Pith. sign in

REVIEW 5 major objections 5 minor 50 references

A Study on Individual Spatiotemporal Activity Generation Method Using MCP-Enhanced Chain-of-Thought Large Language Models

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Pairing chain-of-thought reasoning with a tool-based context protocol lets LLMs generate daily activity-travel chains that match real mobile signaling data, scoring 7.86–8.36 across four base models.

desk verdict A clearly described and honest integration of CoT, MCP, and fine-tuning for activity generation, but the headline quality scores are not yet trustworthy because the objective validation is partly circular and the benchmark comparison is deferred. read the letter →

arxiv 2506.10853 v1 pith:5NJHIZXU submitted 2025-06-12 cs.AI cs.CY

Yu Zhang , Yang Hu , De Wang This is my paper
classification cs.AIcs.CY
keywords LargeLanguageModelsChain-of-ThoughtReasoningModelContextProtocolSpatiotemporalBehaviorUrbanComputingHumanMobilityactivity-travelchainssyntheticdata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large language models are poor urban behavior simulators on their own: they lack spatial cognition, physical-constraint understanding, and individual diversity. This paper tries to establish that pairing chain-of-thought reasoning with a tool-based context protocol fixes those gaps. The method has the model reason through a five-stage decision process, mirroring how people weigh constraints and preferences, while a Model Context Protocol layer supplies real spatial, temporal, environmental, personal, social, and experiential data on demand. Tested on 1,000 synthetic person-days in Shanghai's Lujiazui district, the framework yields daily activity-travel chains whose aggregate distributions resemble anonymized mobile signaling data, with quality scores of 7.86 to 8.36 across four open-source base models, and parallel generation drops from 1.30 to 0.17 minutes per sample as workers increase from 2 to 12. If these results hold, the framework offers a cheaper, more generalizable route to synthetic mobility data for urban planning and transportation forecasting.

What carries the argument

The machinery is the MCP-enhanced chain-of-thought pipeline, a three-layer architecture. The autonomous model learning layer injects two knowledge repositories — a 1,845-entry behavioral knowledge base synthesized from over 1,000 publications and a local database of points of interest, road networks, activity diaries, anonymized mobile signaling data, and social-media posts — through fine-tuning, chain-of-thought demonstration learning, and MCP training. The CoT reasoning layer decomposes each activity decision into five explicit stages modeled on bounded-rationality deliberation. The MCP interaction layer exposes six tool suites as microservices that the model queries through structured messages during generation, which is what grounds the chains in real coordinates, schedules, and constraints. The architecture uses a distributed parallel-generation pipeline with inference optimization, and validation combines subjective criteria scored by an independent LLM judge with an objective term built from Jensen-Shannon divergence on spatial distributions and Kolmogorov-Smirnov statistics on temporal distributions.

What would settle it

Generate activity chains with the framework in a second district using no local training data, then compare the chains against independently collected activity diaries that played no role in building the synthetic training set: if the aggregate spatial and temporal distributions still match the signaling data while individual chains violate basic constraints such as infeasible transition times, persona-contradicting anchor points, or activity sequences absent from the diaries, the behavioral-realism claim fails. The minimal version is recomputing $Q_{\mathrm{obj}}$ on held-out survey records; a large drop would show the score measured training-set replay.

Watch

Extended reading notes

Core claim

The central claim is that explicit, staged reasoning plus tool-based grounding converts an LLM's implicit world knowledge into a usable simulator of individual spatiotemporal behavior. Inside the framework, a five-stage chain-of-thought scaffold handles the decision logic — situational awareness, constraint identification, option generation, multi-factor evaluation, and decision formation — while six MCP tool suites supply the facts: temporal management keeps schedules feasible, spatial navigation anchors activities to real points of interest and routes, environmental perception injects weather and events, personal memory preserves individual habits, social collaboration adds companions, and experience evaluation scores venues. The authors validate the generated chains against anonymized mobile signaling data using Jensen-Shannon divergence on spatial distributions and Kolmogorov-Smirnov statistics on temporal distributions, reporting weighted quality scores of 7.86 to 8.36 across four open-source base models spanning 7B to 13B parameters (GLM-4-9B-0414, DeepSeek-R1-Distill-Qwen-7B, LLAMA-2-13B, and Bloom-7B1). Ablations identify spatial navigation as the load-bearing component, with its removal cutting quality by 75.1 percent, and chain-of-thought as the second pillar, with a 42.4 percent drop when removed. The implied argument is that LLM-based behavioral simulation is limited less by model capability than by missing grounding and missing reasoning structure.

Load-bearing premise

The load-bearing premise is that matching the aggregate spatial and temporal distributions of anonymized mobile signaling data counts as evidence of behavioral realism, even though the same survey and signaling data were used to build the synthetic Markov-chain training set, so the objective score partly measures how well the generator replays the statistics it was trained on.

Editorial extensions

If this is right

  • If the reported quality scores hold, the framework becomes a zero-shot route to synthetic individual mobility data: it deploys without additional training data, which matters for cities where surveys and signaling data are scarce.
  • The ablation ordering implies that spatial grounding and explicit reasoning are mandatory components for LLM activity simulation, while memory, social, and experience tools are optional enhancements.
  • The parallel scaling result — 1.30 to 0.17 minutes per person-day from 2 to 12 workers — implies that generating a city-scale synthetic population is feasible on a single 80 GB GPU node, though acceleration diminishes as hardware bottlenecks dominate.
  • Because quality scores vary by only 0.50 points across models from 7B to 13B parameters, the framework's effectiveness appears tied to the scaffolding rather than any single base model, making it portable to stronger open models as they appear.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: because the Markov-chain synthetic training set was built from the same survey and signaling data used for objective validation, the reported objective scores likely overstate independent behavioral fidelity; a cleaner test would validate against held-out individual activity diaries.
  • Beyond the paper: the logged reasoning traces and tool queries make the framework a natural candidate for counterfactual scenario generation — simulating activity patterns under a new transit line, a weather shock, or a policy change — which the paper lists only as future work.
  • Beyond the paper: the same pipeline could serve as a data-augmentation engine for downstream mobility-prediction models, generating synthetic person-days to train next-location or activity-chain predictors in data-sparse regions.
  • Beyond the paper: the near-zero impact of the social collaboration and experience evaluation tools suggests the current system simulates isolated individuals rather than interacting groups; testing the framework on households or colleague networks would reveal whether it captures collective behavior at all.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes a framework that combines chain-of-thought (CoT) reasoning with the Model Context Protocol (MCP) to generate individual daily activity-travel chains with LLMs. The methodology includes a five-stage CoT reasoning layer, six MCP tool suites, a literature-derived knowledge base, a local spatiotemporal database, fine-tuning and knowledge-injection steps, and a parallel generation pipeline built on Ray and vLLM. Experiments in Shanghai's Lujiazui district report generation quality scores of 7.86-8.36 across four base models, an ablation study identifying spatial navigation and CoT as critical components, and parallel scaling results reducing mean generation time from 1.30 to 0.17 minutes per sample when increasing from 2 to 12 workers.

Significance. If the central claims were established, the framework would be a practically useful synthetic mobility data generator with a modular and reproducible architecture: the manuscript provides detailed algorithmic specifications in Appendices A-C, a public code repository, a substantial synthesized knowledge base of 1,845 entries, and a component ablation that gives useful sensitivity information. However, the headline validation is not currently convincing. The objective quality score is computed against the same mobile-signaling statistics used to construct the synthetic training data, the comparison against established baselines is explicitly deferred, and the subjective score relies on an uncalibrated GPT-4o-mini judge. As a result, the paper currently demonstrates a feasible system architecture and engineering pipeline rather than a validated advance in behavioral realism over existing methods.

major comments (5)
  1. [§4.1.1(3), §4.1.3, Eq. (3)] The objective validation is circular with respect to the central claim. Section 4.1.1 item 3 states that the 2,000 synthetic activity-travel chains used for model training and fine-tuning were generated by time-varying Markov chain modeling constrained by the activity diary and anonymized mobile signaling data. Section 4.1.3 then computes Qobj (Eq. 3) against the same mobile signaling data using JS divergence on KDE hotspots and KS statistics on activity times. A generator fine-tuned to imitate chains drawn from those empirical statistics will match these aggregate distributions by construction, so the Qobj values in Table 1 (8.50-9.10) do not establish independent behavioral realism. The paper's own Section 5.2 concedes that "macroscopic statistical matching approaches cannot effectively assess microscopic behavioral reasonableness." I ask the authors to validate on data not used for training, or to measure microscopic properties (e.g., activity-transition statistics, individual-level timing and sequencing) and to rephrase the claims accordingly.
  2. [§4.1.2(4), footnote 6] The claimed advantage over traditional methods is not supported. The abstract and introduction motivate the work by the limitations of rule-based and statistical approaches, and Contribution (5) claims experimental validation of advantages, but the benchmark comparison against LSTM, TVMC, rule-driven scheduling, and ActivitySim is explicitly deferred in Section 4.1.2 item 4 and footnote 6. Without those comparisons, the results show only that the proposed pipeline can generate outputs that resemble its own training distribution; they do not show superiority over existing baselines. Either the comparisons should be performed or the claims should be restricted to a feasibility demonstration.
  3. [§4.1.3, Eq. (4), Appendix C.1] The subjective component Qsubjective is assessed by GPT-4o-mini using the rubric in Appendix C.1, with no human-annotated calibration or inter-rater agreement reported. Since Qsubjective contributes half of Qtotal through Eq. (4), the headline quality scores depend on an unvalidated LLM judge. I request a human rating subset, agreement statistics, or at least an explicit sensitivity analysis showing that the conclusions do not change when Qsubjective is removed.
  4. [§5.1 vs §3.2.1 and §4.1.1(3)] The claim in Section 5.1 that the approach "enables zero-shot deployment without additional training data" is inconsistent with the methodology: Section 3.2.1 describes fine-tuning and knowledge injection, and Section 4.1.1 item 3 uses 2,000 synthetic chains for model training and fine-tuning optimization. If "zero-shot" refers only to deployment after a one-time training phase, the wording is misleading; if it refers to transfer to new study areas, no evidence is provided. Please clarify and support this claim.
  5. [Table 1 vs Table 2] There is an unexplained numerical inconsistency in the efficiency results. Table 1 reports GLM-4-9B generation time as 2.4±0.9 min/person-day, while Table 2 reports a mean generation time of 1.30 min/sample with 2 parallel workers. If these measurements are not directly comparable, the definitions and test conditions should be stated; as written, the scalability improvement from 1.30 to 0.17 min/sample is difficult to reconcile with the base-model timing.
minor comments (5)
  1. [Abstract and §4.2.1] The abstract says "1,000 generated samples," but Section 4.2.1 reports 500 samples per model for four models, totaling 2,000 samples; please make the sample counts consistent.
  2. [§4.1.3, Eqs. (1)-(3)] The objective quality score in Eq. (3) uses only the JS divergence on spatial hotspots and the KS statistic on temporal distributions, while the text also mentions daily presence probability density curves; the formula should specify how those curves enter Qobj.
  3. [§4.1.3, Eqs. (8)-(9)] The diversity index uses binary DTW distances and is not compared against the diversity of real activity chains; a statement of this limitation would help readers interpret the diversity scores.
  4. [References [12]] Reference [12] appears to cite a Caltech thesis for a deep reinforcement learning agent behavior model in ABM; please verify that this citation accurately supports the claim in Section 2.1.
  5. [Footnote 5] The term "a priori rationality" is introduced without a formal definition; please either define it precisely or replace it with a more standard term.

Circularity Check

1 steps flagged · score 6.0 of 10

Objective Qobj validation is partly circular: the mobile-signaling statistics used to generate the Markov-chain fine-tuning set are the same statistics used as the Qobj reference, so the aggregate-similarity scores are partly forced.

  1. fitted input called prediction [Section 4.1.1 item 3 and Section 4.1.3, Eqs. (1)-(3); Table 1]
    "Synthetic training data: Using the two aforementioned datasets, 2,000 individual activity-travel chains with "a priori rationality" were generated through time-varying Markov chain modeling methods for model training and fine-tuning optimization. [...] Complementing the subjective assessment, objective validation establishes quantitative benchmarks using real-world mobile signaling data."

    The Qobj reference distributions in Eqs. (1)-(3) are computed from the anonymized mobile signaling data. The same mobile signaling data is one of "the two aforementioned datasets" used to generate the 2,000 synthetic activity-travel chains by time-varying Markov chain modeling, and those chains are the LLM's fine-tuning data. A model fine-tuned to imitate Markov-sampled chains whose aggregate statistics were constrained to match the mobile signaling data will trivially obtain high JS/KS similarity to that same data. Thus Qobj = 8.50-9.10 in Table 1 is not independent evidence of behavioral realism; it largely measures replay of the training statistics. The paper's own Sec.

full rationale

The central circular step is the closed validation loop: the same mobile signaling data is used both to build the Markov-chain synthetic fine-tuning set (Sec. 4.1.1 item 3) and as the reference for the objective quality score Qobj (Sec. 4.1.3, Eqs. 1-3). The paper's own footnote defines "a priori rationality" as "predefined statistical constraints derived from the empirical patterns observed in the activity diary and mobile signaling datasets," which confirms that the training chains were constrained to match the very statistics later used for validation. Consequently, the headline Qtotal scores (7.86-8.36) and Qobj scores (8.50-9.10) are partly a measure of distributional replay rather than independent spatiotemporal realism. This is not a case of self-citation or imported uniqueness: the paper's cited background literature is external and not load-bearing. The efficiency results (parallel scaling), diversity index, and ablation rankings are self-contained empirical measurements and are not invalidated by this circularity. However, because the paper's central claim of generating realistic behavior rests substantially on the aggregate similarity scores, and the paper's own Section 5.2 acknowledges that such macroscopic matching cannot assess microscopic behavioral reasonableness, a partial-circularity score of 6 is appropriate. The deferred benchmark comparison (Sec. 4.1.2 item 4) further means no external baseline currently anchors the quality scores. I therefore flag one fitted-input-called-prediction step rather than alleging that the whole framework is definitionally circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical or formal entities are introduced; the MCP tool suite is a software architecture rather than an invented entity with independent empirical handles. The central claim rests mainly on hand-set evaluation weights, undisclosed Markov-chain training parameters, and assumptions that aggregate distributional similarity and LLM-based scoring capture behavioral realism.

free parameters (4)
  • alpha (Qtotal weighting) = 0.5
    Eq. (4) balances subjective and objective quality; hand-set with no sensitivity analysis reported.
  • subjective dimension weights = 25% each
    C.1 defines the comprehensive score as equal 25% weights for time logic, activity purpose, persona matching, and authenticity.
  • adaptive temperature scheduling = 0.7 for CoT, 0.1 for final output
    Section 3.2.4 sets temperatures by design to encourage exploration then precision; no tuning or ablation of this choice is reported.
  • Markov-chain parameters for synthetic training data = not reported
    Section 4.1.1 item 3 generates 2,000 'a priori rational' training chains from survey and mobile data; the transition parameters are undisclosed and directly shape what the LLM learns.
assumptions (4)
  • domain assumption Aggregate distributional similarity is valid evidence of individual behavioral plausibility.
    Eqs. (3)-(4) use JS divergence and KS statistics against mobile signaling data as objective quality; Section 5.2 admits this cannot assess microscopic reasonableness.
  • domain assumption GPT-4o-mini template scoring is a valid proxy for expert human evaluation.
    Section 4.1.3 replaces expert scoring with GPT-4o-mini; no human baseline, calibration, or inter-rater agreement is reported.
  • domain assumption Markov-chain synthetic training data are representative of real human behavior.
    Section 4.1.1 item 3 builds the 2,000 training chains from survey and mobile data; training on these chains embeds the source statistics into the generator.
  • domain assumption MCP tools accurately represent the external urban environment.
    Section 3.2.3 assumes POI, route, weather, memory, and social tools provide sufficient real-world fidelity; no independent verification of tool accuracy is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Study on Individual Spatiotemporal Activity Generation Method Using MCP-Enhanced Chain-of-Thought Large Language Models." pith.science (2026). https://pith.science/paper/5NJHIZXU

@misc{pith2026250610853,
  author       = {Pith},
  title        = {Pith review of: A Study on Individual Spatiotemporal Activity Generation Method Using MCP-Enhanced Chain-of-Thought Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5NJHIZXU}},
  note         = {Machine review of arXiv:2506.10853}
}
read the original abstract

Human spatiotemporal behavior simulation is critical for urban planning research, yet traditional rule-based and statistical approaches suffer from high computational costs, limited generalizability, and poor scalability. While large language models (LLMs) show promise as "world simulators," they face challenges in spatiotemporal reasoning including limited spatial cognition, lack of physical constraint understanding, and group homogenization tendencies. This paper introduces a framework integrating chain-of-thought (CoT) reasoning with Model Context Protocol (MCP) to enhance LLMs' capability in simulating spatiotemporal behaviors that correspond with validation data patterns. The methodology combines human-like progressive reasoning through a five-stage cognitive framework with comprehensive data processing via six specialized MCP tool categories: temporal management, spatial navigation, environmental perception, personal memory, social collaboration, and experience evaluation. Experiments in Shanghai's Lujiazui district validate the framework's effectiveness across 1,000 generated samples. Results demonstrate high similarity with real mobile signaling data, achieving generation quality scores of 7.86 to 8.36 across different base models. Parallel processing experiments show efficiency improvements, with generation times decreasing from 1.30 to 0.17 minutes per sample when scaling from 2 to 12 processes. This work contributes to integrating CoT reasoning with MCP for urban behavior modeling, advancing LLMs applications in urban computing and providing a practical approach for synthetic mobility data generation. The framework offers a foundation for smart city planning, transportation forecasting, and participatory urban design applications.

Figures

Figures reproduced from arXiv: 2506.10853 by the authors.

Figure 1
Figure 1. Evolution and Comparison of LLM-based Urban Activity Generation Methods [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. MCP-Enhanced Chain-of-Thought Theoretical Framework [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. LLM-based System Architecture and Methodology Framework [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Delimitation map of the study area The experimental data consist of: 1. Spatiotemporal Behavioral Knowledge Base: A structured database containing 1,845 knowledge entries that document human spatiotemporal behavioral patterns, which were synthesized from over 1,000 pee…
Figure 5
Figure 5. Figure 5: Schematic of the generation process for this method [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Average Token Distribution Across Base Models for Individual Spatiotemporal Behavior Generation [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Ablation Study Results: Impact of Component Removal on Generation Quality [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 34 canonical work pages

  1. [1]

    A review of human mobility research based on big data and its implication for smart city development.ISPRS International Journal of Geo-Information, 10(1):13, 2020

    Anqi Wang, Anshu Zhang, Edwin HW Chan, Wenzhong Shi, Xiaolin Zhou, and Zhewei Liu. A review of human mobility research based on big data and its implication for smart city development.ISPRS International Journal of Geo-Information, 10(1):13, 2020

    work page 2020

  2. [2]

    Human mobility data and analysis for urban resilience: A systematic review.Environment and Planning B: Urban Analytics and City Science, 49(5):1507–1535, 2022

    Masahiko Haraguchi, Akihiko Nishino, Akira Kodaka, Maura Allaire, Upmanu Lall, Liao Kuei-Hsien, Kaya Onda, Kota Tsubouchi, and Naohiko Kohtake. Human mobility data and analysis for urban resilience: A systematic review.Environment and Planning B: Urban Analytics and City Science, 49(5):1507–1535, 2022

    work page 2022

  3. [3]

    Spatiotemporal data mining: a survey on challenges and open problems.Artificial Intelligence Review, pages 1–48, 2022

    Ali Hamdi, Khaled Shaban, Abdelkarim Erradi, Amr Mohamed, Shakila Khan Rumi, and Flora D Salim. Spatiotemporal data mining: a survey on challenges and open problems.Artificial Intelligence Review, pages 1–48, 2022

    work page 2022

  4. [4]

    Data-driven generation of spatio-temporal routines in human mobility.Data Mining and Knowledge Discovery, 32(3):787–829, 2018

    Luca Pappalardo and Filippo Simini. Data-driven generation of spatio-temporal routines in human mobility.Data Mining and Knowledge Discovery, 32(3):787–829, 2018

    work page 2018

  5. [5]

    Modeling, replicating, and predicting human behavior: A survey.ACM Transactions on Autonomous and Adaptive Systems, 18(2):1–47, 2023

    Andrew Fuchs, Andrea Passarella, and Marco Conti. Modeling, replicating, and predicting human behavior: A survey.ACM Transactions on Autonomous and Adaptive Systems, 18(2):1–47, 2023

    work page 2023

  6. [6]

    Evaluating spatial understand- ing of large language models.arXiv preprint arXiv:2310.14540, 2023

    Yutaro Yamada, Yihan Bao, Andrew K Lampinen, Jungo Kasai, and Ilker Yildirim. Evaluating spatial understand- ing of large language models.arXiv preprint arXiv:2310.14540, 2023

    arXiv 2023

  7. [7]

    Geo-llama: Leveraging llms for human mobility trajectory generation with spatiotemporal constraints.arXiv preprint arXiv:2408.13918, 2024

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

    arXiv 2024

  8. [8]

    Homogenizing effect of large language model (llm) on creative diversity: An empirical comparison of human and chatgpt writing

    Kibum Moon, Adam Green, and Kostadin Kushlev. Homogenizing effect of large language model (llm) on creative diversity: An empirical comparison of human and chatgpt writing. 2024

    work page 2024

Show all 50 references
  1. [9]

    Towards a rigorous science of interpretable machine learning.arXiv preprint arXiv:1702.08608, 2017

    Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning.arXiv preprint arXiv:1702.08608, 2017

    2017 arXiv

  2. [10]

    Agent-based pedestrian modeling, 2001

    Michael Batty. Agent-based pedestrian modeling, 2001

    2001

  3. [11]

    Ubiquity Press, 2016

    Kay W Axhausen, Andreas Horni, and Kai Nagel.The multi-agent transport simulation MATSim. Ubiquity Press, 2016

    2016

  4. [12]

    California Institute of Technology, 2018

    Stephan Tao Zheng.Exploiting Structure for Scalable and Robust Deep Learning. California Institute of Technology, 2018

    2018

  5. [13]

    Multi-agent game abstraction via graph attention neural network

    Yong Liu, Weixun Wang, Yujing Hu, Jianye Hao, Xingguo Chen, and Yang Gao. Multi-agent game abstraction via graph attention neural network. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 7211–7218, 2020

    2020

  6. [14]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology, pages 1–22, 2023

    2023

  7. [15]

    Autogen: Enabling next-gen llm applications via multi-agent conversation.arXiv preprint arXiv:2308.08155, 2023

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. Autogen: Enabling next-gen llm applications via multi-agent conversation.arXiv preprint arXiv:2308.08155, 2023

    2023 arXiv

  8. [16]

    Creating large language model applications utilizing langchain: A primer on developing llm apps fast

    Oguzhan Topsakal and Tahir Cetin Akinci. Creating large language model applications utilizing langchain: A primer on developing llm apps fast. InInternational Conference on Applied Engineering and Natural Sciences, volume 1, pages 1050–1056, 2023

    2023

  9. [17]

    What about people in regional science.Transport Sociology: Social aspects of transport planning, pages 143–158, 1970

    Torsten Hägerstrand. What about people in regional science.Transport Sociology: Social aspects of transport planning, pages 143–158, 1970

    1970

  10. [18]

    Human activity patterns in the city: Things people do in time and in space., 1975

    Mark Granovetter. Human activity patterns in the city: Things people do in time and in space., 1975

    1975

  11. [19]

    The promises of big data and small data for travel behavior (aka human mobility) analysis.Transportation research part C: emerging technologies, 68: 285–299, 2016

    Cynthia Chen, Jingtao Ma, Yusak Susilo, Yu Liu, and Menglin Wang. The promises of big data and small data for travel behavior (aka human mobility) analysis.Transportation research part C: emerging technologies, 68: 285–299, 2016

    2016

  12. [20]

    Discovering urban activity patterns in cell phone data.Transportation, 42:597–623, 2015

    Peter Widhalm, Yingxiang Yang, Michael Ulm, Shounak Athavale, and Marta C González. Discovering urban activity patterns in cell phone data.Transportation, 42:597–623, 2015

    2015

  13. [21]

    Using mobile positioning data to model locations meaningful to users of mobile phones.Journal of urban technology, 17(1):3–27, 2010

    Rein Ahas, Siiri Silm, Olle Järv, Erki Saluveer, and Margus Tiru. Using mobile positioning data to model locations meaningful to users of mobile phones.Journal of urban technology, 17(1):3–27, 2010

    2010

  14. [22]

    Mining daily activity chains from large-scale mobile phone location data

    Ling Yin, Nan Lin, and Zhiyuan Zhao. Mining daily activity chains from large-scale mobile phone location data. Cities, 109:103013, 2021. 21 MCP-Enhanced CoT for Individual Spatiotemporal Behavioral GenerationA PREPRINT

    2021

  15. [23]

    Comprehensive econometric microsimulator for daily activity-travel patterns.Transportation Research Record, 1894(1):57–66, 2004

    Chandra R Bhat, Jessica Y Guo, Sivaramakrishnan Srinivasan, and Aruna Sivakumar. Comprehensive econometric microsimulator for daily activity-travel patterns.Transportation Research Record, 1894(1):57–66, 2004

    2004

  16. [24]

    Albatross: multiagent, rule-based model of activity pattern decisions.Transportation Research Record, 1706(1):136–144, 2000

    Theo Arentze, Frank Hofman, Henk Van Mourik, and Harry Timmermans. Albatross: multiagent, rule-based model of activity pattern decisions.Transportation Research Record, 1706(1):136–144, 2000

    2000

  17. [25]

    Validation of tasha: A 24-h activity scheduling microsimulation model.Transportation Research Part A: Policy and Practice, 42(2):360–375, 2008

    Matthew J Roorda, Eric J Miller, and Khandker MN Habib. Validation of tasha: A 24-h activity scheduling microsimulation model.Transportation Research Part A: Policy and Practice, 42(2):360–375, 2008

    2008

  18. [26]

    Data-driven activity scheduler for agent-based mobility models

    Jan Drchal, Michal ˇCertick`y, and Michal Jakob. Data-driven activity scheduler for agent-based mobility models. Transportation Research Part C: Emerging Technologies, 98:370–390, 2019

    2019

  19. [27]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  20. [28]

    An explanation of in-context learning as implicit bayesian inference.arXiv preprint arXiv:2111.02080, 2021

    Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference.arXiv preprint arXiv:2111.02080, 2021

    2021 arXiv

  21. [29]

    Exploring large language models for human mobility prediction under public events.Computers, Environment and Urban Systems, 112:102153, 2024

    Yuebing Liang, Yichao Liu, Xiaohan Wang, and Zhan Zhao. Exploring large language models for human mobility prediction under public events.Computers, Environment and Urban Systems, 112:102153, 2024

    2024

  22. [30]

    nextlocllm: next location prediction using llms.arXiv preprint arXiv:2410.09129, 2024

    Shuai Liu, Ning Cao, Yile Chen, Yue Jiang, and Gao Cong. nextlocllm: next location prediction using llms.arXiv preprint arXiv:2410.09129, 2024

    2024

  23. [31]

    Agentmove: Predicting human mobility anywhere using large language model based agentic framework.arXiv preprint arXiv:2408.13986, 2024

    Jie Feng, Yuwei Du, Jie Zhao, and Yong Li. Agentmove: Predicting human mobility anywhere using large language model based agentic framework.arXiv preprint arXiv:2408.13986, 2024

    2024 arXiv

  24. [32]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  25. [33]

    Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    2022

  26. [34]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36:68539–68551, 2023

    2023

  27. [35]

    Trafficgpt: Viewing, processing and interacting with traffic foundation models.Transport Policy, 150:95–105, 2024

    Siyao Zhang, Daocheng Fu, Wenzhe Liang, Zhao Zhang, Bin Yu, Pinlong Cai, and Baozhen Yao. Trafficgpt: Viewing, processing and interacting with traffic foundation models.Transport Policy, 150:95–105, 2024

    2024

  28. [36]

    Large language model for participatory urban planning

    Zhilun Zhou, Yuming Lin, Depeng Jin, and Yong Li. Large language model for participatory urban planning. arXiv preprint arXiv:2402.17161, 2024

    2024 arXiv

  29. [37]

    Urbankgent: A unified large language model agent framework for urban knowledge graph construction.arXiv preprint arXiv:2402.06861, 2024

    Yansong Ning and Hao Liu. Urbankgent: A unified large language model agent framework for urban knowledge graph construction.arXiv preprint arXiv:2402.06861, 2024

    2024 arXiv

  30. [38]

    Mapreduce: simplified data processing on large clusters.Communications of the ACM, 51(1):107–113, 2008

    Jeffrey Dean and Sanjay Ghemawat. Mapreduce: simplified data processing on large clusters.Communications of the ACM, 51(1):107–113, 2008

    2008

  31. [39]

    Spark: Cluster computing with working sets

    Matei Zaharia, Mosharaf Chowdhury, Michael J Franklin, Scott Shenker, and Ion Stoica. Spark: Cluster computing with working sets. In2nd USENIX workshop on hot topics in cloud computing (HotCloud 10), 2010

    2010

  32. [40]

    Ray: A distributed framework for emerging {AI} applications

    Philipp Moritz, Robert Nishihara, Stephanie Wang, Alexey Tumanov, Richard Liaw, Eric Liang, Melih Elibol, Zongheng Yang, William Paul, Michael I Jordan, et al. Ray: A distributed framework for emerging {AI} applications. In13th USENIX symposium on operating systems design and ...

    2018

  33. [41]

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 35...

    2020

  34. [42]

    Efficient memory management for large language model serving with pagedattention

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, p...

    2023

  35. [43]

    A behavioral model of rational choice.The quarterly journal of economics, pages 99–118, 1955

    Herbert A Simon. A behavioral model of rational choice.The quarterly journal of economics, pages 99–118, 1955. 22 MCP-Enhanced CoT for Individual Spatiotemporal Behavioral GenerationA PREPRINT A Appendix A: Knowledge Base Architecture for Spatiotemporal Behavior Research Based...

    1955

  36. [44]

    Activity space is highly correlated with cognitive maps

    Basic Theoretical and Empirical Applications of Human Behavior 1.1 Time Geography Theory Applications Human activities are constrained by spatiotemporal factors, with most activities concentrated at key locations. Activity space is highly correlated with cognitive maps. 1.2 Hu...

  37. [45]

    Women spend more time on household tasks

    Actual Patterns of Time Use 2.1 Time Budget Alloca- tion Among Different Social Groups High-income groups work longer hours but have higher leisure quality. Women spend more time on household tasks. Education level positively correlates with reading time. 2.2 Weekday vs. Weeke...

  38. [46]

    Leisure activities value environmental quality

    Real Preferences in Spatial Choice 3.1 Spatial Choice Character- istics for Different Activity Types Shopping driven by price sensitivity. Leisure activities value environmental quality. Social activities often occur in intermediate locations or highly accessible areas. 3.2 Di...

  39. [47]

    Cost weight decreases with in- come

    Actual Behavior in Transportation Mode Choice 4.1 Key Factors and Weights in Transportation Choice Time cost is the primary factor. Cost weight decreases with in- come. Elderly value comfort. Women focus on convenience. 4.2 Transportation Choice Characteristics of Different So...

  40. [48]

    Few activity points constitute most chains

    Actual Patterns of Activity Chains and Sequences 5.1 Typical Activity Chain Patterns and Occurrence Fre- quency Home-work-home chains dominate. Few activity points constitute most chains. Few people contribute complex chains. Shopping- dining combinations are high. 5.2 Activit...

  41. [49]

    Vul- nerable groups have fewer accessible facilities

    Interactive Relationships Between Individual Attributes and Environmental Factors 6.1 Social Economic At- tributes’ Impact on Activity Space Range Low-income car-free households have small activity spaces. Vul- nerable groups have fewer accessible facilities. Women have smalle...

  42. [50]

    Answer:

    Actual Impact of Technological Change on Behavior Patterns 7.1 Smartphone and Naviga- tion App Changes to Route Choice Navigation increases new route attempts, reduces peak time, sup- ports complex path choices, lowers perceived waiting time, affects spatial cognition. 7.2 Soc...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.