arxiv: 2605.10020 · v1 · submitted 2026-05-11 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation

Wilson Wongso , Lihuan Li , Arian Prabowo , Xiachong Lin , Baiyu Chen , Hao Xue , Flora D. Salim

Authors on Pith no claims yet

Pith reviewed 2026-05-12 02:52 UTC · model grok-4.3

classification 💻 cs.LG

keywords trajectory generationdiffusion language modelsroad networkssynthetic data generationzero-shot generalizationmobility trajectoriestopology constraintsblock diffusion

0 comments

The pith

Block diffusion on discrete road segments generates realistic trajectories up to 2.8 times faster than prior topology-aware methods.

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

The paper seeks to show that a diffusion-based language model operating on blocks of discrete road segments can overcome the speed versus accuracy trade-off in trajectory synthesis. Real GPS data is often restricted by privacy, making synthetic alternatives essential for testing urban plans and mobility scenarios. By adding road network structure to the embeddings and sampling, the generated paths stay connected and locally similar to actual ones. This would mean practitioners can produce large volumes of usable data more quickly and apply models across different cities or vehicle types.

Core claim

The central discovery is that trajectories can be generated by first encoding the road network topology into embeddings, then applying block-wise diffusion to denoise sequences of road segments, and finally sampling with topology constraints to ensure the result forms valid paths. This yields strong performance on local similarity metrics on three city datasets, generation speeds up to 2.8 times higher than previous methods, and the ability to transfer zero-shot to new transportation modes and domains.

What carries the argument

Block diffusion language model backbone that processes sequences in parallel blocks, integrated with topology-aware embeddings and constrained sampling to enforce road network coherence.

If this is right

Faster production of synthetic datasets for what-if analyses in transportation and urban planning.
Ability to use a single model for multiple cities and modes through zero-shot transfer.
Better balance of computational efficiency and fidelity to fine-grained trajectory patterns.
Scalable alternative to autoregressive or search-based decoding for graph-structured sequences.

Where Pith is reading between the lines

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

The success of block processing suggests similar efficiency gains could apply to other sequence generation tasks involving discrete graph elements.
If the method generalizes, it could support privacy-preserving data augmentation for machine learning models in mobility prediction.
Extending the constraints beyond topology to include time or speed limits might further improve realism in specific applications.

Load-bearing premise

The premise that discrete road segment sequences, when denoised in blocks and constrained by topology, will avoid artifacts and produce paths as coherent as those from continuous or exhaustive search methods.

What would settle it

A test where the model is applied to a fourth city dataset and the generated trajectories exhibit higher rates of invalid road connections or lower scores on local similarity metrics than reported.

Figures

Figures reproduced from arXiv: 2605.10020 by Arian Prabowo, Baiyu Chen, Flora D. Salim, Hao Xue, Lihuan Li, Wilson Wongso, Xiachong Lin.

**Figure 1.** Figure 1: Proposed TrajDLM architecture. TrajDLM consists of a Block Diffusion Language Model backbone, a Road Network Encoder for topology-aware road embeddings, and a TopologyConstrained Sampling strategy for valid trajectory generation via block-wise denoising. address these limitations, we introduce a topology-constrained sampling strategy that augments the reverse process with ❶ adjacency-aware token restricti… view at source ↗

**Figure 2.** Figure 2: Hausdorff distance versus average generation time per trajectory on Beijing, Porto, and San Francisco. We compare TrajDLM against HOSER, autoregressive (AR), and masked diffusion language model (MDLM) variants based on the same Qwen3-0.6B backbone. (a) DiffTraj (b) HOSER (c) TrajDLM (d) Real [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 4.** Figure 4: Example of the input prompt and target trajectory sequence. The prompt includes the origin and destination road segments, departure time, and trip-level attributes, while the target is a sequence of road segment IDs representing the trajectory to be generated. E Dataset Statistics We train and evaluate TrajDLM on three city-scale trajectory datasets from Beijing, Porto, and San Francisco, following the sam… view at source ↗

**Figure 5.** Figure 5: Heatmap visualizations of generated trajectories across cities. Rows correspond to Beijing, Porto, and San Francisco, while columns compare DiffTraj, HOSER, TrajDLM, and real trajectories. (a) TrajDLM (b) Real [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: Trajectory visualization in Beijing. Green and purple dots indicate the origin and destination, respectively, and the blue line represents the trajectory. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

**Figure 7.** Figure 7: Trajectory visualization in Porto. Green and purple dots indicate the origin and destination, respectively, and the blue line represents the trajectory. (a) TrajDLM (b) Real [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

**Figure 8.** Figure 8: Trajectory visualization in San Francisco. Green and purple dots indicate the origin and destination, respectively, and the blue line represents the trajectory. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

read the original abstract

Generating high-fidelity synthetic GPS trajectories is increasingly important for applications in transportation, urban planning, and what-if scenario simulation, especially as privacy concerns limit access to real-world mobility data. Existing trajectory generation models face a trade-off between efficiency and faithfulness to road network topology: continuous-space methods enable fast generation but ignore the road network, while topology-aware approaches rely on search-based autoregressive decoding that limits generation speed. We propose TrajDLM, a topology-aware trajectory generation framework based on block diffusion language models that bridges this gap. TrajDLM models trajectories as sequences of discrete road segments, combining a block diffusion backbone for efficient denoising, topology-aware embeddings from a road network encoder, and topology-constrained sampling to ensure coherent and realistic trajectories. Across three city-scale datasets, TrajDLM achieves strong performance on fine-grained local similarity metrics while being up to $2.8\times$ faster than prior work, and demonstrates strong zero-shot transfer across domains, including unseen transportation modes. These results highlight the effectiveness of block-wise discrete diffusion as a scalable approach to accurate and efficient trajectory generation. Our code is available at https://github.com/cruiseresearchgroup/TrajDLM/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TrajDLM adapts block diffusion language models to discrete road-segment sequences with topology embeddings and constrained sampling for faster, network-respecting trajectory generation.

read the letter

The main point is that this paper takes block diffusion language models and applies them to trajectories represented as sequences of discrete road segments, adding a graph encoder for topology embeddings and constrained sampling to keep outputs on the actual network. This targets the usual speed-versus-faithfulness trade-off in synthetic mobility data generation. Continuous methods are fast but ignore roads; autoregressive search-based ones respect topology but run slowly. The block-wise denoising plus constraints is the concrete step they take to get both efficiency and structure. They report results on three city-scale datasets with gains on local similarity metrics, up to 2.8 times faster generation, and zero-shot transfer to unseen modes. Code is released, which lets others verify the numbers directly. The approach looks internally consistent and avoids obvious circular fitting or unstated derivations that would undermine the claims. The discrete representation could in principle introduce artifacts that continuous paths avoid, but their local metrics and the stress-test consistency suggest this does not break the results on the tested data. Minor soft spots include limited detail in the abstract on exact baselines and ablations; the full paper presumably supplies those tables. This work is aimed at transportation researchers and urban planners who need large volumes of realistic synthetic trajectories for privacy-safe simulation and planning. Anyone adapting diffusion models to graph-constrained sequences will find the specific integration useful. The method shows clear engagement with the problem and prior baselines, so it deserves a serious referee. I recommend sending it for peer review.

Referee Report

2 major / 3 minor

Summary. The paper introduces TrajDLM, a topology-aware trajectory generation model based on block diffusion language models. Trajectories are represented as discrete sequences of road segments; the architecture combines a block diffusion backbone for efficient parallel denoising, a road network graph encoder for topology-aware embeddings, and constrained sampling during generation to enforce road-network coherence. Empirical evaluation on three city-scale datasets reports competitive or superior performance on fine-grained local similarity metrics, generation speedups of up to 2.8× relative to prior autoregressive or search-based methods, and strong zero-shot transfer to unseen cities and transportation modes. Code is released.

Significance. If the reported gains hold under rigorous scrutiny, the work offers a practical advance in synthetic mobility data generation by reconciling the speed of diffusion-based sampling with topological fidelity. The block-wise discrete diffusion formulation and its integration with graph encoders constitute a reusable architectural pattern for constrained sequence generation. Zero-shot cross-domain transfer and open-source code further increase the potential impact for privacy-preserving urban simulation and transportation research.

major comments (2)

[§4.2, §5.1] §4.2 and §5.1: the topology-constrained sampling procedure is described at a high level but lacks an explicit statement of how the constraint is enforced inside the block-denoising loop (e.g., whether it is a hard mask, a soft penalty, or a post-hoc rejection step). Without this detail it is impossible to determine whether the reported similarity improvements are attributable to the learned model or to the constraint itself.
[Table 2] Table 2 (main results): the 2.8× speedup is reported without accompanying wall-clock timings on identical hardware, batch sizes, or sequence lengths for all baselines. Because generation time in diffusion models depends on the number of denoising steps and block size, the headline efficiency claim cannot be assessed for fairness or reproducibility from the current presentation.

minor comments (3)

[Abstract, §1, §4.1] The abstract and §1 refer to “fine-grained local similarity metrics” without naming them; the precise definitions (e.g., edit distance, Fréchet distance on road-segment sequences, or custom local overlap scores) should appear in §4.1 or an appendix.
[§3.1] §3.1: the notation for block diffusion (e.g., the block size B and the forward-process variance schedule) is introduced without a compact summary equation; adding a single boxed equation would improve readability.
[§5.3] The zero-shot transfer experiments in §5.3 would benefit from an explicit statement of which transportation modes were held out and whether the road-network encoder was frozen or fine-tuned.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment, the recommendation for minor revision, and the constructive comments on methodological clarity and experimental reporting. We address each major comment below and will update the manuscript accordingly to improve reproducibility and transparency.

read point-by-point responses

Referee: [§4.2, §5.1] §4.2 and §5.1: the topology-constrained sampling procedure is described at a high level but lacks an explicit statement of how the constraint is enforced inside the block-denoising loop (e.g., whether it is a hard mask, a soft penalty, or a post-hoc rejection step). Without this detail it is impossible to determine whether the reported similarity improvements are attributable to the learned model or to the constraint itself.

Authors: We agree that the description of the topology-constrained sampling in §4.2 and §5.1 is high-level and would benefit from greater precision. The constraint is enforced as a hard mask inside the block-denoising loop: at each denoising step for a block, the output logits over candidate road segments are element-wise multiplied by a binary adjacency mask derived from the road-network graph encoder, zeroing out probabilities for non-adjacent segments before the softmax and sampling. This mask is recomputed per block using the current partial trajectory prefix and the pre-encoded graph topology. We will revise the manuscript to include this explicit mechanism, a step-by-step description of the loop, and pseudocode in the appendix. This clarification will make clear that the constraint is an integrated component of the generative process rather than a post-processing step. revision: yes
Referee: [Table 2] Table 2 (main results): the 2.8× speedup is reported without accompanying wall-clock timings on identical hardware, batch sizes, or sequence lengths for all baselines. Because generation time in diffusion models depends on the number of denoising steps and block size, the headline efficiency claim cannot be assessed for fairness or reproducibility from the current presentation.

Authors: We acknowledge that the efficiency results in Table 2 would be more convincing with explicit timing details. The 2.8× figure reflects average wall-clock generation time per trajectory measured on the same hardware (NVIDIA A100 80GB GPU) using a batch size of 32 and trajectories of comparable length (approximately 40–60 road segments). TrajDLM uses 100 denoising steps with block size 8, while the autoregressive baselines decode sequentially. We will add a supplementary table listing wall-clock times, hardware, batch sizes, sequence lengths, number of denoising steps, and block sizes for every baseline and dataset to support direct reproducibility and fair comparison. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents TrajDLM as an architectural combination of block diffusion language models, a road network graph encoder for topology-aware embeddings, and topology-constrained sampling during generation. All performance claims (similarity metrics, speedups, zero-shot transfer) are grounded in empirical evaluation on three city-scale datasets rather than any derivation or prediction that reduces to fitted parameters or self-defined quantities by construction. No equations, uniqueness theorems, or ansatzes are invoked that equate outputs to inputs; the method is explicitly a novel synthesis of existing components (diffusion, graph encoding, constrained decoding) with code released for reproduction. This satisfies the criteria for a self-contained, non-circular derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the framework relies on standard diffusion modeling assumptions and road-network graph structure treated as given input.

pith-pipeline@v0.9.0 · 5531 in / 1084 out tokens · 31866 ms · 2026-05-12T02:52:45.384869+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean (J-uniqueness, Aczél classification) washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

TrajDLM models trajectories as sequences of discrete road segments, combining a block diffusion backbone for efficient denoising, topology-aware embeddings from a road network encoder, and topology-constrained sampling
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Block diffusion language model backbone... NELBO objective... topology-constrained sampling

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 5 internal anchors

[1]

Arriola, S

M. Arriola, S. S. Sahoo, A. Gokaslan, Z. Yang, Z. Qi, J. Han, J. T. Chiu, and V . Kuleshov. Block diffusion: Interpolating between autoregressive and diffusion language models. InThe Thirteenth International Conference on Learning Representations, 2025

work page 2025
[2]

Austin, D

J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. Van Den Berg. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

work page 2021
[3]

T. Bie, M. Cao, X. Cao, B. Chen, F. Chen, K. Chen, L. Du, D. Feng, H. Feng, M. Gong, et al. Llada2. 1: Speeding up text diffusion via token editing.arXiv preprint arXiv:2602.08676, 2026

work page arXiv 2026
[4]

T. Bie, M. Cao, K. Chen, L. Du, M. Gong, Z. Gong, Y . Gu, J. Hu, Z. Huang, Z. Lan, et al. Llada2. 0: Scaling up diffusion language models to 100b.arXiv preprint arXiv:2512.15745, 2025

work page internal anchor Pith review arXiv 2025
[5]

Brody, U

S. Brody, U. Alon, and E. Yahav. How attentive are graph attention networks? InInternational Conference on Learning Representations, 2022

work page 2022
[6]

Cao and M

C. Cao and M. Li. Generating mobility trajectories with retained data utility. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 2610–2620, 2021

work page 2021
[7]

J. Cao, T. Zheng, Q. Guo, Y . Wang, J. Dai, S. Liu, J. Yang, J. Song, and M. Song. Holistic semantic representation for navigational trajectory generation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 40–48, 2025

work page 2025
[8]

Chang, J

Y . Chang, J. Qi, Y . Liang, and E. Tanin. Contrastive trajectory similarity learning with dual- feature attention. In2023 IEEE 39th International conference on data engineering (ICDE), pages 2933–2945. IEEE, 2023

work page 2023
[9]

L. Chen, M. T. Özsu, and V . Oria. Robust and fast similarity search for moving object trajectories. InProceedings of the 2005 ACM SIGMOD international conference on Management of data, pages 491–502, 2005

work page 2005
[10]

W. Chen, Y . Zhu, Y . Chang, K. Luo, H. Wen, L. Li, Y . Yu, Q. Wen, C. Chen, K. Zheng, et al. Trajectory data management and mining: A survey from deep learning to the llm era.arXiv preprint arXiv:2403.14151, 2024

work page arXiv 2024
[11]

De Montjoye, C

Y .-A. De Montjoye, C. A. Hidalgo, M. Verleysen, and V . D. Blondel. Unique in the crowd: The privacy bounds of human mobility.Scientific reports, 3(1):1376, 2013

work page 2013
[12]

B. Deng, L. Ding, L. Ji, C. Chen, X. Jing, B. Qu, and D. Yang. Marionette: Fine-grained conditional generative modeling of spatiotemporal human trajectory data beyond imitation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, pages 463–473, 2025

work page 2025
[13]

Dijkstra

E. Dijkstra. A note on two problems in connexion with graphs.Numerische Mathematik, 1(1):269–271, 1959

work page 1959
[14]

J. Feng, Z. Yang, F. Xu, H. Yu, M. Wang, and Y . Li. Learning to simulate human mobility. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 3426–3433, 2020

work page 2020
[15]

Gambs, M.-O

S. Gambs, M.-O. Killijian, and M. N. del Prado Cortez. Next place prediction using mobility markov chains. InProceedings of the first workshop on measurement, privacy, and mobility, pages 1–6, 2012. 10

work page 2012
[16]

M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi. Understanding individual human mobility patterns.nature, 453(7196):779–782, 2008

work page 2008
[17]

Ho and T

J. Ho and T. Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

work page 2021
[18]

Huang, X

D. Huang, X. Song, Z. Fan, R. Jiang, R. Shibasaki, Y . Zhang, H. Wang, and Y . Kato. A variational autoencoder based generative model of urban human mobility. In2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 425–430. IEEE, 2019

work page 2019
[19]

T. S. Jepsen, C. S. Jensen, and T. D. Nielsen. Graph convolutional networks for road networks. In Proceedings of the 27th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, SIGSPATIAL ’19, page 460–463, New York, NY , USA, 2019. Association for Computing Machinery

work page 2019
[20]

Jiang, C

J. Jiang, C. Xu, J. Xu, M. Xu, N. Zheng, and K. Kong. Route planning for locations based on trajectory segments. InProceedings of the 2nd ACM SIGSPATIAL Workshop on Smart Cities and Urban Analytics, pages 1–8, 2016

work page 2016
[21]

Jiang, W

W. Jiang, W. X. Zhao, J. Wang, and J. Jiang. Continuous trajectory generation based on two- stage gan. InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 4374–4382, 2023

work page 2023
[22]

G. Jin, Y . Liang, Y . Fang, Z. Shao, J. Huang, J. Zhang, and Y . Zheng. Spatio-temporal graph neural networks for predictive learning in urban computing: A survey.IEEE transactions on knowledge and data engineering, 36(10):5388–5408, 2023

work page 2023
[23]

Keogh and C

E. Keogh and C. A. Ratanamahatana. Exact indexing of dynamic time warping.Knowledge and information systems, 7(3):358–386, 2005

work page 2005
[24]

T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016
[25]

Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro. Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761, 2020

work page arXiv 2009
[26]

L. Li, H. Xue, S. Ao, Y . Song, and F. Salim. Hit-jepa: A hierarchical self-supervised trajectory embedding framework for similarity computation.arXiv preprint arXiv:2507.00028, 2025

work page arXiv 2025
[27]

L. Li, H. Xue, Y . Song, and F. Salim. T-jepa: A joint-embedding predictive architecture for trajectory similarity computation. InProceedings of the 32nd ACM international conference on advances in geographic information systems, pages 569–572, 2024

work page 2024
[28]

P. Li, J. Wang, H. Zhang, X. Shi, N. Koshizuka, C. Shimizu, and R. Jiang. Trajflow: Nation-wide pseudo GPS trajectory generation with flow matching models. InThe Fourteenth International Conference on Learning Representations, 2026

work page 2026
[29]

T. Li, M. Chen, B. Guo, and Z. Shen. A survey on diffusion language models.arXiv preprint arXiv:2508.10875, 2025

work page arXiv 2025
[30]

X. Liu, H. Chen, and C. Andris. trajgans: Using generative adversarial networks for geo-privacy protection of trajectory data (vision paper). InLocation privacy and security workshop, pages 1–7, 2018

work page 2018
[31]

A. Lou, C. Meng, and S. Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution.arXiv preprint arXiv:2310.16834, 2023

work page internal anchor Pith review arXiv 2023
[32]

Y . Lv, Y . Duan, W. Kang, Z. Li, and F.-Y . Wang. Traffic flow prediction with big data: A deep learning approach.Ieee transactions on intelligent transportation systems, 16(2):865–873, 2014

work page 2014
[33]

S. Nie, F. Zhu, Z. You, X. Zhang, J. Ou, J. Hu, J. Zhou, Y . Lin, J.-R. Wen, and C. Li. Large language diffusion models.arXiv preprint arXiv:2502.09992, 2025. 11

work page internal anchor Pith review Pith/arXiv arXiv 2025
[34]

Y . Qin, H. Wu, W. Ju, X. Luo, and M. Zhang. A diffusion model for poi recommendation.ACM Transactions on Information Systems, 42(2):1–27, 2023

work page 2023
[35]

Rombach, A

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

work page 2022
[36]

S. S. Sahoo, M. Arriola, Y . Schiff, A. Gokaslan, E. Marroquin, J. T. Chiu, A. Rush, and V . Kuleshov. Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

work page 2024
[37]

Sohl-Dickstein, E

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015

work page 2015
[38]

Q. Team. Qwen3 technical report, 2025

work page 2025
[39]

Tizzoni, P

M. Tizzoni, P. Bajardi, A. Decuyper, G. Kon Kam King, C. M. Schneider, V . Blondel, Z. Smoreda, M. C. González, and V . Colizza. On the use of human mobility proxies for modeling epidemics.PLoS computational biology, 10(7):e1003716, 2014

work page 2014
[40]

H. Wang, Q. Zhang, Y . Wu, D. Jin, X. Wang, L. Zhu, and L. Yu. Synthesizing human trajectories based on variational point processes.IEEE Trans. on Knowl. and Data Eng., 36(4):1785–1799, Apr. 2024

work page 2024
[41]

Y . Wang, J. Cao, W. Huang, Z. Liu, T. Zheng, and M. Song. Spatiotemporal gated traffic trajectory simulation with semantic-aware graph learning.Information Fusion, 108:102404, 2024

work page 2024
[42]

Wesolowski, N

A. Wesolowski, N. Eagle, A. J. Tatem, D. L. Smith, A. M. Noor, R. W. Snow, and C. O. Buckee. Quantifying the impact of human mobility on malaria.Science, 338(6104):267–270, 2012

work page 2012
[43]

N. Wu, X. W. Zhao, J. Wang, and D. Pan. Learning effective road network representation with hierarchical graph neural networks. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’20, page 6–14, New York, NY , USA, 2020. Association for Computing Machinery

work page 2020
[44]

D. Xie, F. Li, and J. M. Phillips. Distributed trajectory similarity search.Proceedings of the VLDB Endowment, 10(11):1478–1489, 2017

work page 2017
[45]

Yang and G

C. Yang and G. Gidofalvi. Fast map matching, an algorithm integrating hidden markov model with precomputation.International Journal of Geographical Information Science, 32(3):547 – 570, 2018

work page 2018
[46]

L. Yang, Z. Zhang, Y . Song, S. Hong, R. Xu, Y . Zhao, W. Zhang, B. Cui, and M.-H. Yang. Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

work page 2023
[47]

J. Ye, Z. Xie, L. Zheng, J. Gao, Z. Wu, X. Jiang, Z. Li, and L. Kong. Dream 7b: Diffusion large language models.arXiv preprint arXiv:2508.15487, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[48]

L. Yu, W. Zhang, J. Wang, and Y . Yu. Seqgan: Sequence generative adversarial nets with policy gradient. InProceedings of the AAAI conference on artificial intelligence, volume 31, 2017

work page 2017
[49]

J. Yuan, Y . Zheng, C. Zhang, W. Xie, X. Xie, G. Sun, and Y . Huang. T-drive: driving directions based on taxi trajectories. InProceedings of the 18th SIGSPATIAL International conference on advances in geographic information systems, pages 99–108, 2010

work page 2010
[50]

Y . Yuan, Y . Liu, C. Han, J. Feng, and Y . Li. Breaking data silos: Towards open and scalable mobility foundation models via generative continual learning.arXiv preprint arXiv:2506.06694, 2025

work page arXiv 2025
[51]

Zhang, Q

J. Zhang, Q. Huang, Y . Huang, Q. Ding, and P.-W. Tsai. Dp-trajgan: A privacy-aware trajectory generation model with differential privacy.Future Generation Computer Systems, 142:25–40, 2023. 12

work page 2023
[52]

Zheng, Y

Y . Zheng, Y . Lin, L. Zhao, T. Wu, D. Jin, and Y . Li. Spatial planning of urban communities via deep reinforcement learning.Nature Computational Science, 3(9):748–762, 2023

work page 2023
[53]

Zheng, H

Y . Zheng, H. Su, J. Ding, D. Jin, and Y . Li. Road planning for slums via deep reinforcement learning. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5695–5706, 2023

work page 2023
[54]

Zheng, X

Y . Zheng, X. Xie, W.-Y . Ma, et al. Geolife: A collaborative social networking service among user, location and trajectory.IEEE Data Eng. Bull., 33(2):32–39, 2010

work page 2010
[55]

Z. Zhou, L. Chen, H. Tong, and D. Song. dllm: Simple diffusion language modeling, 2026

work page 2026
[56]

Y . Zhu, Y . Ye, S. Zhang, X. Zhao, and J. Yu. Difftraj: Generating gps trajectory with diffusion probabilistic model.Advances in Neural Information Processing Systems, 36:65168–65188, 2023

work page 2023
[57]

Y . Zhu, J. J. Yu, X. Zhao, Q. Liu, Y . Ye, W. Chen, Z. Zhang, X. Wei, and Y . Liang. Controltraj: Controllable trajectory generation with topology-constrained diffusion model. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4676–4687, 2024

work page 2024
[58]

Y . Zhu, J. J. Yu, X. Zhao, X. Zhou, L. Han, X. Wei, and Y . Liang. Unitraj: Learning a universal trajectory foundation model from billion-scale worldwide traces. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. A Limitations While TrajDLM improves trajectory generation in both fidelity and efficiency, several limitatio...

work page 2025