paper couples this with a local-stability proposition bounding the deviation of one reverse diffusion step by κ_t L_X ||X̃−X|| under a Lipschitz assumption on the noise-prediction network, showing that semantic modulation only mildly perturbs the frozen generator's denoising dynamics. This mechanism is what lets EventOD treat event adaptation a

Load-bearing premise

The LLM must give the correct sign of the event's effect on each POI category and demographic group in each region; because the adjustment multiplies features by exp(α ⊙ v), a wrong direction can be made larger or smaller but never reversed by the learned magnitudes.

What would settle it

Take EventOD on a fresh event type and flip the sign of a randomized subset of the LLM's ternary control entries; if the flipped variant performs as well as the original, the semantic direction is not what drives the improvement — or, more directly, compare against a variant that learns unrestricted signs (no ternary v); if it beats EventOD on regions where the LLM's signs disagree with observed flow changes, the directional-lock assumption is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If EventOD is correct, OD generators pretrained on routine mobility can be repurposed for new disruptive events in hours, without collecting large event-time OD matrices or fine-tuning heavy generative models.
  • The same input-level modulation interface can be attached to different pretrained generators (diffusion or GAN), so one adaptation module can serve multiple backbones.
  • The retrieval-augmented fallback provides a route to event adaptation when event-time supervision is scarce, since it grounds LLM estimates in training-time evidence without test-time leakage.
  • The local stability bound implies the adapted generator stays close to the pretrained prior, which matters for structural fidelity: event flows should look like perturbed routine flows, not arbitrary reconstructions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The framework's separation of directional semantics (LLM) from magnitude calibration (learned factors) suggests a general recipe for adapting frozen generative models to out-of-distribution conditions in other structured prediction tasks, such as traffic demand, energy load, or epidemic spread, whenever an LLM can articulate the direction of change.
  • Editorial inference: Because the modulation is multiplicative in exp(α ⊙ v), a wrong sign in v cannot be corrected by the learned magnitude. A direct testable extension is to allow a learned sign flip (e.g., replace the ternary v with a continuous unconstrained direction) and compare on a novel event type; if the sign-free variant wins, the LLM's directional assumption is the limiting factor.
  • Editorial inference: The paper evaluates two event types (hurricane and pandemic); the framework's generality could be tested on a continuous spectrum of disruptions — heat waves, floods, strikes, large gatherings — where the LLM's event semantics and the learned magnitudes may interact differently.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper proposes EventOD, a framework for adapting a frozen pretrained OD generation model to disruptive events by modulating its input features. An LLM produces ternary control vectors for POI and demographic dimensions; two lightweight MLPs, AlphaNet and BetaNet, learn continuous scaling factors; Eqs. (4) and (6) apply log-space multiplicative modulation. The modules are trained with the generator frozen on event-time OD matrices via Eq. (9). A retrieval-augmented LLM pathway is proposed as a fallback for sparse supervision. Experiments on Hurricane Dorian and COVID-19 county-level tract OD data compare against classical and deep baselines, with ablations, LLM backbone/control-strategy comparisons, a ControlNet-style comparison, and portability to a GAN backbone. The paper claims consistent accuracy/distributional improvements and robustness under limited supervision.

Significance. If the claims hold, EventOD offers a modular, interpretable alternative to fine-tuning pretrained OD generators, with the practical appeal of no generator-parameter updates and portability across architectural families. The paper's strengths are the clean separation of semantic direction from learned magnitude, the inclusion of random/perturbation controls, the component ablations, and the commitment to public code. The central limited-supervision motivation, however, is not adequately validated by the current experiments, and the quantitative comparisons lack error bars.

major comments (2)
  1. [V-A4, V-C, VI] The central 'limited supervision' claim is not tested by the experimental protocol. The hurricane benchmark uses a random 0.8/0.1/0.1 county split (147 training counties) and the pandemic setting uses 160 training/validation counties, each with a complete tract-level event OD matrix as supervision. No experiment varies the amount of supervision (e.g., 5% or 10% of counties, or partial OD observations), and the RAG fallback introduced for 'scenarios with sparse supervision' (Section IV-C) is evaluated only under this abundant split in Table III, where it underperforms the learned modules. Therefore the conclusion in Section VI that EventOD 'remains robust under limited supervision' is unsupported. Please add supervision-ablation curves comparing learned modules vs. the RAG fallback.
  2. [Tables II-VII] All main results are single-run point estimates with no standard deviations, confidence intervals, or significance tests. Several head-to-head differences are small—e.g., Table V: EventOD vs. ControlNet-WeDAN, CPC 0.401 vs 0.391 and JSD-OD 0.199 vs 0.186; Table IV: EventOD α,β vs β, CPC 0.468 vs 0.465. Without repeated-seed variance or paired tests, the claim of 'consistent improvement' is not statistically grounded. Figure 6 repeats each setting five times but does not report the spread; please add error bars/shaded intervals and, where feasible, paired significance tests.
minor comments (6)
  1. [Table IV] Typesetting: entries '0.2320.385' and '0.2290.391' are missing spaces and should read '0.232 0.385' and '0.229 0.391'.
  2. [IV-D] The headings '2) Stability of Semantic Modulation:' and '3) Local Stability of Semantic Modulation' are duplicated/misnumbered; the first appears to be a leftover. Please renumber and unify.
  3. [Abstract/Introduction] The phrase 'without event-specific retraining' is imprecise because AlphaNet and BetaNet are trained on event-time OD labels. Recommend consistently wording the claim as 'without updating generator parameters'.
  4. [Figure 6] Since each split is repeated five times, the figure should show error bars or shaded intervals; otherwise the claimed 'smooth degradation' is difficult to assess.
  5. [IV-D3] Proposition 1 bounds only one reverse diffusion step. The text is appropriately careful to call this 'local', but a sentence noting explicitly that the bound does not extend to the full 1000-step sampling trajectory would prevent over-reading.
  6. [Eqs. (3), (5)] The sign of α_i and β_i is not constrained (s tanh(z) can be negative), so a wrong ternary direction can in principle be compensated by the learned modules. This is not an error, but stating it explicitly would clarify one possible misreading.

Circularity Check

0 steps flagged

No significant circularity: EventOD's reported gains are held-out supervised predictions, not a renaming of fitted inputs or a self-citation tautology.

full rationale

EventOD's pipeline is not circular. The LLM control vectors in Eqs. (1)-(2) are semantic inputs derived from event context, POIs, and demographics; AlphaNet/BetaNet magnitudes are fit by minimizing the MSE in Eq. (9) against ground-truth event OD matrices on training counties, and the evaluation uses county-level held-out splits. The reported CPC/JSD improvements are therefore genuine held-out predictions. The RAG banks in Eqs. (10)-(11) are constructed only from training-set instances, so no test-time leakage or target-derived retrieval occurs. The only salient self-citation is the WeDAN backbone [11], which is a modeling choice rather than a load-bearing evidential premise; Section V-E2 confirms the same modulation improves NetGAN, so the result is not forced by a self-cited uniqueness or ansatz chain. Proposition 1 is a straightforward consequence of Assumption 1 and bounded tanh modulation, not a conclusion that assumes its claim. The paper's 'limited supervision' wording is a validation-gap concern--the hurricane split uses 80% of counties with dense OD labels, and the sparse-supervision fallback is not tested under genuinely sparse supervision--but this is a correctness/evidence issue, not an equation-level circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The burden is mostly supervised fitting: two small MLPs are trained on event OD data to set magnitudes, and the LLM supplies directions. No new physical entities or parameters beyond these; however, the claim of event-adaptive behavior under sparse supervision depends on having enough event-time OD labels to train AlphaNet/BetaNet, and on the correctness of LLM directions.

free parameters (4)
  • AlphaNet parameters Theta_alpha and BetaNet parameters Theta_beta = learned on D_train; not reported
    The MLP weights map regional POI/demographic features and event descriptors to control magnitudes; they are the main fitted component and are trained with Eq. (9).
  • s_alpha, s_beta (learnable scaling scalars) = not reported
    Bounds on control factors in Eqs. (3) and (5); learned end-to-end and central to the bounded-perturbation stability argument.
  • Retrieval top-k = not reported
    k for FAISS retrieval in Eq. (10) is unspecified; affects RAG fallback behavior.
  • Hurricane county selection thresholds (precipitation/wind speed, convex hull expansion) = not reported
    Used to define the 184-county study area in Section V.A.1; hand-chosen data filtering that shapes the benchmark.
axioms (5)
  • domain assumption epsilon_theta is Lipschitz continuous in the conditioning features X (Assumption 1, Eq. 17).
    Unproved regularity assumption; the stability proposition rests on it. Standard for diffusion models but not verified for the trained WeDAN network.
  • domain assumption Routine POI/demographic attributes plus coarse event context delta_i are sufficient to determine the event-conditioned OD matrix.
    Problem statement in Section III-B defines the mapping F; if the event signal is too coarse, the whole input-level adaptation cannot recover OD changes.
  • domain assumption LLM ternary control vectors v_p, v_d encode the correct sign of event-induced changes.
    Used in Eqs. (4) and (6); since modulation is multiplicative exp(alpha⊙v), a wrong sign cannot be corrected by the learned magnitude.
  • domain assumption Pretrained WeDAN generator trained on normal-day data transfers to event-time conditioning via input changes only.
    Frozen-generator design in Section IV-D assumes features that were never seen in pretraining can still produce sensible OD outputs.
  • domain assumption County-level random split and aggregation over the event window produce representative, independent evaluation samples.
    Section V.A.4 uses 0.8/0.1/0.1 random split; counties in the same hurricane may be spatially correlated, so independence is questionable.

pith-pipeline@v1.3.0-alltime-deepseek · 19144 in / 14490 out tokens · 136715 ms · 2026-08-02T09:46:51.692093+00:00 · methodology

0 comments
read the original abstract

Estimating origin-destination (OD) flows under disruptive events is important for disaster response and urban resilience. Existing deep OD models trained on routine mobility often degrade when extreme events abruptly alter regional functions and population activities, while retraining a new generator for each event is impractical under limited event-time supervision. We propose EventOD, an event-adaptive OD generation framework that steers a pretrained OD generator using structured event semantics. EventOD first uses a large language model to infer region-level functional and demographic control vectors from coarse event observations. It then learns two lightweight adaptation modules, AlphaNet and BetaNet, to calibrate the magnitude of these semantic shifts, and further introduces a retrieval-augmented fallback pathway for scenarios with sparse supervision. The resulting event-conditioned features are injected into a pretrained graph diffusion OD model through input-level modulation, enabling event-aware adaptation without updating generator parameters. Experiments on hurricane- and pandemic-induced mobility across U.S. counties show that EventOD consistently improves both reconstruction accuracy and distributional fidelity over strong baselines. Source code is available at https://anonymous.4open.science/r/EventOD-5C11/.

Figures

Figures reproduced from arXiv: 2607.22655 by Can Rong, Jie Feng, Jie Zhao, Peng Lu, Yong Li, Zhihan Hou.

Figure 1
Figure 1. Figure 1: Overall framework of EventOD. which captures demographic and socioeconomic behavioral adjustments. These two dimensions represent the supply and demand sides of urban mobility under external disruptions. We employ large language models as structured semantic extractors to infer discrete control signals that encode the direction of change along each dimension. For each region, the LLM produces a POI control… view at source ↗
Figure 2
Figure 2. Figure 2: Generation of POI and demographic control vector. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Optimization Process of AlphaNet and BetaNet. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pipeline of LLM RAG based factor learning. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Dataset partitioning and workflow. A pretrained OD generator [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity of EventOD to increasingly challenging test splits in the [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance comparison of different LLM backbones in the primary [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance under different semantic control strategies. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 7 linked inside Pith

  1. [1]

    Human mobility data and analysis for urban resilience: A systematic review,

    M. Haraguchi, A. Nishino, A. Kodaka, M. Allaire, U. Lall, L. Kuei-Hsien, K. Onda, K. Tsubouchi, and N. Kohtake, “Human mobility data and analysis for urban resilience: A systematic review,”Environment and Planning B: Urban Analytics and City Science, vol. 49, no. 5, pp. 1507–1535, 2022

    2022

  2. [2]

    Transportation resilience to climate 13 change and extreme weather events–beyond risk and ro- bustness,

    S. A. Markolf, C. Hoehne, A. Fraser, M. V . Chester, and B. S. Underwood, “Transportation resilience to climate 13 change and extreme weather events–beyond risk and ro- bustness,”Transport policy, vol. 74, pp. 174–186, 2019

    2019

  3. [3]

    2F-TP: Learning flexible spatiotemporal dependency for flexible traffic predic- tion,

    J. Zhao, C. Chen, C. Liao, H. Huang, J. Ma, H. Pu, J. Luo, T. Zhu, and S. Wang, “2F-TP: Learning flexible spatiotemporal dependency for flexible traffic predic- tion,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 12, pp. 15 379–15 391, 2022

    2022

  4. [4]

    The effect of human mobility and control measures on the covid-19 epidemic in china,

    M. U. Kraemer, C.-H. Yang, B. Gutierrez, C.-H. Wu, B. Klein, D. M. Pigott, O. C.-. D. W. Group†, L. Du Plessis, N. R. Faria, R. Liet al., “The effect of human mobility and control measures on the covid-19 epidemic in china,”Science, vol. 368, no. 6490, pp. 493– 497, 2020

    2020

  5. [5]

    Citygpt: Empowering urban spatial cognition of large language models,

    J. Feng, T. Liu, Y . Du, S. Guo, Y . Lin, and Y . Li, “Citygpt: Empowering urban spatial cognition of large language models,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, 2025, pp. 591–602

    2025

  6. [6]

    Urban planning in the era of large language models,

    Y . Zheng, F. Xu, Y . Lin, P. Santi, C. Ratti, Q. R. Wang, and Y . Li, “Urban planning in the era of large language models,”Nature computational science, pp. 1–10, 2025

    2025

  7. [7]

    An interdisciplinary survey on origin-destination flows modeling: Theory and tech- niques,

    C. Rong, J. Ding, and Y . Li, “An interdisciplinary survey on origin-destination flows modeling: Theory and tech- niques,”ACM Computing Surveys, vol. 57, no. 1, pp. 1–49, 2024

    2024

  8. [8]

    Event-aware multimodal mobility nowcasting,

    Z. Wang, R. Jiang, H. Xue, F. D. Salim, X. Song, and R. Shibasaki, “Event-aware multimodal mobility nowcasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 4, 2022, pp. 4228– 4236

    2022

  9. [9]

    Eventtsf: Event-aware non-stationary time series forecasting,

    Y . Ge, M. Jin, Y . Zhao, H. Li, B. Du, C. Xu, and S. Pan, “Eventtsf: Event-aware non-stationary time series forecasting,”arXiv preprint arXiv:2508.13434, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Dynamic origin- destination flow imputation using feature-based transfer learning,

    P. Chen, Z. Wang, B. Zhou, and G. Yu, “Dynamic origin- destination flow imputation using feature-based transfer learning,”IEEE Transactions on Intelligent Transporta- tion Systems, 2024

    2024

  11. [11]

    A large-scale dataset and benchmark for commuting origin-destination flow generation,

    C. Rong, J. Ding, Y . Liu, and Y . Li, “A large-scale dataset and benchmark for commuting origin-destination flow generation,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  12. [12]

    Generating sparse origin–destination flows on shared mobility networks using probabilistic graph neural networks,

    Y . Liang, Z. Zhao, and C. Webster, “Generating sparse origin–destination flows on shared mobility networks using probabilistic graph neural networks,”Sustainable Cities and Society, vol. 114, p. 105777, 2024

    2024

  13. [13]

    Cross city traffic flow generation via retrieval augmented diffu- sion model,

    Y . Li, J. Wang, X. Yu, P. Wang, and Q. Huang, “Cross city traffic flow generation via retrieval augmented diffu- sion model,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  14. [14]

    Semob: Semantic synthesis for dynamic urban mobility prediction,

    R. Chen, S. Jiang, and W. Huang, “Semob: Semantic synthesis for dynamic urban mobility prediction,” in Proceedings of the 2025 Conference on Empirical Meth- ods in Natural Language Processing, 2025, pp. 15 346– 15 366

    2025

  15. [15]

    Measuring mobility resilience with network-based simulations of flow dy- namics under extreme events,

    Z. Li, W. Yan, and L. Wang, “Measuring mobility resilience with network-based simulations of flow dy- namics under extreme events,”Transportation Research Part D: Transport and Environment, vol. 135, p. 104362, 2024

    2024

  16. [16]

    Synthetic population: A reliable framework for analysis for agent-based mod- eling in mobility,

    F. Bigi, T. H. Rashidi, and F. Viti, “Synthetic population: A reliable framework for analysis for agent-based mod- eling in mobility,”Transportation Research Record, vol. 2678, no. 11, pp. 1–15, 2024

    2024

  17. [17]

    Daily rhythm of urban space usage: insights from the nexus of urban functions and human mobility,

    F. Du, J. Wang, L. Mao, and J. Kang, “Daily rhythm of urban space usage: insights from the nexus of urban functions and human mobility,”Humanities and Social Sciences Communications, vol. 11, no. 1, pp. 1–10, 2024

    2024

  18. [18]

    Observing the rhythms of daily life: A six-week travel diary,

    K. W. Axhausen, A. Zimmermann, S. Sch ¨onfelder, G. Rindsf ¨user, and T. Haupt, “Observing the rhythms of daily life: A six-week travel diary,”Transportation, vol. 29, no. 2, pp. 95–124, 2002

    2002

  19. [19]

    Development of origin–destination matrices using mobile phone call data,

    M. S. Iqbal, C. F. Choudhury, P. Wang, and M. C. Gonz´alez, “Development of origin–destination matrices using mobile phone call data,”Transportation research part C: emerging technologies, vol. 40, pp. 63–74, 2014

    2014

  20. [20]

    Estimating origin-destination flows using mobile phone location data,

    F. Calabrese, G. Di Lorenzo, L. Liu, and C. Ratti, “Estimating origin-destination flows using mobile phone location data,”IEEE Pervasive Computing, vol. 10, no. 4, pp. 36–44, 2011

    2011

  21. [21]

    Travel demand estimation and network assignment based on cellular network data,

    D. Gundleg ˚ard, C. Rydergren, N. Breyer, and B. Ra- jna, “Travel demand estimation and network assignment based on cellular network data,”Computer Communica- tions, vol. 95, pp. 29–42, 2016

    2016

  22. [22]

    Cellular-based data- extracting method for trip distribution,

    C. Pan, J. Lu, S. Di, and B. Ran, “Cellular-based data- extracting method for trip distribution,”Transportation research record, vol. 1945, no. 1, pp. 33–39, 2006

    1945

  23. [23]

    The p 1 p 2/d hypothesis: on the intercity movement of persons,

    G. K. Zipf, “The p 1 p 2/d hypothesis: on the intercity movement of persons,”American sociological review, vol. 11, no. 6, pp. 677–686, 1946

    1946

  24. [24]

    A universal model for mobility and migration patterns,

    F. Simini, M. C. Gonz ´alez, A. Maritan, and A.-L. Barab´asi, “A universal model for mobility and migration patterns,”Nature, vol. 484, no. 7392, pp. 96–100, 2012

    2012

  25. [25]

    Learning geo-contextual embeddings for com- muting flow prediction,

    Z. Liu, F. Miranda, W. Xiong, J. Yang, Q. Wang, and C. Silva, “Learning geo-contextual embeddings for com- muting flow prediction,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 01, 2020, pp. 808–816

    2020

  26. [26]

    Trip distribution modeling with twitter data,

    N. Pourebrahim, S. Sultana, A. Niakanlahiji, and J.- C. Thill, “Trip distribution modeling with twitter data,” Computers, Environment and Urban Systems, vol. 77, p. 101354, 2019

    2019

  27. [27]

    A machine learning ap- proach to modeling human migration,

    C. Robinson and B. Dilkina, “A machine learning ap- proach to modeling human migration,” inProceedings of the 1st ACM SIGCAS Conference on Computing and Sustainable Societies, 2018, pp. 1–8

    2018

  28. [28]

    A deep gravity model for mobility flows generation,

    F. Simini, G. Barlacchi, M. Luca, and L. Pappalardo, “A deep gravity model for mobility flows generation,” Nature communications, vol. 12, no. 1, p. 6576, 2021

    2021

  29. [29]

    Goddag: Generating origin-destination flow for new cities via domain adver- sarial training,

    C. Rong, J. Feng, and J. Ding, “Goddag: Generating origin-destination flow for new cities via domain adver- sarial training,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 10, pp. 10 048–10 057, 2023

    2023

  30. [30]

    Spatio- temporal diffusion model for cellular traffic generation,

    X. Liu, X. Xu, Z. Liu, Z. Li, and K. Wu, “Spatio- temporal diffusion model for cellular traffic generation,” IEEE Transactions on Mobile Computing, 2025. 14

    2025

  31. [31]

    Spatio-temporal knowledge driven diffusion model for mobile traffic generation,

    H. Chai, X. Qi, and Y . Li, “Spatio-temporal knowledge driven diffusion model for mobile traffic generation,” IEEE Transactions on Mobile Computing, 2025

    2025

  32. [32]

    Decad: A decomposition-enhanced diffusion framework for mobile traffic generation from satellite im- agery,

    Y . Xu, M. Jiang, T. Zhu, L. Sun, H. Yu, J. Wang, and H. Sheng, “Decad: A decomposition-enhanced diffusion framework for mobile traffic generation from satellite im- agery,”IEEE Transactions on Mobile Computing, 2026

    2026

  33. [33]

    Exploring large language models for human mobility prediction under public events,

    Y . Liang, Y . Liu, X. Wang, and Z. Zhao, “Exploring large language models for human mobility prediction under public events,”Computers, Environment and Urban Sys- tems, vol. 112, p. 102153, 2024

    2024

  34. [34]

    Event- aware analysis of cross-city visitor flows using large language models and social media data,

    X. Wang, Z. Zhao, R. Wang, and Y . Xu, “Event- aware analysis of cross-city visitor flows using large language models and social media data,”arXiv preprint arXiv:2505.03847, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Causalmob: Causal human mobility prediction with llms-derived human intentions toward public events,

    X. Yang, H. Ge, J. Wang, Z. Fan, R. Jiang, R. Shibasaki, and N. Koshizuka, “Causalmob: Causal human mobility prediction with llms-derived human intentions toward public events,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1, 2025, pp. 1773–1784

    2025

  36. [36]

    Predicting human mobility in disasters via llm-enhanced cross-city learning,

    Y . Tang, H. Wang, X. Fan, and Y . Li, “Predicting human mobility in disasters via llm-enhanced cross-city learning,”arXiv preprint arXiv:2507.19737, 2025

    arXiv 2025

  37. [37]

    Llecat: Llm enhanced causality-aware traffic accidents post-effects prediction,

    X. Yang, Y . Tao, H. Ge, Z. Fan, R. Akerkar, and N. Koshizuka, “Llecat: Llm enhanced causality-aware traffic accidents post-effects prediction,”IEEE Transac- tions on Intelligent Transportation Systems, 2025

    2025

  38. [38]

    Self-correcting llm-controlled diffusion models,

    T.-H. Wu, L. Lian, J. E. Gonzalez, B. Li, and T. Dar- rell, “Self-correcting llm-controlled diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 6327–6336

    2024

  39. [39]

    Mastering text-to-image diffusion: Recaptioning, plan- ning, and generating with multimodal llms

    L. Yang, Z. Yu, C. Meng, M. Xu, S. Ermon, and B. Cui, “Mastering text-to-image diffusion: Recaptioning, plan- ning, and generating with multimodal llms.” inIcml, vol. 3, no. 6, 2024, p. 7

    2024

  40. [40]

    Genartist: Multimodal llm as an agent for unified image generation and editing,

    Z. Wang, A. Li, Z. Li, and X. Liu, “Genartist: Multimodal llm as an agent for unified image generation and editing,” Advances in Neural Information Processing Systems, vol. 37, pp. 128 374–128 395, 2024

    2024

  41. [41]

    Promptenhancer: A simple approach to enhance text-to-image models via chain-of-thought prompt rewriting,

    L. Wang, X. Xing, Y . Cheng, Z. Zhao, D. Li, T. Hang, J. Tao, Q. Wang, R. Li, C. Chenet al., “Promptenhancer: A simple approach to enhance text-to-image models via chain-of-thought prompt rewriting,”arXiv preprint arXiv:2509.04545, 2025

    arXiv 2025

  42. [42]

    Llmcontrol: Grounded control of text-to-image diffusion-based synthesis with multimodal llms,

    J. Wang, R. Chen, and H. Cui, “Llmcontrol: Grounded control of text-to-image diffusion-based synthesis with multimodal llms,”arXiv preprint arXiv:2507.19939, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Multimodal llm-guided semantic correction in text- to-image diffusion,

    Z. Lv, J. Chen, Q. Tian, K. Yin, S. Zhang, and F. Wu, “Multimodal llm-guided semantic correction in text- to-image diffusion,”arXiv preprint arXiv:2505.20053, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Dynamic-i2v: Exploring image- to-video generation models via multimodal llm,

    P. Liu, X. Ren, F. Liu, Q. Xie, Q. Zheng, Y . Zhang, H. Lu, and Y . Yang, “Dynamic-i2v: Exploring image- to-video generation models via multimodal llm,”arXiv preprint arXiv:2505.19901, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Exploring mllm-diffusion information transfer with metacanvas,

    H. Lin, X. Pan, Z. Huang, J. Hou, J. Wang, W. Chen, Z. He, F. Juefei-Xu, J. Sun, Z. Fanet al., “Exploring mllm-diffusion information transfer with metacanvas,” arXiv preprint arXiv:2512.11464, 2025

    arXiv 2025

  46. [46]

    Llplace: The 3d indoor scene layout generation and editing via large language model,

    Y . Yang, J. Lu, Z. Zhao, Z. Luo, J. J. Yu, V . Sanchez, and F. Zheng, “Llplace: The 3d indoor scene layout generation and editing via large language model,”arXiv preprint arXiv:2406.03866, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Programmable-room: Interactive textured 3d room meshes generation empowered by large language models,

    J. Kim, J. Park, K. Kong, and S.-J. Kang, “Programmable-room: Interactive textured 3d room meshes generation empowered by large language models,”IEEE Transactions on Multimedia, vol. 27, pp. 6358–6368, 2025

    2025

  48. [48]

    The faiss library,

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazar ´e, M. Lomeli, L. Hosseini, and H. J ´egou, “The faiss library,”IEEE Transactions on Big Data, 2025

    2025

  49. [49]

    Adding con- ditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding con- ditional control to text-to-image diffusion models,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3836–3847

    2023

  50. [50]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  51. [51]

    Physics-informed neural ode for post-disaster mobility recovery,

    J. Li, H. Wang, and X. Chen, “Physics-informed neural ode for post-disaster mobility recovery,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 1587–1598

    2024

  52. [52]

    Global historical climatology network daily (ghcn- daily),

    NOAA National Centers for Environmental Informa- tion, “Global historical climatology network daily (ghcn- daily),” https://www.ncei.noaa.gov/, 2019, accessed: 2025-10-30

    2019

  53. [53]

    Safegraph data: Point-of-interest and mobility data,

    SafeGraph Inc., “Safegraph data: Point-of-interest and mobility data,” https://www.safegraph.com/, 2020, ac- cessed: 2025-10-30

    2020

  54. [54]

    A county-level dataset for informing the united states’ response to covid-19,

    B. D. Killeen, J. Y . Wu, K. Shah, A. Zapaishchykova, P. Nikutta, A. Tamhane, S. Chakraborty, J. Wei, T. Gao, M. Thieset al., “A county-level dataset for informing the united states’ response to covid-19,”arXiv preprint arXiv:2004.00756, 2020

    Pith/arXiv arXiv 2004

  55. [55]

    Multiscale dynamic human mobility flow dataset in the us during the covid-19 epidemic,

    Y . Kang, S. Gao, Y . Liang, M. Li, J. Rao, and J. Kruse, “Multiscale dynamic human mobility flow dataset in the us during the covid-19 epidemic,”Scientific data, vol. 7, no. 1, p. 390, 2020

    2020

  56. [56]

    Origin–destination matrix es- timation and prediction from socioeconomic variables using automatic feature selection procedure-based ma- chine learning model,

    P. Rodr ´ıguez-Rueda, J. Ruiz-Aguilar, J. Gonz ´alez- Enrique, and I. Turias, “Origin–destination matrix es- timation and prediction from socioeconomic variables using automatic feature selection procedure-based ma- chine learning model,”Journal of urban planning and development, vol. 147, no. 4, p. 04021056, 2021

    2021