Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Pairing dual-layer LLMs with an enhanced GRASP planner yields higher-scoring, fully feasible multi-day city itineraries than pure-LLM or solver baselines.

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 · grok-4.5

2026-07-14 10:12 UTC pith:DAKMHPAX

load-bearing objection Competent multi-day hybrid that actually improves feasibility and one-iteration efficiency on real city POIs, but the headline S_avg gains are partly circular because the planner optimizes the same PE-PS score used for evaluation. the 3 major comments →

arxiv 2607.10651 v1 pith:DAKMHPAX submitted 2026-07-12 cs.AI

Embark Now: User Demand Oriented Framework for Multi-day Urban Travel Itinerary Planning

Rongbo Qi , Yaqi Zhang , Shijun Yan , Xuemeng Liu , Xiangrui Cai , Chunyao Song This is my paper
classification cs.AI
keywords travel itinerary planningmulti-day urban travelpersonalized travel plansLLM preference extractionGRASP heuristicPOI scoringhybrid planning
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.

Planning multi-day trips inside a large city is hard because there are many points of interest, travelers want different paces and meals, and every stop has opening hours and travel times. Pure language models can read a natural-language request but often invent places that do not exist or violate the schedule; classical optimizers need rigid parameter forms and can be slow. This paper claims that a two-stage hybrid fixes both problems: a dual-layer LLM first turns free text into structured preference parameters and personalised POI scores, then an enhanced GRASP heuristic builds and refines day-by-day routes under those constraints. On real Beijing and Tianjin POI databases plus 5 040 simulated preference cases the hybrid raises average itinerary score by at least 4.5 % and 11 % over strong baselines, keeps day-count failures near zero, guarantees every listed place is real, and finishes in fewer iterations than multi-pass alternatives. A sympathetic reader cares because the same pipeline lets a traveller obtain a usable, preference-aligned multi-day plan from one sentence of natural language.

Core claim

The authors establish that UDOIP—an end-to-end framework that first uses a dual-layer LLM pipeline to extract structured user parameters and personalised POI scores, then feeds those signals into a GRASP planner enhanced by geographic clustering and same-type substitution—produces multi-day urban itineraries that simultaneously satisfy hard spatiotemporal constraints, align with diverse free-text preferences, and outscore both pure-LLM generators and adapted solver-assisted baselines on two real city datasets.

What carries the argument

UDOIP’s dual-layer LLM input generator (overall assignment plus preference-classification fine-tuning) together with the preference-aware GRASP that inserts POIs via the PE-PS score s(p,u)=r_p·log2(n_r_p+1)·sim(p,u) and improves routes by clustering-and-substitution; this pipeline converts one natural-language request into a feasible, scored multi-day plan.

Load-bearing premise

The claim rests on the premise that LLM-simulated user requests and the PE-PS score used both for planning and for evaluation fairly represent real travellers and fairly rank methods that do not optimise that same score.

What would settle it

Re-run the identical Beijing/Tianjin evaluation on a held-out set of genuine multi-day travel requests written by human tourists (not LLM-generated) and recompute S_avg with an independent human preference ranking; if the hybrid’s gains disappear or reverse, the central claim fails.

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

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

3 major / 5 minor

Summary. The paper proposes UDOIP, a hybrid framework for Multi-day Urban Travel Itinerary Planning (MUTIP). A dual-layer LLM pipeline maps a single natural-language request into structured planner inputs (prioritized POIs L_Pu, time parameters including wu, dining bounds), and an enhanced Fuzzy-GRASP planner with PE-PS scoring (Eqs. 1–4), alternating construction/local search (Eq. 6), and a Clustering-and-Substitution strategy produces multi-day, constraint-feasible itineraries. On Beijing and Tianjin POI databases and 5,040 LLM-simulated preference cases, the authors report S_avg gains of at least 4.52% and 11.09% over LLM-only and adapted PTS baselines, near-zero day-count failures, 100% database-grounded POIs, preference-variability improvements (Kruskal–Wallis/Dunn), and ablation gains in spatial metrics and runtime from clustering+substitution (Table 7).

Significance. If the claims hold under fair evaluation, the work is a solid systems contribution at the LLM–OR interface: it targets multi-day single-city planning with opening hours, dining, pace, and mandatory POIs more completely than many recent hybrids (Table 1), and the dual-layer preference pipeline plus Clustering-and-Substitution are concrete engineering ideas. Strengths include real city POI data, ablations (Table 7), hallucination checks (Fig. 4), preference-variability tests, and a DeepSeek-R1 case study with maps and runtime. The main significance is practical controllability and efficiency for interactive itinerary tools rather than a new theoretical guarantee.

major comments (3)
  1. Primary metric circularity (§5.1.2 Eq. 7; §4.2.1 Eqs. 1–4; Alg. A.2 lines 22–24; Alg. 1 line 19). S_avg is the sum of PE-PS scores s(p,u), and the same s(p,u) is used to build RCL* and to select best_itinerary. LLM-only and adapted PTS baselines do not optimize PE-PS, so the abstract’s headline S_avg lifts (4.52%/11.09%) partly measure optimization of the evaluation objective rather than independent quality. Spatial metrics and R_days are secondary and do not fully underwrite the primary quality claim. Report a PE-PS-independent quality metric (held-out human ratings, fixed public ratings without sim(p,u), or constraint-satisfaction/coverage scores) for all methods, or re-score baselines under the same PE-PS objective with an equalized optimizer.
  2. Evaluation relies entirely on LLM-simulated demands (§5.1.1, Table 5: 2,640+2,400 cases). PE-PS similarity is computed from L_Pu extracted from the same style of LLM text used to generate the corpus, so preference-alignment and H statistics can be self-reinforcing. Without real-user or held-out human preference labels, the claim that UDOIP better captures diverse traveler needs is not fully supported. Add a real-user study or human preference ranking on a subset of itineraries.
  3. Baseline fairness and adaptation (§5.1.2, Table 6). PTS is adapted by replacing review-history re-ranking with keyword matching against the POI DB; LLM-only methods use a single-shot prompt without repair. Near-zero H for PTS and large S_avg gaps may reflect unequal access to PE-PS and constraint machinery rather than pure planner quality. Either give all methods the same PE-PS scores and mandatory-POI list as inputs, or clearly separate “preference extraction quality” from “planner quality” in the claims.
minor comments (5)
  1. Abstract vs Table 7: the abstract’s 17.95%/26.07% “computed metrics” and 4.64%/25.55% time-efficiency figures are hard to reconstruct from the per-metric Improvement rows; define the aggregation explicitly.
  2. Free parameters β, γ, RCL sizes, max_iter/fixed_iter, and wu ranges are listed but lack sensitivity analysis or default values in the main text.
  3. Fig. 4 similarity-threshold matching for “real-world POIs” needs a precise definition of the embedding match and threshold selection.
  4. Notation: tvisit_(p,u) vs tvisit_p and RCL* sizing are slightly inconsistent across §3–4 and the algorithms appendix.
  5. Related-work Table 1 is useful; a short paragraph on failure modes of SMT/MILP hybrids on multi-day city instances would strengthen the motivation for GRASP.

Circularity Check

3 steps flagged

Primary quality claim (S_avg gains) is partly by construction: the planner maximizes the same PE-PS score that defines the headline metric, disadvantaging non-optimizing baselines.

specific steps
  1. other [Eqs. 1–4, Eq. 7; Alg. A.2 lines 22–24; Alg. 1 line 19; Abstract / Table 6]
    "s(p,u) = r_p · log2(nr_p + 1) · sim(p,u) ... Construct the RCL* by sorting the POIs according to their scores and selecting the top RCL*_SIZE elements. ... if the ∑_k∈K ∑_i∈I s(p,u)_i Y^k_i of the current_itinerary exceeds that of the previously best_itinerary then Update best_itinerary ... S_avg = 1/N ∑_i (∑_j s(ij,u)) ... improving the average total itinerary score by at least 4.52% and 11.09%"

    The custom PE-PS score is simultaneously (i) the ranking objective inside RCL* construction, (ii) the selection criterion for best_itinerary, and (iii) the definition of the primary reported quality metric S_avg. Methods that do not optimize PE-PS (LLM-only baselines, adapted PTS) are therefore scored by a metric the proposed planner is explicitly maximizing. Headline S_avg gains over those baselines are partly by construction, not independent evidence of superior itinerary quality.

  2. self definitional [Sec. 4.1 dual-layer classifier; Sec. 5.1.2 Preference Variability metrics H1–H5; Table 6; Figs. 5–6]
    "The second layer, known as the classification fine-tuning layer, provides macro-level oversight ... categorizing user preferences in areas such as time and dining, and adjusting values that significantly deviate from expected ranges. ... We employed the Kruskal-Wallis test ... to evaluate the variability in results based on different categories of time and dining preferences, demonstrating the algorithm’s capability to capture user preferences. ... UDOIP (ours) ... H1 1393.58 ... H5 841.86"

    The dual-layer LLM is designed to classify demands into the same six time / two dining categories and to force parameters into category-specific ranges. Preference-capture is then demonstrated by Kruskal-Wallis H (and Dunn p-values) measuring output separation across those same author-defined categories. Higher H for UDOIP is partly the classifier’s own differentiation mechanism measured against its own labels, not an external test of preference understanding.

  3. fitted input called prediction [Sec. 4.2.1 PE-PS Formulas (1)–(4); Sec. 5.1.1 simulated user demands Table 5; case study Fig. 7]
    "emb_u = Average([emb_p1u, emb_p2u, …, emb_plu]) ... sim(p,u) = emb_p · emb_u / (∥emb_p∥ · ∥emb_u∥) ... s(p,u) = r_p · log2(nr_p + 1) · sim(p,u) ... simulated real user demand data constructed based on defined rules by LLMs, which spans all preference categories considered by the algorithm ... Total = 2640 (Beijing) + 2400 (Tianjin)"

    User embedding emb_u is the average of embeddings of the prioritized POI list L_Pu extracted from the demand text. The 5,040 evaluation cases are themselves LLM-simulated demands that already encode those preference types and L_Pu lists. PE-PS similarity therefore rewards POIs near the same LLM-extracted L_Pu that generated the case, and S_avg / case-study ‘alignment’ claims partly restate that construction rather than independently predicting held-out preference satisfaction.

full rationale

This is an engineering/systems paper, not a first-principles derivation paper, so classical self-citation or uniqueness circularity is absent. The load-bearing circularity is evaluation-objective alignment. The planner builds RCL* by sorting on the custom PE-PS score s(p,u) (Eqs. 1–4) and selects best_itinerary by maximizing sum s(p,u) (Alg. 1 line 19). The abstract’s primary quality claim is improvement in average total itinerary score S_avg, which Eq. 7 defines as exactly that same sum. LLM-only and adapted PTS baselines do not optimize PE-PS, so S_avg gains of ≥4.52%/11.09% are partly forced by metric design rather than independent quality evidence. Preference-variability H statistics similarly partly reflect the dual-layer classifier’s own category labels and parameter adjustments. Spatial metrics (D_total, A_overlap, θ_turns, etc.), R_days, and 100% database-grounded POIs are independent and non-circular, which is why the score is 5 rather than 8–10: the central claim is only partially reduced to its inputs, and secondary metrics still provide external grounding. No self-citation load-bearing chain or uniqueness theorem is involved.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

As an empirical systems paper the central claim rests on standard combinatorial-optimization and embedding assumptions plus several hand-chosen algorithmic knobs and the untested fidelity of LLM-simulated demand. No new physical entities are postulated; the free parameters are the usual GRASP/LLM hyperparameters that control construction and local search.

free parameters (6)
  • β (waiting-time weight in Flocal)
    Hyperparameter in Eq. 6 balancing waiting time vs travel time; range [0,1] stated but no sensitivity or selection procedure given.
  • γ (substitution time-saving threshold)
    Threshold that decides whether a cross-cluster same-type POI may replace a non-essential POI; value not reported.
  • RCL_SIZE / RCL*_SIZE
    Sizes of the restricted candidate lists that control greediness vs randomness; treated as fixed inputs without ablation.
  • wu (visit-duration compactness)
    User-specific multiplier on suggested visit times (Eq. 5); initialized by LLM and dynamically reduced; exact reduction schedule not fully specified.
  • max_iter / fixed_iter
    Outer and inner iteration budgets of Algorithm 1; experiments emphasize 1-iteration results but values are free choices.
  • k (number of geographic clusters = days)
    K-means cluster count set equal to trip length; assumes geographic separation is always desirable.
axioms (5)
  • domain assumption Travel times t_traffic between POIs obtained from map services are accurate and static for planning purposes.
    Used throughout construction and local search; no uncertainty model.
  • ad hoc to paper Public ratings r_p and rating counts n_r_p together with cosine similarity of LLM embeddings of name+description yield a score that correctly ranks POIs for a given user.
    PE-PS (Eqs. 1–4) is introduced without external validation against real choice data.
  • ad hoc to paper LLM-simulated natural-language demands spanning the enumerated preference types are representative of real multi-day urban travelers.
    All 5,040 evaluation cases are generated this way (Section 5.1.1).
  • domain assumption A dual-layer LLM (overall assignment + classification fine-tuning) produces algorithm inputs that stay inside feasible ranges after one interaction.
    Core of the preference pipeline; second layer only clamps to hand-defined category ranges.
  • standard math Standard GRASP construction + 2-opt-style local search with the Flocal objective yields near-optimal multi-day routes under the stated constraints.
    Builds on Fuzzy-GRASP (Expósito et al. 2019) without new optimality guarantees.
invented entities (3)
  • UDOIP (User Demand-Oriented Itinerary Planner) no independent evidence
    purpose: End-to-end hybrid framework name for the dual-layer LLM + enhanced GRASP pipeline.
    Organizational label; no independent existence outside the paper.
  • PE-PS (Personalized LLM-enhanced POI Scoring) no independent evidence
    purpose: Formula that multiplies rating, log rating-count, and embedding similarity to produce user-specific POI scores used both for ranking and for S_avg.
    New scoring function introduced in Section 4.2.1; validated only by internal case study.
  • Clustering-and-Substitution strategy inside GRASP no independent evidence
    purpose: Pre-cluster POIs by geography then allow same-type cross-cluster swaps when priority POIs are missed.
    Algorithmic device (Algorithm A.4) that improves spatial metrics; not a physical entity.

pith-pipeline@v1.1.0-grok45 · 33264 in / 3873 out tokens · 43055 ms · 2026-07-14T10:12:32.571488+00:00 · methodology

0 comments
read the original abstract

In large urban areas, planning multi-day travel itineraries is challenging due to the abundance of Points of Interest (POIs), diverse user preferences, and constraints such as opening hours. Effective solutions must dynamically accommodate diverse traveler requirements while optimizing for satisfaction and feasibility within limited computation time. This paper addresses these challenges through introducing an innovative framework that integrates Large Language Models (LLMs) to dynamically capture user requirements with precision and flexibility, and an enhanced Greedy Randomized Adaptive Search Procedure (GRASP) algorithm as a well-suited preference-aware planner to generate feasible multi-day itineraries. The effectiveness of our integrated approach is demonstrated through extensive experiments on two real-world urban datasets from Beijing and Tianjin. Our framework significantly outperforms state-of-the-art (SOTA) methods, improving the average total itinerary score by at least 4.52% and 11.09% across 5,040 user cases with diverse preferences in the two datasets. Furthermore, through end-to-end algorithmic enhancements, it achieves notable average improvements of 17.95% and 26.07% in the computed metrics, while also delivering substantial gains in time efficiency -- realizing average performance increases of 4.64% and 25.55% within shorter computation times compared to suboptimal methods that require multiple iterations. These outcomes underscore our method's superiority in delivering both enhanced itinerary quality and computational efficiency over existing methodologies.

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

44 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    2025 , howpublished =

    OpenAI , title =. 2025 , howpublished =

    2025

  2. [2]

    Mariam Orabi and Imad Afyouni and Zaher Al Aghbari , title =. Inf. Process. Manag. , volume =. 2025 , url =. doi:10.1016/J.IPM.2024.103970 , timestamp =

    work page doi:10.1016/j.ipm.2024.103970 2025

  3. [3]

    To the Globe (TTG): Towards Language-Driven Guaranteed Travel Planning

    Da JU and Song Jiang and Andrew Cohen and Aaron Foss and Sasha Mitts and Arman Zharmagambetov and Brandon Amos and Xian Li and Justine T. Kao and Maryam Fazel. To the Globe. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2410.16456 , eprinttype =. 2410.16456 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.16456 2024

  4. [4]

    Zhang and Han Bao and Hanwei Xu and Haocheng Wang and Haowei Zhang and Honghui Ding and Huajian Xin and Huazuo Gao and Hui Li and Hui Qu and J

    DeepSeek-AI and Aixin Liu and Bei Feng and Bing Xue and Bingxuan Wang and Bochao Wu and Chengda Lu and Chenggang Zhao and Chengqi Deng and Chenyu Zhang and Chong Ruan and Damai Dai and Daya Guo and Dejian Yang and Deli Chen and Dongjie Ji and Erhang Li and Fangyun Lin and Fucong Dai and Fuli Luo and Guangbo Hao and Guanting Chen and Guowei Li and H. Zhang...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.19437 2024

  5. [5]

    DeepSeek-AI and Daya Guo and Dejian Yang and Haowei Zhang and Junxiao Song and Ruoyu Zhang and Runxin Xu and Qihao Zhu and Shirong Ma and Peiyi Wang and Xiao Bi and Xiaokang Zhang and Xingkai Yu and Yu Wu and Z. F. Wu and Zhibin Gou and Zhihong Shao and Zhuoshu Li and Ziyi Gao and Aixin Liu and Bing Xue and Bingxuan Wang and Bochao Wu and Bei Feng and Che...

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.12948 2025

  6. [6]

    Airam Exp. A fuzzy. Expert Syst. Appl. , volume =. 2019 , url =. doi:10.1016/J.ESWA.2019.03.004 , timestamp =

    work page doi:10.1016/j.eswa.2019.03.004 2019

  7. [7]

    Expert Syst

    Takwa Tlili and Saoussen Krichen , title =. Expert Syst. Appl. , volume =. 2021 , url =. doi:10.1016/J.ESWA.2021.115723 , timestamp =

    work page doi:10.1016/j.eswa.2021.115723 2021

  8. [8]

    Expert Syst

    Remigijus Paulavicius and Linas Stripinis and Simona Sutaviciute and Dmitrij Kocegarov and Ernestas Filatovas , title =. Expert Syst. Appl. , volume =. 2023 , url =. doi:10.1016/J.ESWA.2023.120580 , timestamp =

    work page doi:10.1016/j.eswa.2023.120580 2023

  9. [9]

    Symmetry , volume =

    Xiao Zhou and Jian Peng and Bowei Wen and Mingzhan Su , title =. Symmetry , volume =. 2023 , url =. doi:10.3390/SYM15122168 , timestamp =

    work page doi:10.3390/sym15122168 2023

  10. [10]

    Selective clustered tourist trip design problem with time windows under intuitionistic fuzzy score and exponential travel times , journal =

    Tusan Derya and Kumru Didem Atalay and Esra Dinler and Baris Ke. Selective clustered tourist trip design problem with time windows under intuitionistic fuzzy score and exponential travel times , journal =. 2024 , url =. doi:10.1016/J.ESWA.2024.124792 , timestamp =

    work page doi:10.1016/j.eswa.2024.124792 2024

  11. [11]

    Expert Syst

    Angelo Aliano Filho and Reinaldo Morabito , title =. Expert Syst. Appl. , volume =. 2024 , url =. doi:10.1016/J.ESWA.2023.122437 , timestamp =

    work page doi:10.1016/j.eswa.2023.122437 2024

  12. [12]

    Sustainable group tourist trip planning: An adaptive large neighborhood search algorithm , journal =

    Mansoureh Hasannia Kolaee and Armin Jabbarzadeh and Seyed Mohammad Javad Mirzapour Al. Sustainable group tourist trip planning: An adaptive large neighborhood search algorithm , journal =. 2024 , url =. doi:10.1016/J.ESWA.2023.121375 , timestamp =

    work page doi:10.1016/j.eswa.2023.121375 2024

  13. [13]

    Journal of Cleaner Production , volume=

    Multi-objective sustainability tourist trip design: An innovative approach for balancing tourists' preferences with key sustainability considerations , author=. Journal of Cleaner Production , volume=. 2024 , publisher=

    2024

  14. [14]

    Forty-first International Conference on Machine Learning,

    Jian Xie and Kai Zhang and Jiangjie Chen and Tinghui Zhu and Renze Lou and Yuandong Tian and Yanghua Xiao and Yu Su , title =. Forty-first International Conference on Machine Learning,. 2024 , url =

    2024

  15. [15]

    Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies,

    Yilun Hao and Yongchao Chen and Yang Zhang and Chuchu Fan , title =. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies,. 2025 , url =. doi:10.18653/V1/2025.NAACL-LONG.176 , timestamp =

    doi:10.18653/v1/2025.naacl-long.176 2025

  16. [16]

    CoRR , volume =

    Atharva Gundawar and Mudit Verma and Lin Guan and Karthik Valmeekam and Siddhant Bhambri and Subbarao Kambhampati , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2405.20625 , eprinttype =. 2405.20625 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.20625 2024

  17. [17]

    CoRR , volume =

    Tom. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2406.10196 , eprinttype =. 2406.10196 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.10196 2024

  18. [18]

    CoRR , volume =

    Yansong Ning and Rui Liu and Jun Wang and Kai Chen and Wei Li and Jun Fang and Kan Zheng and Naiqiang Tan and Hao Liu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2509.21842 , eprinttype =. 2509.21842 , timestamp =

    doi:10.48550/arxiv.2509.21842 2025

  19. [19]

    arXiv preprint arXiv:2601.04698 , year=

    TourPlanner: A Competitive Consensus Framework with Constraint-Gated Reinforcement Learning for Travel Planning , author=. arXiv preprint arXiv:2601.04698 , year=

    arXiv

  20. [20]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing:

    Yihong Tang and Zhaokai Wang and Ao Qu and Yihao Yan and Zhaofeng Wu and Dingyi Zhuang and Jushi Kai and Kebing Hou and Xiaotong Guo and Jinhua Zhao and Zhan Zhao and Wei Ma , title =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing:. 2024 , url =. doi:10.18653/V1/2024.EMNLP-INDUSTRY.104 , timestamp =

    doi:10.18653/v1/2024.emnlp-industry.104 2024

  21. [21]

    Pieter Vansteenwegen and Wouter Souffriau and Dirk Van Oudheusden , title =. Eur. J. Oper. Res. , volume =. 2011 , url =

    2011

  22. [22]

    Tourism Management , volume=

    Using a heuristic algorithm to design a personalized day tour route in a time-dependent stochastic environment , author=. Tourism Management , volume=. 2018 , publisher=

    2018

  23. [23]

    Grazia Speranza , title =

    Nicola Bianchessi and Renata Mansini and M. Grazia Speranza , title =. Int. Trans. Oper. Res. , volume =. 2018 , url =. doi:10.1111/ITOR.12422 , timestamp =

    work page doi:10.1111/itor.12422 2018

  24. [24]

    Variable neighborhood search to solve the generalized orienteering problem , journal =

    Adolfo. Variable neighborhood search to solve the generalized orienteering problem , journal =. 2021 , url =. doi:10.1111/ITOR.12800 , timestamp =

    work page doi:10.1111/itor.12800 2021

  25. [25]

    Solving the stochastic time-dependent orienteering problem with time windows , journal =

    C. Solving the stochastic time-dependent orienteering problem with time windows , journal =. 2016 , url =. doi:10.1016/J.EJOR.2016.05.031 , timestamp =

    work page doi:10.1016/j.ejor.2016.05.031 2016

  26. [26]

    Pardalos , title =

    Wanzhe Hu and Mahdi Fathi and Panos M. Pardalos , title =. Appl. Soft Comput. , volume =. 2018 , url =. doi:10.1016/J.ASOC.2018.08.026 , timestamp =

    work page doi:10.1016/j.asoc.2018.08.026 2018

  27. [27]

    1999 , url =

    Jo. 1999 , url =. doi:10.1109/12.769433 , timestamp =

    doi:10.1109/12.769433 1999

  28. [28]

    Pantziou , title =

    Damianos Gavalas and Charalampos Konstantopoulos and Konstantinos Mastakas and Grammati E. Pantziou , title =. J. Heuristics , volume =. 2014 , url =. doi:10.1007/S10732-014-9242-5 , timestamp =

    work page doi:10.1007/s10732-014-9242-5 2014

  29. [29]

    CoRR , volume =

    OpenAI , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2303.08774 , eprinttype =. 2303.08774 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774 2023

  30. [30]

    2024 , howpublished =

    OpenAI , title =. 2024 , howpublished =

    2024

  31. [31]

    Forty-first International Conference on Machine Learning,

    Subbarao Kambhampati and Karthik Valmeekam and Lin Guan and Mudit Verma and Kaya Stechly and Siddhant Bhambri and Lucas Saldyt and Anil Murthy , title =. Forty-first International Conference on Machine Learning,. 2024 , url =

    2024

  32. [32]

    CoRR , volume =

    Karthik Valmeekam and Kaya Stechly and Subbarao Kambhampati , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2409.13373 , eprinttype =. 2409.13373 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2409.13373 2024

  33. [33]

    Smart Cities , volume=

    On-site trip planning support system based on dynamic information on tourism spots , author=. Smart Cities , volume=. 2020 , publisher=

    2020

  34. [34]

    Solving the team orienteering problem with time windows and mandatory visits by multi-start simulated annealing , journal =

    Shih. Solving the team orienteering problem with time windows and mandatory visits by multi-start simulated annealing , journal =. 2017 , url =. doi:10.1016/J.CIE.2017.10.020 , timestamp =

    work page doi:10.1016/j.cie.2017.10.020 2017

  35. [35]

    International Journal of Logistics Systems and Management , volume=

    -GRASP with bi-directional path relinking for the bi-objective orienteering problem , author=. International Journal of Logistics Systems and Management , volume=. 2018 , publisher=

    2018

  36. [36]

    Sustainability , volume=

    Multi-objective fuzzy tourist trip design problem with heterogeneous preferences and sustainable itineraries , author=. Sustainability , volume=. 2021 , publisher=

    2021

  37. [37]

    Sujan Piya and Chefi Triki and Abdulwahab Al Maimani and Mahdi Mokhtarzadeh , title =. Comput. Ind. Eng. , volume =. 2023 , url =. doi:10.1016/J.CIE.2022.108839 , timestamp =

    work page doi:10.1016/j.cie.2022.108839 2023

  38. [38]

    Statistical science , volume=

    Simulated annealing , author=. Statistical science , volume=. 1993 , publisher=

    1993

  39. [39]

    The corsini encyclopedia of psychology , pages=

    Kruskal-wallis test , author=. The corsini encyclopedia of psychology , pages=. 2010 , publisher=

    2010

  40. [40]

    The Stata Journal , volume=

    Nonparametric pairwise multiple comparisons in independent groups using Dunn's test , author=. The Stata Journal , volume=. 2015 , publisher=

    2015

  41. [41]

    Sajal Halder and Kwan Hui Lim and Jeffrey Chan and Xiuzhen Zhang , title =. Appl. Soft Comput. , volume =. 2024 , url =. doi:10.1016/J.ASOC.2023.111200 , timestamp =

    work page doi:10.1016/j.asoc.2023.111200 2024

  42. [42]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

    Zijian Shao and Jiancan Wu and Weijian Chen and Xiang Wang , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2025 , url =

    2025

  43. [43]

    On planning sightseeing tours with TripBuilder , journal =

    Igo Ramalho Brilhante and Jos. On planning sightseeing tours with TripBuilder , journal =. 2015 , url =. doi:10.1016/J.IPM.2014.10.003 , timestamp =

    work page doi:10.1016/j.ipm.2014.10.003 2015

  44. [44]

    Yerin Yhee and Seunghun Shin and Eunji Lee and Jungkeun Kim , title =. Inf. Process. Manag. , volume =. 2023 , url =. doi:10.1016/J.IPM.2023.103281 , timestamp =

    work page doi:10.1016/j.ipm.2023.103281 2023