Pith. sign in

REVIEW 5 major objections 6 minor 43 references

This paper claims that feeding population-level mobility scaling laws back into LLM prompts makes simulated human agents collectively realistic, not just individually plausible.

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 22:45 UTC pith:CAD6XXQR

load-bearing objection Solid idea, honest held-out evaluation on Beijing, but the NYC 'ground truth' is a simulation and the manuscript has several internal inconsistencies; worth refereeing after revision. the 5 major comments →

arxiv 2602.16726 v2 pith:CAD6XXQR submitted 2026-02-17 cs.MA

Bridging Individual and Collective Realism in LLM-Based Human Mobility Simulation via Mobility Scaling-Law Guidance

Hua Yan , Heng Tan , Yu Yang This is my paper
classification cs.MA
keywords human mobility simulationlarge language modelsmobility scaling lawsprompt optimizationMonte Carlo Tree Searchmulti-objective optimizationcollective behaviorpopulation-level simulation
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 fill a gap in LLM-based human mobility simulation: individual agents reason like people, but the simulated population fails to reproduce the stable statistical regularities of real mobility—the distribution of trip distances, activity ranges, and visit frequencies. The authors argue that these mobility scaling laws can act as a feedback signal. Their framework, M2LSimu, measures the gap between simulated and target distributions, asks an LLM to propose group-level prompt adjustments, searches the adjustment space with Monte Carlo Tree Search under a multi-objective reward, and spreads the optimized prompts to similar profiles. On two public datasets it reports that this closes most of the gap, improving even metrics that were not directly optimized—travel distance, origin–destination similarity, circadian rhythm, exploration, and preferential return—by 11% to 64% over the best baseline. The payoff is a path to realistic, privacy-preserving, city-scale simulation without needing detailed private trajectories.

Core claim

The central claim is that collective realism in LLM-based human mobility simulation is achievable by treating empirical mobility scaling laws as population-level guidance for individual prompt construction. Concretely, the paper shows that computing mobility measures (radius of gyration, stay duration, visitation frequency, travel distance) from both the target shared data and the simulated trajectories, and then iteratively adjusting each individual's prompt through coarse-grained group-level strategies, yields simulated populations whose aggregate distributions match real-world scaling laws. The authors report that this joint optimization works across three types of shared data—coarse-grai

What carries the argument

The central machinery is the mobility measure itself: stable empirical regularities (e.g., the truncated power-law of travel distance, the radius-of-gyration distribution, and the power-law exponent ζ of visitation frequency) that survive coarse-graining and can be computed from shared data. M2LSimu wraps these measures in a closed loop—simulate, measure the distribution gap, let an LLM propose group-level prompt adjustments, re-simulate, evaluate with a geometric-mean multi-objective reward—and uses Monte Carlo Tree Search, aided by a global action-value filter, to find a budget-limited sequence of adjustments. The final ingredient is profile-similarity scaling: the adjusted prompt set, lea

Load-bearing premise

The New York City half of the evaluation treats a synthetic global simulation dataset, itself validated only against the same scaling laws the method aims to reproduce, as ground truth for real city mobility; if that proxy is not faithful to actual New York movement, the claimed second-city generalisation collapses.

What would settle it

Re-run the NYC experiment with observed mobility flows—e.g., transit station ridership or cell-tower counts—as ground truth instead of the synthetic proxy; if the generated trajectories reproduce the synthetic dataset's distributions but diverge from the real observed flows, the generalisation claim fails.

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

If this is right

  • A simulator guided by aggregate scaling laws can generate trajectories that match population-level reality without ever training on—or exposing—individual private trajectories, so the same method could operate under restrictive data-sharing policies.
  • Because the reward is a geometric mean over multiple mobility measures, the optimization explicitly avoids fixing one dimension at the cost of another; the paper shows guidance in spatial and temporal dimensions also improves spatial-temporal behavior such as exploration and preferential return.
  • The 30% subset search plus profile-similarity extension means the prompt-adjustment overhead is roughly one extra simulation run, making the approach practical for city-scale populations rather than only small cohorts.
  • Even purely statistical summaries (scalar exponents and cutoffs, no trajectories) were enough to outperform all baselines, implying the method can be applied where only published scaling-law parameters are available.
  • The success of coarse-grained shared data suggests that data providers can release mobility statistics at low spatial and temporal resolution without losing the guiding signal.

Where Pith is reading between the lines

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

  • If the scaling-law universality that social physics reports holds across cities, the same prompt-adjustment loop could transfer from one city to another with only a change in target distributions, making it a candidate for zero-shot cross-city simulation.
  • The paper's reliance on profile similarity to spread prompts implies a strong assumption that mobility behavior is predictable from demographic profile; a direct test is to compare same-profile individuals' actual visited places, and if profiles explain less mobility variance than assumed, the cost-saving extension would degrade.
  • The method's reward is population-level, so it cannot detect individual-level implausibilities that wash out in aggregate distributions; a natural extension is to add a per-agent realism constraint, such as home-work anchoring, and see whether collective realism is preserved.
  • The MCTS optimization is agnostic to the type of measure; one could feed it other collective constraints (e.g., commute-time distributions, carbon-emission totals) to steer simulations toward policy-relevant targets, effectively turning LLM agents into a controllable urban simulator.

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

5 major / 6 minor

Summary. The paper proposes M2LSimu, an LLM-based human mobility simulation framework that uses population-level mobility measures (radius of gyration, stay duration, visitation frequency, travel distance) derived from shared data as guidance to iteratively adjust individual-level prompts. Prompt adjustment is formulated as a Markov decision process and solved with Monte Carlo Tree Search, with a multi-objective reward based on the geometric mean of distances to target mobility distributions. The authors evaluate on a Beijing dataset and a New York City dataset (WorldMove), comparing against several LLM-based baselines and reporting improvements on metrics not directly used as guidance, such as travel distance, OD similarity, circadian rhythm, exploration, and preferential return. The paper also studies three types of shared data and an extension procedure from a 30% user subset to the full population.

Significance. If the results hold, the paper would make a useful contribution by introducing a general mechanism for injecting population-level constraints into LLM-based individual mobility simulation without retraining. The formulation is clean, the idea of using mobility scaling laws as a feedback signal is plausible, and the authors are careful to distinguish guidance metrics from held-out metrics when computing headline improvements. The MCTS-based optimization and the three shared-data settings are practical and could be reused by others. However, the strength of the empirical claims depends critically on the validity of the New York City ground truth, which is itself a simulation, and on the statistical reliability of single-run comparisons.

major comments (5)
  1. [Section 4.1, Tables 1-2, Abstract] The New York City 'ground truth' is WorldMove, described in Section 4.1 as 'a global simulation dataset that has been validated under scaling laws of human mobility [40]'. It is therefore synthetic, not observed NYC mobility. The abstract claims M2LSimu 'significantly outperforms state-of-the-art LLM-based methods on two public datasets,' and the limitations section says evaluation is 'limited to two cities,' implying both are real. The NYC experiment demonstrates only that M2LSimu can fit the output of another simulator; it does not provide independent evidence of generalization to a second real city. Even if WorldMove reproduces aggregate scaling laws, it may not capture the detailed distributions, OD structure, or behavioral heterogeneity of real NYC mobility. This is a load-bearing issue for the two-dataset generalization claim. The authors should either validate against real observe
  2. [Section 4.2, Tables 1-2] No error bars, confidence intervals, or repeated seeds are reported for any experiment. LLM generation is stochastic, and MCTS involves random exploration constants and random subset sampling; JSD differences as small as 0.005 between methods (e.g., Table 1 M2LSimu vs. M2LSimu w SD3 on Duration) may be within run-to-run variability. The claim of 'significant outperformance' is not statistically supported as presented. At minimum, report means and standard deviations across at least three independent runs for all tables and figures, and perform a significance test or state that differences are descriptive.
  3. [Table 1, Table 2, Sections 4.3.1 and 4.4.1] The gray-shading semantics are contradictory. The Table 1 caption says 'The gray-shaded parts indicate the metrics used for evaluation under the specific shared data setting.' Section 4.3.1 says for M2LSimu the non-guidance metrics 'were used for validation, shown in gray.' Section 4.4.1 says 'we use the metrics not included as guidance for evaluation, as shown in gray.' For M2LSimu w SD2, the guidance metrics are travel distance and stay duration, but the text appears to mark the validation metrics in gray. If the gray shading changes meaning across rows or across tables, this must be stated explicitly. As written, a reader cannot tell which numbers support the headline 'improvements on metrics not directly optimized' claim.
  4. [Section 3.4, Eq. (2)-(4), Section 4.3.1] The reward function directly minimizes divergence from the target mobility measures (radius of gyration, stay duration, visitation frequency) by using those exact measures in the reward. Consequently, improvements on those metrics reported in Table 1 are by construction and cannot be interpreted as emergent collective behavior. The paper partially addresses this by excluding guidance metrics from the headline improvement percentages, which is commendable. However, Sections 4.3.1 and 4.5 still use phrases such as 'using mobility measures as guidance can effectively guide individual generation, and produce better collective behavioral patterns' without distinguishing built-in optimization from genuine emergence. Please rephrase these claims to clarify that guidance metrics are optimized targets, and emphasize that the evidence for emergence rests solely on the non-guided metrics.
  5. [Section 3.6, Figure 8] The scaling of optimized prompts from a 30% subset to the full population via profile similarity is supported only by Figure 8, which shows two metrics on the NYC dataset with no error bars or statistical comparison. If the 30% subset is not representative of the population, or if profile similarity does not guarantee behavioral similarity, the optimized prompts may not transfer. This is a key scalability claim of the method. The authors should provide repeated-seed results for Figure 8 and ideally a sensitivity analysis over the subset fraction and the similarity threshold.
minor comments (6)
  1. [Title and Abstract] The abstract names the framework 'COMPASS', but the rest of the paper calls it 'M2LSimu'. The title also refers to 'Mobility Scaling-Law Guidance' while the full-text title is 'Guiding LLM-Based Human Mobility Simulation with Mobility Measures from Shared Data'. These inconsistencies should be fixed.
  2. [Section 4.1] The description of the Beijing dataset is brief. It says the data were 'collected via a social networking platform' [24]; please specify the spatial/temporal resolution and how the 1,200 individuals were sampled. Also, the paper says 'the dataset does not contain user identifiers' for NYC, but then states that NYC trajectories are used as ground truth and user profiles are simulated; clarify whether any profile information is used for validation.
  3. [Section 4.2.2] The baseline CitySim is described as using 'LLM-powered agents with personas, memory, and long-term goals' but the paper does not state whether CitySim was adapted to the two datasets in the same way as the other baselines. Please describe the prompt/input setup for each baseline.
  4. [Section 4.5, Figure 7] Figure 7 shows only Distance, Circadian, and Exploration for the ablation, but the text discusses spatial, temporal, and spatial-temporal dimensions. Ensure the figure includes all relevant metrics or clearly indicate which metrics are selected.
  5. [Appendix A.2, Figure 11] The paper states that travel distances are clipped to 600 km 'during plotting only', but the fitted cutoff and exponent values shown in the figure may still be affected by the clipping. Clarify whether the parameter fits were computed before or after clipping.
  6. [References] Reference [24] is cited for the Beijing dataset but also for the CoPB baseline; please ensure the dataset source is clearly distinguished from the baseline method. Also, reference [40] is a preprint; indicate if a peer-reviewed version is available.

Circularity Check

0 steps flagged

No significant circularity: reward-guided metrics are excluded from headline claims.

full rationale

The paper's headline improvements are evaluated on metrics that are not directly optimized by the reward function. Section 3.3 defines the reward as a geometric mean of distances to the target mobility measures, so gains on the guided measures (radius, stay duration, visitation frequency on Beijing; distance and duration on NYC) are by construction. However, Section 4.2.3 explicitly states: 'performance improvements are evaluated only on metrics excluding the guidance measures,' and the reported gains (64.08% travel distance, 15.09% OdSim, 11.29% circadian, 47.69% exploration, 63.80% return on Beijing; 11.70% OdSim and 24.63% circadian on NYC) are on validation metrics not present in the reward. These are not reductions of the optimization objective. The only notable concern is external validity, not circularity: the NYC 'ground truth' is itself the synthetic WorldMove dataset (Section 4.1), so the second city is not an independent real-world test, and the Limitations section understates this by saying only 'Our evaluation is currently limited to two cities.' This weakens the generalization claim but is not an Eq. X = Eq. Y construction. No load-bearing self-citation or imported uniqueness theorem is used. Hence no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities; its invented machinery is the prompt-adjustment protocol. The main cost is the many hand-set hyperparameters and untested behavioral assumptions about the LLM's responsiveness to prompt edits.

free parameters (5)
  • reward weighting coefficient mu = 0.5
    Section 4.2.1; chosen by hand to balance Wasserstein and CCDF L1 terms in reward.
  • prompt-update fraction k = 10%
    Section 4.2.1; fraction of individuals re-prompted per transition; no sensitivity analysis.
  • MCTS depth / candidates / simulations / exploration constants = depth=10, candidates=3, simulations=50, c=1.4, c_g=1.0
    Section 4.2.1; hand-set, no ablation.
  • search subset size = 30% of users
    Section 4.2.1; subset used to find prompts then scaled to full population; scaling validity only lightly tested.
  • preliminary adjustment fraction = 5%
    Section 2.2; used in the preliminary single-measure study.
axioms (5)
  • domain assumption Human mobility follows stable truncated power-law scaling in travel distance, radius of gyration, and visitation frequency.
    Section 2.1 relies on prior scaling-law results [10,28] as targets for guidance; if these are not universal, the guidance targets are wrong.
  • domain assumption Coarse-grained shared trajectories preserve the mobility measures needed for guidance.
    Section 2.3 shows this only for one Beijing dataset; assumed for other shared data types and cities.
  • ad hoc to paper An LLM can translate distribution gaps into effective per-group prompt adjustments.
    Section 3.3 action generation relies on LLM judgment; no direct verification that the proposed strategies cause the intended behavioral changes.
  • ad hoc to paper Prompts optimized on a 30% subset transfer to similar users via profile similarity.
    Section 3.6; only one comparison (Fig 8) without error bars.
  • domain assumption The WorldMove simulation dataset can serve as real-world ground truth for NYC mobility.
    Section 4.1 describes NYC data as a 'global simulation dataset... validated under scaling laws'; it is not observed mobility data.

pith-pipeline@v1.3.0-alltime-deepseek · 17220 in / 13597 out tokens · 128723 ms · 2026-08-02T22:45:27.852097+00:00 · methodology

0 comments
read the original abstract

Geospatial applications such as urban planning, epidemic forecasting, and transportation demand modeling depend on individual mobility data, but such data are costly to collect, uneven in coverage, and privacy-sensitive. Human mobility simulation offers a scalable alternative. A recent line of work treats large language models (LLMs) as human agents, modeling individual cognitive processes to generate realistic trajectories. Yet because each agent is simulated in isolation, these methods provide no population-level coordination mechanism, and the collective regularities of real mobility - how trip distances, visited locations, and flows distribute across a population - fail to emerge. We close this gap with COMPASS, which turns empirical mobility scaling laws into a feedback signal that guides prompt construction. COMPASS starts from coarse, population-level adjustments driven by these scaling laws and progressively refines them into individual prompts, jointly satisfying multiple aggregate objectives while keeping individual trajectories realistic. Across two public datasets, COMPASS outperforms state-of-the-art LLM-based simulators.

Figures

Figures reproduced from arXiv: 2602.16726 by Heng Tan, Hua Yan, Yu Yang.

Figure 1
Figure 1. Figure 1: Core idea of M2LSimu. insufficient activity around midday or at night, this does not im￾ply that all adjusted individuals should both travel short distances and be highly active during those periods. Moreover, adjustments aimed at satisfying one mobility measure must be coordinated with other measures to avoid unintended conflicts. (ii) Each adjustment requires re-running the LLM-based simulation for multi… view at source ↗
Figure 2
Figure 2. Figure 2: Travel distance distri￾butions (Real vs. Simulation). Real-world data Simulation Method 1 2 Exponent [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Travel distance (Real vs. Coarse-grained). 2.3 Preservation of human mobility scaling laws in coarse-grained data In this section, we examine whether coarse-grained trajectories preserve the scaling laws of human mobility observed in real-world data. Following [37], we map the geographic coordinates of real￾world data onto a 500 m × 500 m spatial grid and compute all dis￾tances in the grid space. Time is d… view at source ↗
Figure 6
Figure 6. Figure 6: Framework of M2LSimu. between a specific target measure and the corresponding simulated one. From a population-level perspective, this LLM examines the simulated data, partitions individuals into multiple groups based on their mobility behaviors, and proposes corresponding prompt adjustment strategies for each group (see an example in [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of spatial and temporal guidance on spatial￾temporal performance. OdSim Circadian 0.00 0.05 0.10 0.15 0.20 JSD M2LSimu w SD2 M2LSimu w/o E [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Stay duration distribu￾tions (Real vs. Coarse-grained). Real-world data Coarse-grained data Method 1 2 Exponent [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Effect of prompt ad￾justment on travel distance. Real-world data M2LSimu M2LSimu w/o PA Method 0.0 0.5 1.0 Exponent alpha [PITH_FULL_IMAGE:figures/full_fig_p010_11.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

43 extracted references · 13 linked inside Pith

  1. [1]

    Prabin Bhandari, Antonios Anastasopoulos, and Dieter Pfoser. 2024. Urban mobility assessment using llms. InProceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 67–79

    2024

  2. [2]

    Nicolas Bougie and Narimawa Watanabe. 2025. Citysim: Modeling urban behav- iors and city dynamics with large-scale llm-driven agent simulation. InProceed- ings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track. 215–229

    2025

  3. [3]

    Dirk Brockmann, Lars Hufnagel, and Theo Geisel. 2006. The scaling laws of human travel.Nature439, 7075 (2006), 462–465

    2006

  4. [4]

    Census Bureau

    U.S. Census Bureau. 2025. American Community Survey (ACS). https://www. census.gov/programs-surveys/acs.html

    2025

  5. [5]

    Jiale Cheng, Xiao Liu, Kehan Zheng, Pei Ke, Hongning Wang, Yuxiao Dong, Jie Tang, and Minlie Huang. 2024. Black-box prompt optimization: Aligning large language models without model training. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 3201–3219

    2024

  6. [6]

    Yuwei Du, Jie Feng, Jian Yuan, and Yong Li. 2025. CAMS: A CityGPT-Powered Agentic Framework for Urban Human Mobility Simulation.arXiv preprint arXiv:2506.13599(2025)

    Pith/arXiv arXiv 2025

  7. [7]

    Zipei Fan, Xuan Song, Yinghao Liu, Zhiwen Zhang, Chuang Yang, Quanjun Chen, Renhe Jiang, and Ryosuke Shibasaki. 2020. Human mobility based individual-level epidemic simulation platform.SIGSPATIAL Special12, 1 (2020), 34–40

    2020

  8. [8]

    Jie Feng, Zeyu Yang, Fengli Xu, Haisu Yu, Mudan Wang, and Yong Li. 2020. Learning to simulate human mobility. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 3426–3433

    2020

  9. [9]

    Haoyu Geng, Guanjie Zheng, Zhengqing Han, Hua Wei, and Zhenhui Li. 2022. HMES: A Scalable Human Mobility and Epidemic Simulation System with Fast Intervention Modeling. In2022 IEEE Smartworld, Ubiquitous Intelligence & Com- puting, Scalable Computing & Communications, Digital Twin, Privacy Comput- ing, Metaverse, Autonomous & Trusted Vehicles (SmartWorld/...

    2022

  10. [10]

    Marta C Gonzalez, Cesar A Hidalgo, and Albert-Laszlo Barabasi. 2008. Under- standing individual human mobility patterns.nature453, 7196 (2008), 779–782

    2008

  11. [11]

    Wenyang Hu, Yao Shu, Zongmin Yu, Zhaoxuan Wu, Xiaoqiang Lin, Zhongxiang Dai, See-Kiong Ng, and Bryan Kian Hsiang Low. 2024. Localized zeroth-order prompt optimization.Advances in Neural Information Processing Systems37 (2024), 86309–86345

    2024

  12. [12]

    Yasaman Jafari, Dheeraj Mekala, Rose Yu, and Taylor Berg-Kirkpatrick. 2024. Morl-prompt: An empirical analysis of multi-objective reinforcement learning for discrete prompt optimization.arXiv preprint arXiv:2402.11711(2024)

    Pith/arXiv arXiv 2024

  13. [13]

    WANG JIAWEI, Renhe Jiang, Chuang Yang, Zengqing Wu, Ryosuke Shibasaki, Noboru Koshizuka, Chuan Xiao, et al . 2024. Large language models as urban residents: An llm agent framework for personal mobility generation.Advances in Neural Information Processing Systems37 (2024), 124547–124574

    2024

  14. [14]

    Chenlu Ju, Jiaxin Liu, Shobhit Sinha, Hao Xue, and Flora Salim. 2025. Trajllm: A modular llm-enhanced agent-based framework for realistic human trajectory simulation. InCompanion Proceedings of the ACM on Web Conference 2025. 2847– 2850

    2025

  15. [15]

    Pierre-Yves Lajoie, Bobak Hamed Baghi, Sachini Herath, Francois Hogan, Xue Liu, and Gregory Dudek. 2024. PEOPLEx: Pedestrian opportunistic positioning leveraging IMU, UWB, BLE and WiFi. InICC 2024-IEEE International Conference on Communications. IEEE, 3518–3523

    2024

  16. [16]

    Siyu Li, Toan Tran, Haowen Lin, John Krumm, Cyrus Shahabi, Lingyi Zhao, Khurram Shafique, and Li Xiong. 2024. Geo-llama: Leveraging llms for human mobility trajectory generation with spatiotemporal constraints.arXiv preprint arXiv:2408.13918(2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Yifan Liu, Xishun Liao, Haoxuan Ma, Brian Yueshuai He, Chris Stanford, and Jiaqi Ma. 2024. Human Mobility Modeling with Household Coordination Activi- ties under Limited Information via Retrieval-Augmented LLMs.arXiv preprint arXiv:2409.17495(2024)

    Pith/arXiv arXiv 2024

  18. [18]

    Xinyi Mou, Xuanwen Ding, Qi He, Liang Wang, Jingcong Liang, Xinnong Zhang, Libo Sun, Jiayu Lin, Jie Zhou, Xuanjing Huang, et al. 2024. From individual to society: A survey on social simulation driven by large language model-based agents.arXiv preprint arXiv:2412.03563(2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Farshid Nooshi and Suining He. 2025. Multi-Agent Reinforcement Learning for Dynamic Mobility Resource Allocation with Hierarchical Adaptive Grouping. arXiv preprint arXiv:2507.20377(2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Department of Transportation

    U.S. Department of Transportation. 2022. 2022 National Household Travel Survey User’s Guide. Nine -wave national household travel behavior survey since 1969; rich socio-demographic, trip mode/purpose data

    2022

  21. [21]

    Kun Ouyang, Reza Shokri, David S Rosenblum, and Wenzhuo Yang. 2018. A non-parametric generative model for human trajectories.. InIJCAI, Vol. 18. 3812– 3817

    2018

  22. [22]

    Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, et al. 2025. Agentsociety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society.arXiv preprint arXiv:2502.08691(2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Christian M Schneider, Vitaly Belik, Thomas Couronné, Zbigniew Smoreda, and Marta C González. 2013. Unravelling daily human mobility motifs.Journal of The Royal Society Interface10, 84 (2013), 20130246

    2013

  24. [24]

    Chenyang Shao, Fengli Xu, Bingbing Fan, Jingtao Ding, Yuan Yuan, Meng Wang, and Yong Li. 2024. Chain-of-planned-behaviour workflow elicits few-shot mobil- ity generation in llms.arXiv preprint arXiv:2402.09836(2024)

    Pith/arXiv arXiv 2024

  25. [25]

    Chengshuai Shi, Kun Yang, Zihan Chen, Jundong Li, Jing Yang, and Cong Shen

  26. [26]

    Somanshu Singla, Zhen Wang, Tianyang Liu, Abdullah Ashfaq, Zhiting Hu, and Eric Xing. 2024. Dynamic rewarding with prompt optimization enables tuning- free self-alignment of language models. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 21889–21909

    2024

  27. [27]

    Ankita Sinha, Wendi Cui, Kamalika Das, and Jiaxin Zhang. 2024. Survival of the Safest: Towards Secure Prompt Optimization through Interleaved Multi-Objective Evolution.arXiv preprint arXiv:2410.09652(2024)

    Pith/arXiv arXiv 2024

  28. [28]

    Chaoming Song, Tal Koren, Pu Wang, and Albert-László Barabási. 2010. Modelling the scaling properties of human mobility.Nature physics6, 10 (2010), 818–823

    2010

  29. [29]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. Mpnet: Masked and permuted pre-training for language understanding.Advances in neural information processing systems33 (2020), 16857–16867

    2020

  30. [30]

    Heng Tan, Yukun Yuan, Shuxin Zhong, and Yu Yang. 2023. Joint rebalancing and charging for shared electric micromobility vehicles with energy-informed demand. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2392–2401

    2023

  31. [31]

    Eran Toch, Boaz Lerner, Eyal Ben-Zion, and Irad Ben-Gal. 2019. Analyzing large-scale human mobility data: a survey of machine learning methods and applications.Knowledge and Information Systems58, 3 (2019), 501–523

    2019

  32. [32]

    Prashant Trivedi, Souradip Chakraborty, Avinash Reddy, Vaneet Aggarwal, Am- rit Singh Bedi, and George K Atia. 2025. Align-pro: A principled approach to prompt optimization for llm alignment. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 27653–27661

    2025

  33. [33]

    Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian Luo, Jiayou Zhang, Nebojsa Jojic, Eric P Xing, and Zhiting Hu. 2023. Promptagent: Strategic planning with language models enables expert-level prompt optimization.arXiv preprint arXiv:2310.16427(2023)

    Pith/arXiv arXiv 2023

  34. [34]

    Wayne Wu, Honglin He, Jack He, Yiran Wang, Chenda Duan, Zhizheng Liu, Quanyi Li, and Bolei Zhou. 2024. Metaurban: An embodied ai simulation platform for urban micromobility.arXiv preprint arXiv:2407.08725(2024)

    Pith/arXiv arXiv 2024

  35. [35]

    Wayne Wu, Honglin He, Chaoyuan Zhang, Jack He, Seth Z Zhao, Ran Gong, Quanyi Li, and Bolei Zhou. 2025. Towards autonomous micromobility through scalable urban simulation. InProceedings of the Computer Vision and Pattern Recognition Conference. 27553–27563

    2025

  36. [36]

    Yurong Wu, Yan Gao, Bin Benjamin Zhu, Zineng Zhou, Xiaodi Sun, Sheng Yang, Jian-Guang Lou, Zhiming Ding, and Linjun Yang. 2024. Strago: Harnessing strategic guidance for prompt optimization.arXiv preprint arXiv:2410.08601 (2024)

    Pith/arXiv arXiv 2024

  37. [37]

    Takahiro Yabe, Kota Tsubouchi, Toru Shimizu, Yoshihide Sekimoto, Kaoru Sezaki, Esteban Moro, and Alex Pentland. 2024. YJMob100K: City-scale and longitudinal dataset of anonymized human mobility trajectories.Scientific Data11, 1 (2024), 397

    2024

  38. [38]

    Sheng Yang, Yurong Wu, Yan Gao, Zineng Zhou, Bin Benjamin Zhu, Xiaodi Sun, Jian-Guang Lou, Zhiming Ding, Anbang Hu, Yuan Fang, et al. 2024. Ampo: Automatic multi-branched prompt optimization.arXiv preprint arXiv:2410.08696 (2024)

    Pith/arXiv arXiv 2024

  39. [39]

    Haoqi Yuan, Yuhui Fu, Feiyang Xie, and Zongqing Lu. 2024. Pre-trained multi-goal transformers with prompt optimization for efficient online adaptation.Advances in Neural Information Processing Systems37 (2024), 55086–55114

    2024

  40. [40]

    Yuan Yuan, Yuheng Zhang, Jingtao Ding, and Yong Li. 2025. WorldMove, a global open data for human mobility.arXiv preprint arXiv:2504.10506(2025)

    arXiv 2025

  41. [41]

    Guang Zhao, Byung-Jun Yoon, Gilchan Park, Shantenu Jha, Shinjae Yoo, and Xiaoning Qian. 2025. Pareto prompt optimization. InThe Thirteenth International Conference on Learning Representations

    2025

  42. [42]

    Yuanshao Zhu, Yongchao Ye, Shiyao Zhang, Xiangyu Zhao, and James Yu. 2023. Difftraj: Generating gps trajectory with diffusion probabilistic model.Advances in Neural Information Processing Systems36 (2023), 65168–65188. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Hua Yan, Heng Tan, and Yu Yang A Appendix A.1 Supplementary analysis: preservation...

    2023

  43. [2024]

    Advances in Neural Information Processing Systems37 (2024), 99646–99685

    Efficient prompt optimization through the lens of best arm identification. Advances in Neural Information Processing Systems37 (2024), 99646–99685

    2024