Pith. sign in

REVIEW 5 major objections 5 minor 45 references

MobCache caches LLM reasoning steps in latent space, cutting mobility simulation cost by 42–93% while matching the fidelity of full LLM-based methods.

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:PNM33CGE

load-bearing objection MobCache is a real step toward affordable LLM-based mobility simulation, but the fidelity claim for recombined reasoning chains rests on an evaluator trained against the LLM's own continuations — worth fixing before publication. the 5 major comments →

arxiv 2602.16727 v2 pith:PNM33CGE submitted 2026-02-17 cs.AI cs.LG

Mobility-Aware Cache Framework for Scalable LLM-Based Human Mobility Simulation

Hua Yan , Heng Tan , Yingxue Zhang , Yu Yang This is my paper
classification cs.AI cs.LG
keywords human mobility simulationlatent-space reasoningreconstructible cachelarge language modelsefficiencyknowledge distillationmobility lawsagent-based 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 establish that a cache of latent-space reasoning steps, rather than final answers, can make LLM-based human mobility simulation drastically cheaper without sacrificing fidelity. It introduces MobCache, which stores step-by-step reasoning from a fine-tuned LLM as embeddings, recombines them per user through a learned validity gate, and decodes them with a small distilled model. If correct, this means simulations of hundreds of thousands of agents become affordable, and the reasoning itself can be reused across cities and across other mobility simulators. The load-bearing premise is that latent embeddings retain enough logical structure for a trained evaluator to judge which recombinations are valid.

Core claim

MobCache replaces the expensive per-agent LLM rollout with three components: a reconstructible cache of latent-space reasoning chains generated once by a fine-tuned LLM; a latent-space evaluator that scores candidate branching steps by similarity to the model's own continuations; and a lightweight decoder distilled from the LLM under a mobility-law penalty that matches jump-distance and other statistical distributions. On two public mobility datasets, MobCache cuts inference time by at least 42%, raises throughput by at least 28%, and in a case study cuts cost by 93%, while staying within the quality margin of full LLM-based methods on radius of gyration, stay duration, jump length, location

What carries the argument

The central object is the reconstructible cache: a set of reasoning chains stored as latent-space embeddings, organized as a tree so new chains can branch from any node. Three mechanisms make it work: latent-space reasoning training, in which the LLM produces step embeddings instead of text tokens while the final answer remains in language; a latent-space evaluator, a learned scorer trained on similarity to the teacher model's next-step embeddings that gates which branches are logically acceptable; and a lightweight decoder, a small model trained by distillation plus a mobility-law KL constraint to map latent chains to natural-language activities.

Load-bearing premise

The latent-space evaluator is trained with labels that measure similarity to the fine-tuned LLM's own next-step generations, so if those continuations are not a reliable ground truth for logical validity, the cache can recombine chains that read fluently but violate mobility logic.

What would settle it

Construct deliberately invalid reasoning chains—for example, a chain that teleports a person from home to a location 30 km away with no travel step, or schedules holiday shopping on a date with no holiday—and check whether the latent-space evaluator scores them above the acceptance threshold used in inference. If the evaluator approves them, the logical-consistency gate is not functioning.

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

If this is right

  • At 42–93% lower cost per trajectory, population-scale mobility simulation becomes feasible on modest budgets, enabling larger and more detailed agent-based studies.
  • Because the cache stores reasoning rather than locations, a cache built in one city can be reused in another with only minor quality loss and a small cache-miss overhead.
  • The framework can wrap an existing LLM-based mobility simulator, accelerating it without retraining the underlying model, as long as that model exposes step-by-step reasoning.
  • Quality metrics (radius of gyration, stay duration, jump length, location frequency, OD similarity) remain comparable to state-of-the-art LLM methods, so the efficiency gain is not bought with a large fidelity loss.

Where Pith is reading between the lines

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

  • If the evaluator is reliable, the same latent-cache design could generalize to other structured reasoning domains—travel planning, scheduling, logistics—where stepwise reasoning can be cached and recombined.
  • The mobility-law-constrained distillation offers a template for preserving domain statistics when compressing LLM reasoning into small decoders; other generation tasks with known statistical invariants could adopt it.
  • The cross-city experiment's 6% cache-miss rate suggests that performance depends on profile coverage; a more diverse seed population or a hierarchical cache could push miss rates lower.
  • Because the evaluator learns from the teacher LLM's own continuations, it may inherit the teacher's systematic biases; a robustness check against human-annotated logic violations would strengthen the gate.

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 / 5 minor

Summary. The paper proposes MobCache, a framework for accelerating LLM-based human mobility simulation by caching latent-space reasoning embeddings instead of final responses. A fine-tuned LLM produces latent reasoning chains; a learned latent-space evaluator scores candidate recombinations; a lightweight decoder with mobility-law-constrained distillation converts the resulting chains into textual activities, which are mapped to coordinates with a gravity model. Experiments on Beijing and NYC datasets report large efficiency gains (e.g., 42–80% improvements in inference time, throughput, and cost) while claiming quality comparable to state-of-the-art LLM methods, plus a case study showing that MbCache can accelerate Urban-Mobility-LLM.

Significance. If the efficiency and fidelity claims hold, this work would meaningfully improve the scalability of LLM-based human mobility simulation, potentially enabling simulations with tens of millions of agents. The idea of caching and recombining latent reasoning steps, rather than caching final outputs, is interesting and complementary to token-level KV caching. The paper provides a complete system, experiments on two datasets, ablations, and a cross-city transfer study. However, the fidelity half of the central claim is not yet established: the latent evaluator is trained with self-referential labels, and the quality comparisons lack statistical support. The efficiency results are promising but need a fairer cost accounting.

major comments (5)
  1. [§3.4.2] The latent-space evaluator's supervision is b_t = sim(r_t, r̂_t), where r̂_t is generated by the same fine-tuned LLM used to build the cache. The evaluator therefore learns "what the LLM would generate next," not an independent criterion of logical validity or mobility awareness. This is circular: the §2.2 motivating example shows that fluent recombinations can be logically inconsistent, and the LLM's own continuations may share the same biases. Since the evaluator is the sole gate for branching (§3.5), the "maintaining fidelity" claim is not established. I suggest using external validity labels (e.g., rule-based mobility constraints or human annotations) or at least validating the evaluator against a set of known-incoherent chains.
  2. [§4.2.2, §4.3] The exploration rate (0.5) and λ (0.05) are selected on the same 10,000 test trajectories used for the reported quality metrics. This risks optimistic bias in the quality numbers. Please use a separate validation split or nested cross-validation, and report performance over the full grid of tested values. Without this, the "comparable performance" claim is not properly supported.
  3. [§4.3, Tables 1–2] All quality JSDs are reported as point estimates without error bars, repeated runs, or significance tests. Differences between methods are often small (e.g., Table 1: radius JSD 0.0333 vs. 0.0247; Table 2: radius JSD 0.1109 vs. 0.0542). The conclusion that MobCache is "comparable" to baselines is therefore not statistically grounded. Please provide confidence intervals or paired tests over multiple seeds/runs.
  4. [§4.2.4, Tables 1–2] The efficiency comparison mixes local LLaMA-3.2-3B (MobCache, Geo-LLaMA) with API-based GPT-4o baselines, and the cost accounting excludes the one-time cost of generating the 13k synthetic trajectories, fine-tuning the LLM, training the evaluator/decoder, and building the cache. Part of the reported speedup may be due to model size rather than the caching mechanism. Please report amortized training/cache-construction costs, or compare with a local LLM baseline under identical hardware, to isolate the contribution.
  5. [§3.5, §4.3] The final mapping from textual activities to real coordinates uses a "gravity model," but no equation, parameters, or fitting procedure is provided. Since radius of gyration, jump length, location frequency, and OdSim are all computed on these mapped coordinates, an unexplained mapping is load-bearing for the quality metrics. Please specify the gravity-model form, how it is calibrated, and its sensitivity.
minor comments (5)
  1. [§4.4] The phrase "66.93% reduction in inference speed" should be "reduction in inference time" or "66.93% faster inference."
  2. [Tables 1–3] The column header "Effiency" is misspelled; should be "Efficiency."
  3. [§3.3.2] "COT" is written inconsistently; use "CoT" throughout.
  4. [§4.5] Figures 7–8 are referenced qualitatively; please include numeric values or error bars for the cross-city comparison.
  5. [§3.5, §4.2.2] The profile-similarity threshold and the random search-round sampling (1–3) are not specified; state the threshold value and provide a random seed for reproducibility.

Circularity Check

2 steps flagged

Latent evaluator's 'validity' labels are the fine-tuned LLM's own continuations, making the recombination gate self-referential; the mobility-law constraint is likewise teacher-distilled. Efficiency and external JSD results remain independent checks.

specific steps
  1. self definitional [Section 3.4.2, 'Construct new reasoning chains', label construction (b_t = sim(r_t, r̂_t))]
    "For each training example (q, R_{1:t-1}), we use the fine-tuned LLM (details in Section. 3.3) to generate the next latent-space reasoning steps: r̂_t. (2) Compute similarity-based labels. Given a candidate reasoning r_t, its supervision label b_t is computed as: b_t = sim(r_t, r̂_t), where sim(·,·) is a similarity function (e.g., embedding cosine similarity or model-based scoring)."

    The evaluator is introduced to judge whether a branch is 'valid' and 'consistent with human reasoning' because latent embeddings are not interpretable. But the supervision label is defined as similarity to the next-step embedding generated by the same fine-tuned LLM that produced the cached chains. 'Valid' is therefore operationalized as 'close to what this LLM would generate next'; no independent mobility-logic criterion enters. At inference (§3.5) this evaluator selects the most plausible branch, so the recombination gate scores candidates by self-similarity to the very model whose biases are cached. The paper's motivating contradictions (holiday shopping on Nov 2, distance from workplace on a weekend) would only be caught if the LLM's own continuation were inconsistent; b_t never expose

  2. self definitional [Section 3.4.3, 'Lightweight decoder', Mobility law constraint (L_law = KL(z_τ(h_light) ∥ p_teacher(d(Y))))]
    "Given the activity sequences Y_teacher generated by the teacher decoder, we apply a statistics function d(·) to extract mobility law features (e.g., jump distances) and obtain the corresponding distribution p_teacher(d(Y)). ... The predicted distribution z_τ(h_light) is encouraged to match the teacher distribution p_teacher(d(Y)), by minimizing the Kullback–Leibler (KL) divergence: L_law = KL(z_τ(h_light)∥p_teacher(d(Y)))."

    The component is called a 'mobility law constraint,' but the target distribution p_teacher(d(Y)) is computed from the teacher decoder's own generated activity sequences, not from real-world mobility data or an independently stated law. The student is trained to match the teacher's empirical statistics, which is standard distillation. Thus the 'mobility law' is, by construction, the teacher model's distribution. This does not force the final JSD results against real data, but it means the spatial/temporal consistency claim inherits the teacher's biases rather than being anchored to an external mobility law.

full rationale

The paper's headline efficiency claims are not circular: inference time, token throughput, monetary cost, and cost reductions are measured against external baselines and are not forced by the training equations. The quality comparisons against held-out real trajectories (JSD of radius, duration, jump length, location frequency, OD similarity) are also external checks, so the central 'comparable performance' claim has independent content. The circularity is local and component-level. First, the latent-space evaluator—the sole gate for branch recombination—is trained with labels b_t = sim(r_t, r̂_t), where r̂_t is generated by the same fine-tuned LLM used to build the cache. Its notion of branch validity is therefore self-referential: it measures similarity to the cached generator's own continuations rather than an independent standard of mobility logic. Second, the 'mobility law' distillation target is the teacher decoder's own output distribution, so the law constraint is a renaming of standard teacher-student distillation. These self-referential objectives weaken the fidelity argument for recombined chains, but because the final evaluation is anchored to real-world distributions and external baselines, the derivation as a whole is not equivalent to its inputs. A score of 4 reflects one meaningful self-referential component plus a secondary self-referential training target, while the central efficiency and external quality findings remain independently testable.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on the assumption that latent-space reasoning embeddings can be recombined safely, and that the LLM's own next-step generations are a valid criterion for evaluating that recombination. These are design assumptions rather than independently verified facts. Free parameters are mostly hyperparameters chosen on the test set, and the gravity model is left unspecified.

free parameters (7)
  • exploration_rate = 0.5
    Selected from [0.3, 0.5, 0.7] based on efficiency/quality trade-off (§4.2.2, §4.9).
  • mobility_loss_weight_lambda = 0.05
    Selected from [0.01, 0.03, 0.05, 0.07] based on 'best performance' (§4.2.2).
  • profile_similarity_threshold
    Predefined threshold for cache retrieval is not specified in the paper (§3.5); its value affects how often the LLM is invoked.
  • search_rounds = sampled 1-3
    Number of branching iterations in 'explore alternative chains' is randomly sampled between 1 and 3 (§4.2.2), affecting exploration quality/efficiency.
  • similarity_function_sim
    The similarity function used for evaluator labels and branch candidate selection is not concretely specified (§3.4.2).
  • cache_size = 13,000 trajectories
    Cache constructed from 13,000 synthetic trajectories; size affects hit rate and diversity.
  • gravity_model_parameters
    Location mapping uses a gravity model, but its parameters are not reported (§3.5, §4.2.2).
axioms (5)
  • domain assumption LLM-generated step-by-step reasoning can produce realistic mobility trajectories
    Adopted from prior work (CoPB [32], Urban-Mobility-LLM [1]); the whole approach builds on this.
  • domain assumption Latent-space reasoning embeddings retain semantic/behavioral structure that supports recombination
    The paper asserts this (Section 3.3.2) but does not prove it; the empirical results are the only evidence.
  • ad hoc to paper Similarity in the latent embedding space is a valid proxy for logical compatibility of reasoning steps
    Used to build the evaluator labels and to select branch candidates (§3.4.2, §3.5).
  • ad hoc to paper The fine-tuned LLM's own next-step generations define what counts as a valid reasoning step
    Evaluator labels are computed as similarity to r̂_t generated by the fine-tuned LLM (§3.4.2).
  • domain assumption The gravity model is an appropriate mapping from textual activities to geographic coordinates
    Invoked at the final mapping stage without justification or parameter fitting (§3.5, §4.2.2).

pith-pipeline@v1.3.0-alltime-deepseek · 15480 in / 13846 out tokens · 117886 ms · 2026-08-02T22:45:19.235600+00:00 · methodology

0 comments
read the original abstract

Simulating large-scale human mobility is fundamental to understanding population movement patterns and supporting real-world geospatial applications such as urban planning, epidemic response, and transportation analysis. Recent works treat large language models (LLMs) as human agents to simulate realistic mobility behaviors using structured reasoning, but their high computational cost limits scalability. To address this, we design a mobility-aware cache framework named MobCache that leverages reconstructible caches to enable efficient large-scale human mobility simulations. It consists of: (1) a reasoning component that encodes each reasoning step as a latent-space embedding and uses a latent-space evaluator to enable the reuse and recombination of reasoning steps; and (2) a decoding component that employs a lightweight decoder trained with mobility law-constrained distillation to translate latent-space reasoning chains into natural language, thereby improving simulation efficiency while maintaining fidelity. Experiments show that MobCache significantly improves efficiency across multiple dimensions while maintaining performance comparable to state-of-the-art LLM-based methods.

Figures

Figures reproduced from arXiv: 2602.16727 by Heng Tan, Hua Yan, Yingxue Zhang, Yu Yang.

Figure 1
Figure 1. Figure 1: Core idea of MobCache. 2 Motivation 2.1 Why cache diversity matters We run a simple experiment to show that a limited diversity cache will influence simulation quality. We randomly select 10,000 trajec￾tories from a real-world dataset (described in the section 4) as the ground truth. Following a common group-based approach: LLM￾archetypes [6], we first build several clusters based on people’s profile attri… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of stay duration distributions. 0 1 2 3 4 5 6 7 8 9 1011121314 Location 0 0.4 0.8 1.2 1.6 2.0 Visit count (k) Real data Group-based method [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Framework of MobCache. To achieve latent-space reasoning, we then progressively replace textual reasoning steps with latent-space reasoning embeddings from earlier to later steps. Specifically, at each stage 𝑡, the first 𝑡 reasoning steps are replaced by latent embeddings 𝑅1:𝑡 , while the remaining reasoning steps and the final answer are still supervised in text. This continues until all reasoning is perf… view at source ↗
Figure 5
Figure 5. Figure 5: Efficiency comparison. Urban-Mobility-LLM Urban-Mobility-LLM (cache) 0.00 0.05 0.10 0.15 0.20 JSD Radius Duration Jump Locfreq [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of Mob￾Cache with LLM-archetypes. 0.3 0.5 0.7 Explore rate 1.0 1.5 2.0 2.5 Inference time (s) [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-city: efficiency. Cache (NYC) for NYC Cache (BJ) for NYC 0.00 0.05 0.10 0.15 0.20 JSD Radius Odsim Jump Locfreq [PITH_FULL_IMAGE:figures/full_fig_p007_7.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

45 extracted references · 14 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]

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Ruther- ford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bog- dan Damoc, Aidan Clark, et al. 2022. Improving language models by retrieving from trillions of tokens. InInternational conference on machine learning. PMLR, 2206–2240

    2022

  3. [3]

    Census Bureau

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

    2025

  4. [4]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2023. Frugalgpt: How to use large language models while reducing cost and improving performance.arXiv preprint arXiv:2305.05176(2023)

    Pith/arXiv arXiv 2023

  5. [5]

    ChicagoGov. 2023. Scooter Sharing in Chicago. https://www.chicago.gov/city/ en/depts/cdot/supp_info/escooter-share-pilot-project.html

    2023

  6. [6]

    Ayush Chopra, Shashank Kumar, Nurullah Giray-Kuru, Ramesh Raskar, and Arnau Quera-Bofarull. 2024. On the limits of agency in agent-based models. arXiv preprint arXiv:2409.10568(2024)

    Pith/arXiv arXiv 2024

  7. [7]

    Camille Couturier, Spyros Mastorakis, Haiying Shen, Saravan Rajmohan, and Victor Rühle. 2025. Semantic Caching of Contextual Summaries for Efficient Question-Answering with Language Models.arXiv preprint arXiv:2505.11271 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    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

  9. [9]

    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

  10. [10]

    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

  11. [11]

    In Gim, Guojun Chen, Seung-seob Lee, Nikhil Sarda, Anurag Khandelwal, and Lin Zhong. 2024. Prompt cache: Modular attention reuse for low-latency inference. Proceedings of Machine Learning and Systems6 (2024), 325–338

    2024

  12. [12]

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

    2008

  13. [13]

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769(2024)

    Pith/arXiv arXiv 2024

  14. [14]

    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

  15. [15]

    Chao Jin, Zili Zhang, Xuanlin Jiang, Fangyue Liu, Xin Liu, Xuanzhe Liu, and Xin Jin. 2024. Ragcache: Efficient knowledge caching for retrieval-augmented generation.arXiv preprint arXiv:2404.12457(2024)

    Pith/arXiv arXiv 2024

  16. [16]

    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

  17. [17]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th symposium on operating systems principles. 611–626

    2023

  18. [18]

    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

  19. [19]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33 (2020), 9459–9474

    2020

  20. [20]

    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

  21. [21]

    Shu Liu, Asim Biswal, Audrey Cheng, Xiangxi Mo, Shiyi Cao, Joseph E Gonza- lez, Ion Stoica, and Matei Zaharia. 2024. Optimizing llm queries in relational workloads.CoRR(2024)

    2024

  22. [22]

    Yuhan Liu, Hanchen Li, Kuntai Du, Jiayi Yao, Yihua Cheng, Yuyang Huang, Shan Lu, Michael Maire, Henry Hoffmann, Ari Holtzman, et al. 2023. Cachegen: Fast context loading for language model applications.CoRR(2023)

    2023

  23. [23]

    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

  24. [24]

    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

  25. [25]

    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

  26. [26]

    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

  27. [27]

    Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. 2023. MemGPT: Towards LLMs as Operating Systems. (2023)

    2023

  28. [28]

    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

  29. [29]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (Eds.). Association...

    doi:10.18653/v1/d19-1410 2019

  30. [30]

    Yangjun Ruan, Neil Band, Chris J Maddison, and Tatsunori Hashimoto. 2025. Reasoning to learn from latent thoughts.arXiv preprint arXiv:2503.18866(2025)

    arXiv 2025

  31. [31]

    Abhishek Shah. 2025. Navigating the LLM Cost Maze: A Q2 2025 Pricing and Limits Analysis. https://ashah007.medium.com/navigating-the-llm-cost-maze- a-q2-2025-pricing-and-limits-analysis-80e9c832ef39. Accessed: 2025-07-31

    2025

  32. [32]

    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

  33. [33]

    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

  34. [34]

    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

  35. [35]

    Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, and Ruihua Song

  36. [36]

    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

  37. [37]

    Yu Wang, Tongya Zheng, Shunyu Liu, Zunlei Feng, Kaixuan Chen, Yunzhi Hao, and Mingli Song. 2024. Spatiotemporal-augmented graph neural networks for human mobility simulation.IEEE Transactions on Knowledge and Data Engineering 36, 11 (2024), 7074–7086

    2024

  38. [38]

    Wayne Wu, Honglin He, Jack He, Yiran Wang, Chenda Duan, Zhizheng Liu, Quanyi Li, and Bolei Zhou. 2024. Metaurban: An embodied ai simulation platform Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Hua Yan, Heng Tan, Yingxue Zhang, and Yu Yang for urban micromobility.arXiv preprint arXiv:2407.08725(2024)

    Pith/arXiv arXiv 2024

  39. [39]

    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

  40. [40]

    Yuwei Yan, Qingbin Zeng, Zhiheng Zheng, Jingzhe Yuan, Jie Feng, Jun Zhang, Fengli Xu, and Yong Li. 2024. Opencity: A scalable platform to simulate urban activities with massive llm agents.arXiv preprint arXiv:2410.21286(2024)

    Pith/arXiv arXiv 2024

  41. [41]

    Dingqi Yang, Daqing Zhang, Vincent W Zheng, and Zhiyong Yu. 2014. Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs.IEEE Transactions on Systems, Man, and Cybernetics: Systems45, 1 (2014), 129–142

    2014

  42. [42]

    Zhaofan Zhang, Yanan Xiao, Lu Jiang, Dingqi Yang, Minghao Yin, and Pengyang Wang. 2024. Spatial-temporal interplay in human mobility: A hierarchical rein- forcement learning approach with hypergraph representation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 9396–9404

    2024

  43. [43]

    Gang Zhao, Ximing Zhang, Chenji Lu, Hui Zhao, Tianshu Wu, Pengjie Wang, Jian Xu, and Bo Zheng. 2025. Explainable LLM-driven Multi-dimensional Distillation for E-Commerce Relevance Learning. InCompanion Proceedings of the ACM on Web Conference 2025. 631–640

    2025

  44. [44]

    Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Jeff Huang, Chuyue Sun, Cody_Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. 2023. Efficiently Programming Large Language Models using SGLang. (2023). A Appendix A.1 Prompt example Daily mobility activity generation prompt example Profile Profile:{profile}; Date:{today_date}; Nearby...

    2023

  45. [2025]

    Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains.arXiv preprint arXiv:2505.16552(2025)

    arXiv 2025