arxiv: 2605.10782 · v1 · submitted 2026-05-11 · 💻 cs.AI

Recognition: 2 theorem links

· Lean Theorem

TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding

Lihuan Li , Wilson Wongso , Baiyu Chen , Hao Xue , Ruiyi Yang , Yifan Duan , Xiachong Lin , Yang Song

show 1 more author

Flora Salim

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:45 UTC · model grok-4.3

classification 💻 cs.AI

keywords urban trajectorieslanguage groundingmulti-task benchmarktrajectory generationtrajectory retrievaltrajectory captioningtravel intent taxonomyurban mobility

0 comments

The pith

TrajPrism benchmark shows geometry-only models leave large gaps on language-trajectory tasks that language-aware models can close.

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

The paper introduces TrajPrism to evaluate models that must handle both geometric trajectories and natural-language descriptions of travel intent on the same real urban data. It defines three aligned tasks—instruction-conditioned generation, language-driven retrieval, and trajectory captioning—built from 300K trajectories across three cities and 2.1M instances created with a four-dimensional travel-intent taxonomy. Proof-of-concept models demonstrate that baselines using only spatial coordinates underperform once language enters the input or output, establishing the need for joint modeling. This matters because everyday mobility decisions combine paths with constraints and preferences that pure geometry cannot express.

Core claim

TrajPrism unifies instruction-conditioned trajectory generation, language-driven semantic trajectory retrieval, and trajectory captioning on 300K real trajectories from Porto, San Francisco, and Beijing. Language annotations are produced under a four-dimensional travel-intent taxonomy and judge-filtered to yield 2.1M task instances. The models TrajAnchor, TrajFuse, and TrajRap instantiate the tasks and show that geometry-only trajectory baselines leave a large gap on the protocol, especially where language forms part of the input-output interface.

What carries the argument

The four-dimensional travel-intent taxonomy used to generate and filter language annotations that link travel intents, constraints, and preferences to observed trajectories, together with the unified evaluation protocol that jointly scores trajectory fidelity, retrieval quality, and language groundedness.

If this is right

Instruction-based generation of urban routes requires joint language and geometry modeling to match real traveler behavior.
Semantic retrieval of trajectories from natural-language queries outperforms spatial-only matching on the benchmark.
Accurate trajectory captioning depends on grounding descriptions in both path shape and semantic intent rather than geometry alone.
The annotation pipeline and code can be reused on new cities that supply compatible trajectory data and map resources.
Models must be evaluated on all three tasks together to measure true language-trajectory alignment.

Where Pith is reading between the lines

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

Hybrid language-geometry architectures may become standard for any AI system that plans or explains movement in cities.
The benchmark could support training of multimodal models that accept spoken or written mobility requests and output verifiable routes.
Extending the taxonomy or adding cities would test whether the observed language advantage generalizes beyond the current three locations.
If the gap persists across larger models, it suggests fundamental limits to purely geometric representations of human travel.

Load-bearing premise

The judge-filtered language annotations generated under the four-dimensional travel-intent taxonomy accurately and consistently capture the travel intents, constraints, and preferences present in the underlying real-world trajectories.

What would settle it

Re-annotating a held-out sample of trajectories with the same taxonomy and measuring inter-annotator agreement or correlation with independent human descriptions of the same trips; low agreement or weak correlation would show the annotations do not reliably represent the underlying intents.

Figures

Figures reproduced from arXiv: 2605.10782 by Baiyu Chen, Flora Salim, Hao Xue, Lihuan Li, Ruiyi Yang, Wilson Wongso, Xiachong Lin, Yang Song, Yifan Duan.

**Figure 2.** Figure 2: Dataset generation pipeline. Steps 1–4 constitute the Reverse Intent Reconstruction (RIR) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: TrajAnchor pipeline (Task 1). Step 1 retrieves similar training trajectories; Step 2 extracts and grounds spatial constraints via LLM; Step 3 generates the route via chain Dijkstra. Bottom: Porto example with ground truth on the left. control stages covering noun/location/phase grounding, GIS terminology correction, LLM-based hallucination verification, lexical diversity enforcement, and punctuation saniti… view at source ↗

**Figure 4.** Figure 4: TrajFuse architecture (Task 2). A dual-encoder framework fuses geometric trajectory embeddings (from a fine-tuned TrajCL encoder) with H3-cell semantic embeddings and aligns them with text query embeddings via contrastive learning for cross-modal trajectory retrieval. TrajFuse (Task 2) addresses language-driven trajectory retrieval via a dual-encoder contrastive framework ( [PITH_FULL_IMAGE:figures/full_f… view at source ↗

**Figure 5.** Figure 5: TrajRap pipeline (Task 3). Similar training trajectories are retrieved and their gold captions serve as few-shot examples, which are fed alongside the test trajectory’s structural features to an LLM for factual captioning. Start GT end Good (Jac=0.923) Jac=0.923 Dest Haus=0.202 km Start GT end Average (Jac=0.429) Jac=0.429 Dest Haus=0.785 km Start GT end Poor (Jac=0.038) Jac=0.038 Dest Haus=2.062 km Ground… view at source ↗

**Figure 6.** Figure 6: Task 1 qualitative comparison: good, moderate, and poor predictions. Each column shows the ground-truth trajectory (blue) and the TrajAnchor prediction (red) for one Porto test case. Left: the predicted route closely follows the ground truth with correct destination and waypoints. Middle: the destination is approximately correct but the predicted route diverges in the middle segment. Right: the destination… view at source ↗

**Figure 7.** Figure 7: Task 1 Jaccard (H3) by instruction style across three cities. Performance is broken down by the three instruction variants (Literal, Concise, Chatty) for each baseline and TrajAnchor. A.3 TrajPrism Construction Pipeline Details A.3.1 Trajectory and Map Inputs [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

**Figure 8.** Figure 8: Task 2 retrieval MRR@10 by intent focus dimension across three cities. Each axis corresponds to one of the ten intent subcategories in the TrajPrism taxonomy [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

**Figure 9.** Figure 9: Distribution of the number of phases per trajectory after H3-based compression. The [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

**Figure 10.** Figure 10: Conditional distribution of fine-grained semantic subcategories in TrajPrism. Percentages [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗

**Figure 11.** Figure 11: Distribution of personas used for generating navigation instructions in TrajPrism. Uniform [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗

**Figure 12.** Figure 12: Distribution of the number of intent scenarios assigned to each generated trajectory. The [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗

**Figure 13.** Figure 13: Token length distribution of generated navigation instructions across three stylistic variants [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗

**Figure 14.** Figure 14: Qualitative example from San Francisco (traj_id: 475). Left: ground-truth trajectory on [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗

**Figure 15.** Figure 15: Qualitative example from Beijing (traj_id: 9331). Left: ground-truth trajectory on the [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗

**Figure 16.** Figure 16: Qualitative example from Porto (traj_id: 1373490). Top-left: ground-truth trajectory. [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗

read the original abstract

Urban mobility is naturally expressed both as trajectories in space and as natural-language descriptions of travel intent, constraints, and preferences. However, prior work rarely evaluates these two modalities together on the same real-world trajectories: trajectory modeling often stays geometry-centric, while language-centric mobility benchmarks frequently target route planning and tool use rather than fine-grained, verifiable alignment between text and the underlying route. We introduce TrajPrism, a multi-task benchmark for language-trajectory alignment that unifies (i) instruction-conditioned trajectory generation, (ii) language-driven semantic trajectory retrieval, and (iii) trajectory captioning, together with an evaluation protocol that measures trajectory fidelity, retrieval quality, and language groundedness. We construct TrajPrism by pairing real urban trajectories with judge-filtered language annotations generated under a four-dimensional travel-intent taxonomy. The benchmark contains 300K selected trajectories across Porto, San Francisco, and Beijing, yielding 2.1M task instances from three instruction variants, three retrieval queries, and one caption per trajectory. We further develop proof-of-concept models for each task: TrajAnchor for instruction-conditioned trajectory generation, TrajFuse for semantic trajectory retrieval, and TrajRap for trajectory captioning. These models instantiate the proposed tasks and show that geometry-only trajectory baselines leave a large gap on our protocol, especially where language is part of the input-output interface. We release TrajPrism with code and a reproducible annotation pipeline that is designed to be portable across cities, given compatible trajectory inputs and map resources.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

TrajPrism gives a clean new multi-task benchmark on real trajectories with language, but the claimed gaps over geometry baselines hinge on unshown annotation quality.

read the letter

The paper's core contribution is a benchmark that puts the same 300K real trajectories from Porto, San Francisco, and Beijing into three tasks at once: generating trajectories from instructions, retrieving them from language queries, and captioning them. It also ships a portable annotation pipeline and POC models for each task. That unification on actual urban data is the part that stands out and could be useful for anyone trying to move beyond pure geometry or pure route-planning setups. The release of code and the pipeline is a practical plus for reproducibility across cities with similar inputs. The abstract shows they ran the POC models and report that geometry-only baselines fall short, especially on the language-involved tasks. That direction makes sense as a starting point for showing why joint modeling matters. The main soft spot is the language annotations themselves. They come from a four-dimensional taxonomy and judge-filtering, but the text gives no numbers on agreement rates, error against re-annotation, or mismatch checks. If those annotations carry noise or systematic bias, the performance gaps could trace back to how the data was labeled rather than to genuine trajectory-language alignment. The soundness score in the report flags exactly this gap in quantitative support. The circularity risk looks low since the trajectories are external and the tasks are defined separately. This is the kind of paper that belongs in a reading group for people working on multimodal mobility models or trajectory benchmarks. Readers who need a ready evaluation suite for instruction-conditioned generation or language-driven retrieval will get the most out of it, provided the annotation details hold up under scrutiny. It deserves a serious referee round because the benchmark construction and task definitions are substantive enough to warrant external checks on the data pipeline and the baseline comparisons. I would send it to review with a request for the missing annotation validation stats.

Referee Report

2 major / 1 minor

Summary. The paper introduces TrajPrism, a multi-task benchmark for language-grounded urban trajectory understanding. It constructs 300K real trajectories from Porto, San Francisco, and Beijing paired with judge-filtered language annotations under a four-dimensional travel-intent taxonomy, producing 2.1M task instances across instruction-conditioned trajectory generation, language-driven semantic retrieval, and trajectory captioning. Proof-of-concept models (TrajAnchor, TrajFuse, TrajRap) are developed and used to show that geometry-only trajectory baselines leave a large performance gap on the proposed protocol, especially on tasks involving language in the input-output interface. The benchmark, code, and reproducible annotation pipeline are released.

Significance. If the judge-filtered annotations prove faithful to the underlying real-world trajectories, TrajPrism would offer a useful standardized resource for evaluating multimodal alignment in urban mobility, bridging geometry-centric trajectory modeling with language-based intent understanding. The multi-city scale, multi-task design, and emphasis on releasing a portable pipeline are strengths that could facilitate follow-on work.

major comments (2)

[Benchmark Construction / Annotation Pipeline] The manuscript provides no quantitative validation of the judge-filtered annotations (e.g., inter-judge agreement rates, error rates on a re-annotated sample, or systematic mismatch analysis against the 300K trajectories). This is load-bearing for the central claim that geometry-only baselines leave a large gap, because noisy or inconsistent annotations under the four-dimensional taxonomy could artifactually inflate language-aware model performance while making geometry-only baselines appear deficient by construction.
[Proof-of-Concept Models and Experiments] The description of the POC models and evaluation protocol supplies no concrete quantitative metrics, baseline implementation details, or error analysis supporting the claimed performance gap. Without these, it is not possible to assess the magnitude, statistical significance, or robustness of the reported differences between geometry-only and language-grounded approaches.

minor comments (1)

[Introduction] The abstract and introduction could more explicitly define the four-dimensional travel-intent taxonomy (e.g., list the dimensions and their values) to improve readability for readers unfamiliar with the annotation scheme.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important aspects of benchmark validation and experimental transparency. We address each major comment below and commit to revisions that strengthen the manuscript without altering its core contributions.

read point-by-point responses

Referee: [Benchmark Construction / Annotation Pipeline] The manuscript provides no quantitative validation of the judge-filtered annotations (e.g., inter-judge agreement rates, error rates on a re-annotated sample, or systematic mismatch analysis against the 300K trajectories). This is load-bearing for the central claim that geometry-only baselines leave a large gap, because noisy or inconsistent annotations under the four-dimensional taxonomy could artifactually inflate language-aware model performance while making geometry-only baselines appear deficient by construction.

Authors: We agree that the absence of quantitative annotation validation is a substantive gap in the current manuscript. The description of the judge-filtered process and four-dimensional taxonomy is provided, but no inter-judge agreement statistics, re-annotation error rates, or mismatch analysis appear in the text. In the revised version we will add a new subsection under Benchmark Construction that reports these metrics on a sampled subset of trajectories (e.g., Cohen’s kappa or raw agreement percentages across judges, error rates from independent re-annotation, and a breakdown of any systematic mismatches with the underlying GPS data). These statistics will be computed from the existing annotation logs and pipeline we release, directly supporting the reliability of the 2.1M task instances. revision: yes
Referee: [Proof-of-Concept Models and Experiments] The description of the POC models and evaluation protocol supplies no concrete quantitative metrics, baseline implementation details, or error analysis supporting the claimed performance gap. Without these, it is not possible to assess the magnitude, statistical significance, or robustness of the reported differences between geometry-only and language-grounded approaches.

Authors: We acknowledge that the experimental reporting in the manuscript is insufficiently detailed for independent assessment of the performance gaps. While the abstract and main text state that geometry-only baselines leave a large gap, the full quantitative tables, exact baseline implementations (architectures, hyperparameters, training details), statistical significance tests, and error analysis are not presented at the required level of concreteness. In the revision we will expand the Experiments section with complete per-task metric tables (including means, variances, and significance where appropriate), precise descriptions of all baselines and POC models (TrajAnchor, TrajFuse, TrajRap), and a dedicated error analysis subsection that examines representative failure cases for both geometry-only and language-grounded models. This will allow readers to evaluate the magnitude and robustness of the reported differences. revision: yes

Circularity Check

0 steps flagged

No circularity: benchmark built from external trajectories and annotations; claims rest on empirical protocol

full rationale

The paper constructs TrajPrism by pairing 300K real trajectories from Porto, San Francisco, and Beijing with judge-filtered language annotations under a four-dimensional taxonomy, yielding 2.1M task instances. It defines three tasks (instruction-conditioned generation, semantic retrieval, captioning) and instantiates them with proof-of-concept models (TrajAnchor, TrajFuse, TrajRap). The central claim—that geometry-only baselines leave a large gap—is an empirical comparison on the proposed protocol measuring fidelity, retrieval quality, and language groundedness. No equations, fitted parameters, or derivations are present that reduce to the inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. The chain is self-contained against external real-world data and the newly defined evaluation protocol.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the constructed language annotations faithfully represent travel intent; no free parameters, additional axioms beyond standard data-processing assumptions, or invented entities are described in the abstract.

axioms (1)

domain assumption The four-dimensional travel-intent taxonomy adequately captures relevant aspects of urban travel for the purpose of creating language annotations.
Invoked when generating and filtering the language annotations paired with trajectories.

pith-pipeline@v0.9.0 · 5601 in / 1170 out tokens · 39516 ms · 2026-05-12T04:45:41.499850+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

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

We construct TrajPrism by pairing real urban trajectories with judge-filtered language annotations generated under a four-dimensional travel-intent taxonomy... three tasks: instruction-conditioned trajectory generation, language-driven semantic trajectory retrieval, and trajectory captioning
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

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

TrajAnchor... retrieval + constraint extraction + chain Dijkstra... TrajFuse fuses geometric and semantic representations... TrajRap retrieval-augmented captioning

What do these tags mean?

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

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 3 internal anchors

[1]

gpt-oss-120b & gpt-oss-20b Model Card

S. Agarwal, L. Ahmad, J. Ai, S. Altman, A. Applebaum, E. Arbus, R. K. Arora, Y . Bai, B. Baker, H. Bao, et al. gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

Banerjee and A

S. Banerjee and A. Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. InProceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72, 2005

work page 2005
[3]

J. Cao, T. Zheng, Q. Guo, Y . Wang, J. Dai, S. Liu, J. Yang, J. Song, and M. Song. Holistic semantic representation for navigational trajectory generation.arXiv preprint arXiv:2501.02737, 2025

work page arXiv 2025
[4]

Chang, J

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

work page 2023
[5]

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

work page 2005
[6]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen, et al. Gemini 2.5: Pushing the frontier with advanced rea- soning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[7]

L. Gong, Y . Lin, X. Zhang, Y . Lu, X. Han, Y . Liu, S. Guo, Y . Lin, and H. Wan. Mobility-llm: Learning visiting intentions and travel preference from human mobility data with large language models.Advances in Neural Information Processing Systems, 37:36185–36217, 2024

work page 2024
[8]

J. Han, Y . Ning, Z. Yuan, H. Ni, F. Liu, T. Lyu, and H. Liu. Large language model powered intel- ligent urban agents: Concepts, capabilities, and applications.arXiv preprint arXiv:2507.00914, 2025

work page arXiv 2025
[9]

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, 2022

work page 2022
[10]

Keogh and C

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

work page 2005
[11]

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

work page arXiv 2025
[12]

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

work page 2024
[13]

X. Li, K. Zhao, G. Cong, C. S. Jensen, and W. Wei. Deep representation learning for trajectory similarity computation. In2018 IEEE 34th international conference on data engineering (ICDE), pages 617–628. IEEE, 2018

work page 2018
[14]

Z. Li, L. Xia, J. Tang, Y . Xu, L. Shi, L. Xia, D. Yin, and C. Huang. Urbangpt: Spatio-temporal large language models. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 5351–5362, 2024

work page 2024
[15]

C.-Y . Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004

work page 2004
[16]

T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014. 11

work page 2014
[17]

Y . Lin, Z. Zhou, Y . Liu, H. Lv, H. Wen, T. Li, Y . Li, C. S. Jensen, S. Guo, Y . Lin, et al. Unite: A survey and unified pipeline for pre-training spatiotemporal trajectory embeddings.IEEE Transactions on Knowledge and Data Engineering, 37(3):1475–1494, 2024

work page 2024
[18]

Z. Ma, Z. Tu, X. Chen, Y . Zhang, D. Xia, G. Zhou, Y . Chen, Y . Zheng, and J. Gong. More than routing: Joint gps and route modeling for refine trajectory representation learning. In Proceedings of the ACM Web Conference 2024, pages 3064–3075, 2024

work page 2024
[19]

Nomic embed: Training a reproducible long context text embedder.arXiv preprint arXiv:2402.01613, 2024

Z. Nussbaum, J. X. Morris, B. Duderstadt, and A. Mulyar. Nomic embed: Training a repro- ducible long context text embedder.arXiv preprint arXiv:2402.01613, 2024

work page arXiv 2024
[20]

B. A. Plummer, L. Wang, C. M. Cervantes, J. C. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. InProceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015

work page 2015
[21]

Qwen3.5: Towards native multimodal agents, February 2026

Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026

work page 2026
[22]

Z. Song, J. Zhang, C. Qin, C. Wang, C. Chen, L. Xu, K. Liu, X. Chu, and H. Zhu. Mobilitybench: A benchmark for evaluating route-planning agents in real-world mobility scenarios.arXiv preprint arXiv:2602.22638, 2026

work page arXiv 2026
[23]

J. Wang, R. Jiang, C. Yang, Z. Wu, M. Onizuka, R. Shibasaki, N. Koshizuka, and C. Xiao. Large language models as urban residents: An llm agent framework for personal mobility generation. Advances in Neural Information Processing Systems, 37:124547–124574, 2024

work page 2024
[24]

Y . Wang, C. Yang, J. Wang, X. Xu, J. Xu, D. Li, C. Xiao, and R. Jiang. Ellmob: Event-driven human mobility generation with self-aligned llm framework.arXiv preprint arXiv:2603.07946, 2026

work page arXiv 2026
[25]

Wongso, H

W. Wongso, H. Xue, and F. Salim. Genup: Generative user profilers as in-context learners for next poi recommender systems. InProceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems, pages 436–439, 2025

work page 2025
[26]

Wongso, H

W. Wongso, H. Xue, and F. D. Salim. Massive-steps: Massive semantic trajectories for understanding poi check-ins–dataset and benchmarks.arXiv preprint arXiv:2505.11239, 2025

work page arXiv 2025
[27]

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

work page 2017
[28]

J. Xie, K. Zhang, J. Chen, T. Zhu, R. Lou, Y . Tian, Y . Xiao, and Y . Su. Travelplanner: A benchmark for real-world planning with language agents.arXiv preprint arXiv:2402.01622, 2024

work page arXiv 2024
[29]

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

work page 2025
[30]

D. Yao, G. Cong, C. Zhang, and J. Bi. Computing trajectory similarity in linear time: A generic seed-guided neural metric learning approach. In2019 IEEE 35th international conference on data engineering (ICDE), pages 1358–1369. IEEE, 2019

work page 2019
[31]

BERTScore: Evaluating Text Generation with BERT

T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi. Bertscore: Evaluating text generation with bert.arXiv preprint arXiv:1904.09675, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1904
[32]

E. Zhao, P. Awasthi, Z. Chen, S. Gollapudi, and D. Delling. Semantic routing via autoregressive modeling.Advances in Neural Information Processing Systems, 37:10060–10087, 2024

work page 2024
[33]

Zheng, W.-L

L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in neural information processing systems, 36:46595–46623, 2023

work page 2023
[34]

Zheng, X

Y . Zheng, X. Xie, W.-Y . Ma, et al. Geolife: A collaborative social networking service among user, location and trajectory.IEEE Data Eng. Bull., 33(2):32–39, 2010. 12

work page 2010
[35]

Zheng, L

Y . Zheng, L. Zhang, X. Xie, and W.-Y . Ma. Mining interesting locations and travel sequences from gps trajectories. InProceedings of the 18th international conference on World wide web, pages 791–800, 2009

work page 2009
[36]

S. Zhou, Y . Chen, S. Shang, L. Chen, B. He, and R. Shibasaki. Blurred encoding for trajectory representation learning.arXiv preprint arXiv:2511.13741, 2025

work page arXiv 2025
[37]

Trips from PraçaGuilherme Gomes Fernandes in Cedofeitato Campo dos Mártiresda Pátriathat stay in park-like areas and avoid main shopping streets

Y . Zhu, J. J. Yu, X. Zhao, X. Zhou, L. Han, X. Wei, and Y . Liang. Unitraj: Learning a universal trajectory foundation model from billion-scale worldwide traces.arXiv preprint arXiv:2411.03859, 2024. 13 A Appendix A.1 Evaluation Metrics We provide formal definitions of all evaluation metrics used in TrajPrism. A.1.1 Task 1: Navigation Instruction Followi...

work page arXiv 2024
[38]

a quiet green area

Destination 1.1 Exact Anchor Specific physical destination (POI or road name) 1.2 Fuzzy Semantic Conceptual destination (e.g. “a quiet green area”)

work page
[39]

Waypoint 2.1 Strict Sequential Must pass through specific named road segments 2.2 Flexible / Feature Stop described by semantic features only 2.3 Pass-through Zone Cross an area type without stopping

work page
[40]

through chaos but main road

Route Pref. 3.1 Semantic Constraints Affinity or avoidance (e.g. parks, industrial zones) 3.2 Topological / Direct. Fluency, permeability, or directional preference 3.3 Orthogonal Comp. Semantic vs. topology (e.g. “through chaos but main road”)

work page
[41]

drive to the station, stop for fuel on the way, and avoid the highway

Temporal/Pace 4.1 Time-of-Day Route choice driven by time or day of week 4.2 Pace / Duration Urgency, leisure, or deadline-driven constraints the remaining scenarios from Dimensions 2–4, producing realistic multi-constraint requests such as “drive to the station, stop for fuel on the way, and avoid the highway”. Figure 12 shows the resulting distribution ...

work page