Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

This paper formally defines itinerary modification as a task and builds iTIMO, a synthetic dataset of need-to-modify itineraries generated by perturbing real-world itineraries.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 10:14 UTC pith:7F3MF5U4

load-bearing objection A solid first benchmark for itinerary repair, with one load-bearing construct-validity question the appendix doesn't fully close. the 4 major comments →

arxiv 2601.10609 v5 pith:7F3MF5U4 submitted 2026-01-15 cs.IR

iTIMO: An LLM-empowered Synthesis Dataset for Travel Itinerary Modification

Zhuoxuan Huang , Yunshan Ma , Hongyu Zhang , Hua Ma , Zhu Sun This is my paper
classification cs.IR
keywords itinerary modificationsynthetic datasetlarge language modelstravel recommender systemsintent-driven perturbationdata generation pipelineitinerary repair
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.

Travel plans rarely survive contact with reality: users add, drop, or swap attractions before and during a trip. The authors argue this repair function is missing from travel recommender research because no existing dataset contains itineraries that need fixing. They define the itinerary modification task, then generate paired data by instructing a large language model to perturb real itineraries with three operations (ADD, REPLACE, DELETE) under three intents (popularity, distance, diversity), filtering the results through hybrid metrics. The resulting iTIMO dataset enables benchmarking, and the experiments show that small fine-tuned models can rival much larger reasoning models, while combining retrieval with fine-tuning only helps if the two are aligned.

Core claim

The paper claims to be the first to formulate itinerary modification as a formal task, defined as the inverse of an intent-driven perturbation: a perturbation generates a need-to-modify itinerary by applying an operation that disrupts specified attributes, and modification reverts it. To create training and evaluation data at scale, it constructs iTIMO from three public real-world itinerary collections by prompting a large language model to perform these perturbations, validating each one with metrics: category-diversity change for the diversity intent, and Hellinger distance or Kendall's tau-b for popularity and spatial distance. Benchmarking on iTIMO yields the paper's empirical claims: cu

What carries the argument

The machinery is the intent-driven perturbation formulation together with a metric-based quality filter. The task definitions encode modification as the inverse of perturbation. The pipeline drives a large language model with function-calling tools (exact distance computation and distribution/ranking metrics) and a memory module that discourages repeating POIs and positions, then filters outputs using the hybrid metrics with a Hellinger threshold of 0.1. These metrics are the load-bearing validator: they decide which perturbations count as need-to-modify and later score modification accuracy.

Load-bearing premise

A perturbation counts as need-to-modify only if it trips the paper's metric thresholds (Hellinger distance above 0.1 or a rank-order change, plus a category-diversity change), so any real-world modification with smaller or differently typed shifts is systematically absent from the dataset.

What would settle it

Collect a set of real user modification requests (for example, from travel forums or a diary study), encode each as an itinerary pair, and compute what fraction triggers the paper's metric conditions. If most real changes fall below the threshold, the benchmark omits the common case.

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

If this is right

  • iTIMO supplies the first benchmark where itinerary repair can be measured, letting travel recommenders move from static planning to adaptive modification.
  • Supervised fine-tuning is a promising route: small open models fine-tuned on iTIMO achieve competitive modification accuracy, lowering the barrier to building repair-capable assistants.
  • ADD is the hardest operation and DELETE the easiest, pointing future work toward multi-step selection and insertion-position reasoning.
  • RAG helps base models, but fine-tuned models only benefit when training and inference prompts align; naive combination can hurt.
  • Full fine-tuning beats low-rank adaptation on small datasets, while low-rank adaptation becomes competitive as data grows.

Where Pith is reading between the lines

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

  • If the metric-based filtering generalizes, the same perturb-then-repair pipeline could synthesize revision data for other sequential planning domains, such as meal planning or daily scheduling, where need-to-modify examples are also scarce.
  • The Hellinger threshold of 0.1 is justified by a lower-bound proof for extreme distribution shifts only; real user modifications with subtle shifts may be underrepresented. A study that elicits actual user modification requests and checks what fraction triggers the same metric conditions would test this directly.
  • The intents in iTIMO inherit whatever preferences are latent in the source photo-sharing itineraries; because the paper itself notes inconsistent itinerary definitions in one city source, some of those preferences may be noisy.
  • The paired structure of iTIMO (perturbed versus original) makes it a natural testbed for studying when models detect that something is off before repairing 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

4 major / 4 minor

Summary. The paper formalizes the itinerary modification task (Definitions 1–2), proposes an LLM-based pipeline to construct the iTIMO dataset by perturbing real-world itineraries from Toronto, Melbourne, and Florence with ADD/REPLACE/DELETE operations under popularity, distance, and diversity intents, and introduces hybrid metrics (Hellinger distance, Kendall's tau-b, category-diversity ratio) to validate perturbations. It then benchmarks ten base LLMs and reasoning models, with RAG and SFT variants, on Mod and APR metrics, reporting that SFT enables 8B models to compete with much larger reasoning models. The appendices supply full prompts, a tool contract, dataset statistics, and lower-bound proofs for the Hellinger threshold.

Significance. The paper has real strengths: precise task definitions, carefully specified metrics, a detailed and reproducible pipeline with released prompts/dataset/code, and a broad benchmark. The exact-match Mod metric provides an external anchor. If the construct-validity concern is resolved, iTIMO would fill a genuine gap in travel recommendation research. However, the central 'first valid testbed' claim turns on an unvalidated threshold rule: the same H>0.1/tau_b<1/CD-change conditions are used to generate, filter, and score the data, and the appendix proof covers only three hand-built distribution families rather than the empirical population. The paper is a useful contribution, but the validity case is incomplete and needs additional evidence.

major comments (4)
  1. [§4.2 and Appendix A.6] The threshold theta=0.1 is load-bearing but not validated. Lemmas A.1–A.3 lower-bound H only for three hand-constructed extreme distribution families; they do not show that H>0.1, tau_b<1, or a CD change is necessary or sufficient for a genuine need-to-modify request in the actual itinerary population. The ADD bound gives n<=49, not n<=50 as stated. Since the same conditions filter iTIMO (§5.3) and define APR (§6.1), all benchmark conclusions inherit this rule. Please add human annotation of filtered vs. rejected perturbations, inter-annotator agreement, and a sensitivity sweep over theta (e.g., 0.05–0.2) and the tau_b/CD rules.
  2. [§6.1] APR is partly circular. It awards success iff hinted axes shift and non-hinted axes stay invariant, using exactly the H/tau_b/CD conditions that define need-to-modify during perturbation and filtering. Thus APR measures self-consistency with the generator's metric conditions, not semantic repair quality. The exact-match Mod metric is an independent anchor, but it only tests recovery of the original itinerary and does not validate naturalness. Please report human judgments on a sample of modification outputs and quantify APR successes that are Mod failures; otherwise the RQ1–RQ4 comparisons are partly self-referential.
  3. [Appendix A.1 / §5.3] The paper itself acknowledges inconsistent Florence itinerary definitions and truncates long itineraries with a sliding window of size 21. This arbitrary preprocessing weakens the premise that base itineraries faithfully encode user travel plans; the same concern applies to Flickr-derived Toronto/Melbourne data. Please report the fraction of truncated itineraries, sensitivity to window size, and a qualitative check that truncated sub-itineraries remain coherent plans. This bears directly on the 'real-world itineraries implicitly encode user intents' premise.
  4. [§4.3 and Boxes 8–10] Spatial distance 'disruption' uses dataset-global low/medium/high thresholds, and those thresholds are injected into the modification prompts. Perceived disruptiveness of distances is context-dependent (urban density, transport mode), so a global threshold may mislabel realistic modifications. Please justify the threshold choice, test sensitivity, and ideally validate against human perception.
minor comments (4)
  1. [Table 1 / §2.1] 'Flicker' should be 'Flickr'; reference [50] has 'Proceddings' typo.
  2. [Figure 3 / Table 4] The caption/legend glyphs for 'existing POI', 'perturbation position', and model-type markers appear to be missing in the rendered text; please ensure the symbols render correctly.
  3. [§5.2] The preliminary study reports only Melbourne, yet claims consistent trends across Toronto and Florence; include at least summary statistics or state why the trends are consistent.
  4. [Tables 4–7] No variance or significance tests are reported. Given small gaps in Table 5 (e.g., Borda 180 vs. 173.5), add seed-level variance or statistical tests for the main claims.

Circularity Check

1 steps flagged

Local circularity in V3.2 FM's reported 1.00 perturbation accuracy (enforced by tool-echo contract), while the central dataset/benchmark claim remains externally anchored.

specific steps
  1. self definitional [Section 5.2, Table 2, Appendix Box 4 (V3.2 FM ADD system prompt)]
    "Call stats_from_categories with the CANON arrays and domain=["High","Medium","Low"], thresholds={"hellinger":0.1,"tau_b":1.0}. ... A) Atomic Intents Builder (MUST OVERWRITE) ... pop_flag := stats_from_categories(popularity).disruption ... Overwrite JSON["Intents"] with exactly these (or ["UNKNOWN"] if none true)."

    Table 2 reports 'Pert. Acc.' as perturbation accuracy 'assessing the alignment between perturbation results and intents.' For V3.2 FM, however, the intents are not independently predicted by the model: the tool computes H, tau-b, and CD against the same thresholds used by the scorer, the prompt contract requires every disruption boolean to mirror the tool output exactly, and the Atomic Intents Builder overwrites the JSON Intents with the tool flags. Thus Pert. Acc. = 1.00 for V3.2 FM is true by construction; reporting it as a measured model capability is a restatement of the prompt contract rather than an empirical finding.

full rationale

Most of iTIMO's construction is self-contained and externally anchored: base itineraries come from three public datasets, the modification ground truth is the original real-world itinerary, and the benchmark's Mod metric requires exact recovery, so the central task has independent content. The pipeline transparently defines perturbation intents via the Section 4 metrics (H>0.1, tau_b<1, CD change), and APR consistently uses that same definition; this is a construct-validity choice rather than a circular derivation, since APR is not used to generate the ground truth. The one genuine circularity is localized to Table 2 / Section 5.2: V3.2 FM's perturbation accuracy of 1.00 is enforced by the tool-echo and Atomic Intents Builder contract, so presenting it as a measured capability is tautological. Appendix Lemmas A.1-A.3 only lower-bound H for three hand-constructed extreme families and do not validate theta=0.1 as a separator of real user modification intents; that is a weakness in external validity, not a circularity. No load-bearing self-citation chains or imported uniqueness theorems appear; self-citations are incidental. Score 4 reflects one prediction reduced by construction while the central dataset/benchmark claim stands on independent ground.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The dataset's validity rests on a small set of hand-chosen thresholds and domain assumptions. The Hellinger threshold θ=0.1 is the most consequential free parameter: it is used both to decide which perturbations enter the dataset and to judge whether a model's fix 'passes' the APR metric, so any miscalibration propagates through the entire benchmark. The spatial and popularity level cutoffs are data-derived but their exact values are not reported. The scope restriction to three operations and three intents, and the treatment of Flickr-derived itineraries as ground truth, are domain assumptions the paper itself flags as limitations.

free parameters (6)
  • Hellinger disruption threshold θ = 0.1
    Hand-set in §4.2; used both for dataset construction and APR evaluation. Appendix A.6 proves only that for three constructed extreme families H ≥ 1/sqrt(2n), so any θ ≤ ~0.15 would pass for max length 21; the choice of exactly 0.1 is not uniquely justified.
  • Popularity level cutoffs (low/medium/high) = not reported
    §4.2 defines levels 'based on the visit frequency' but the cutoff values are not given; they determine the popularity distributions and hence every popularity-disruption label.
  • Spatial distance level cutoffs (low/medium/high) = not reported
    §4.3 derives cutoffs from the statistical distribution of Haversine distances across all itineraries; exact quantiles not stated; they determine all spatial-disruption labels.
  • Florence sliding-window truncation size = 21
    Appendix A.1: long Florence itineraries truncated to window 21 to match Melbourne/Toronto max length; changes the composition and difficulty of the Florence sub-dataset.
  • Target intent count |z| sampling = uniform over {1,2,3}
    §5.1: a design choice that shapes the distribution of the number of intents in the dataset; not independently justified.
  • RAG retrieved-example count K = 3
    §6.1: experimental setting; minor influence on the RAG comparisons rather than the dataset itself.
axioms (6)
  • domain assumption Popularity, spatial distance, and category diversity are the key travel-preference factors and suffice as modification intents.
    Invoked in §1 and §3 with citation [8]; it defines the intent set Z. If real modification intents are more varied (e.g., temporal, budget, companionship), the dataset systematically misses them; the paper itself defers temporal constraints to future work.
  • ad hoc to paper ADD/REPLACE/DELETE exhaust the operations needed to model realistic modification requests.
    Footnoted in §1 ('leave other operations (e.g., shuffle) ... as future work'); the dataset's coverage of realistic fixes is bounded by this scope choice.
  • domain assumption Real-world Flickr-derived itineraries are reliable ground truth that implicitly encodes user intents.
    Base data from Toronto [25], Melbourne [45], Florence [30] (YFCC100M-derived); Appendix A.1 itself notes Florence's itinerary definition is inconsistent, so source quality is uneven.
  • domain assumption Tool-computed statistics are ground truth, and the LLM's verbatim mirroring of tool outputs yields correct perturbation labels.
    V3.2 FM's design (§5.2, Boxes 4–6) replaces model estimation with tool values; dataset quality therefore rests on the correctness of the tool implementations and prompt compliance, not on an independent human check (the manual filter in §5.3 is unquantified).
  • ad hoc to paper A perturbation that triggers the metric conditions (H > 0.1 or τ_b < 1 or CD change) corresponds to a genuine need-to-modify scenario.
    §4 defines disruption exactly this way; Appendix A.6 justifies the threshold only for three constructed families, so the equivalence between metric-firing and real modification need is an adopted postulate.
  • standard math Haversine distance with Earth radius ≈ 6371 km models geographic proximity adequately.
    §4.3 and prompt boxes; standard and uncontroversial for city-scale data.

pith-pipeline@v1.3.0-alltime-deepseek · 52140 in / 21616 out tokens · 196252 ms · 2026-08-03T10:14:24.813595+00:00 · methodology

0 comments
read the original abstract

Addressing itinerary modification is crucial for enhancing the travel experience as it is a frequent requirement during traveling. However, existing research mainly focuses on fixed itinerary planning, leaving modification underexplored due to the scarcity of need-to-modify itinerary data. To bridge this gap, we formally define the itinerary modification task and propose a general pipeline to construct the corresponding dataset, namely iTIMO. This pipeline frames the generation of need-to-modify itinerary data as an intent-driven perturbation task. It instructs large language models to perturb real-world itineraries using three operations: REPLACE, ADD, and DELETE. Each perturbation is grounded in three intents: disruptions of popularity, spatial distance, and category diversity. Furthermore, hybrid evaluation metrics are introduced to ensure perturbation effectiveness. We conduct comprehensive benchmarking on iTIMO to analyze the capabilities and limitations of state-of-the-art LLMs. Overall, iTIMO provides a comprehensive testbed for the modification task, and empowers the evolution of traditional travel recommender systems into adaptive frameworks capable of handling dynamic travel needs. Dataset, code and supplementary materials are available at https://github.com/zelo2/iTIMO.

Figures

Figures reproduced from arXiv: 2601.10609 by Hongyu Zhang, Hua Ma, Yunshan Ma, Zhuoxuan Huang, Zhu Sun.

Figure 1
Figure 1. Figure 1: Illustrations of itinerary modification scenarios across two [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall pipeline of iTIMO dataset construction. The top [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of perturbation positions for V3.2 and our [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. AlterAtlas: Shifting Travel Planning from AI Generation to Validation via Persona-Driven Simulations

    cs.HC 2026-07 conditional novelty 6.0

    AlterAtlas replaces one-shot AI itinerary generation with an interactive validation loop where persona-driven simulations expose route-level constraints and guide iterative revision.

Reference graph

Works this paper leans on

143 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauff- mann, et al. 2024. Phi-4 technical report.arXiv preprint arXiv:2412.08905(2024)

    Pith/arXiv arXiv 2024

  2. [2]

    Ashmi Banerjee, Adithi Satish, Fitri Nur Aisyah, Wolfgang Wörndl, and Yashar Deldjoo. 2025. SynthTRIPs: A knowledge-grounded framework for benchmark data generation for personalized tourism recommenders. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 3743–3752

    2025

  3. [3]

    Shihao Cai, Jizhi Zhang, Keqin Bao, Chongming Gao, Qifan Wang, Fuli Feng, and Xiangnan He. 2025. Agentic feedback loop modeling improves recommendation and user simulation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2235–2244

    2025

  4. [4]

    He Chang, Chenchen Ye, Zhulin Tao, Jie Wu, Zhengmao Yang, Yunshan Ma, Xianglin Huang, and Tat-Seng Chua. 2024. A comprehensive evaluation of large language models on temporal event forecasting.arXiv preprint arXiv:2407.11638 (2024)

    Pith/arXiv arXiv 2024

  5. [5]

    Soumyabrata Chaudhuri, Pranav Purkar, Ritwik Raghav, Shubhojit Mallick, Man- ish Gupta, Abhik Jana, and Shreya Ghosh. 2025. Tripcraft: A benchmark for spatio-temporally fine grained travel planning.arXiv preprint arXiv:2502.20508 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    Aili Chen, Xuyang Ge, Ziquan Fu, Yanghua Xiao, and Jiangjie Chen. 2024. Travelagent: An AI assistant for personalized travel planning.arXiv preprint arXiv:2409.08069(2024)

    Pith/arXiv arXiv 2024

  7. [7]

    Lei Chen, Jie Cao, Weichao Liang, and Qiaolin Ye. 2024. Geography-aware heterogeneous graph contrastive learning for travel recommendation.ACM Transactions on Spatial Algorithms and Systems10, 3 (2024), 1–22

    2024

  8. [8]

    Lei Chen, Jie Cao, Haicheng Tao, and Jia Wu. 2023. Trip reinforcement rec- ommendation with graph-based representation learning.ACM Transactions on Knowledge Discovery from Data17, 4 (2023), 1–20

    2023

  9. [9]

    Lei Chen, Guixiang Zhu, Weichao Liang, and Youquan Wang. 2023. Multi- objective reinforcement learning approach for trip recommendation.Expert Systems with Applications226 (2023), 120145

    2023

  10. [10]

    Wei Chen, Yuxuan Liang, Yuanshao Zhu, Yanchuan Chang, Kang Luo, Haomin Wen, Lei Li, Yanwei Yu, Qingsong Wen, Chao Chen, et al. 2024. Deep learning for trajectory data management and mining: A survey and beyond.arXiv preprint arXiv:2403.14151(2024)

    arXiv 2024

  11. [11]

    DeepSeek-AI Team. 2025. Introducing DeepSeek-V3.2-Exp. DeepSeek API Docs News. https://api-docs.deepseek.com/news/news250929

    2025

  12. [12]

    Bin Deng, Yizhe Feng, Zeming Liu, Qing Wei, Xiangrong Zhu, Shuai Chen, Yuanfang Guo, and Yunhong Wang. 2025. Retail: Towards real-world travel planning for large language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 14881–14913

    2025

  13. [13]

    Peter Emerson. 2013. The original Borda count and partial voting.Social Choice and Welfare40, 2 (2013), 353–358

    2013

  14. [14]

    Kenneth Enevoldsen, Isaac Chung, Imene Kerboua, Márton Kardos, Ashwin Mathur, David Stap, Jay Gala, Wissam Siblini, Dominik Krzemiński, Genta Indra Winata, et al. 2025. MMTEB: Massive multilingual text embedding benchmark. arXiv preprint arXiv:2502.13595(2025)

    arXiv 2025

  15. [15]

    Fuli Feng, Xiangnan He, Yiqun Liu, Liqiang Nie, and Tat-Seng Chua. 2018. Learn- ing on partial-order hypergraphs. InProceedings of the 2018 World Wide Web Conference. 1523–1532

    2018

  16. [16]

    Ines Gasmi, Makram Soui, Khaoula Barhoumi, and Mourad Abed. 2024. Rec- ommendation rules to personalize itineraries for tourists in an unfamiliar city. Applied Soft Computing150 (2024), 111084

    2024

  17. [17]

    Runquan Gui, Zhihai Wang, Jie Wang, Chi Ma, Huiling Zhen, Mingxuan Yuan, Jianye HAO, Defu Lian, Enhong Chen, and Feng Wu. 2025. HyperTree planning: Enhancing LLM reasoning via hierarchical thinking. InProceedings of the 42nd International Conference on Machine Learning

    2025

  18. [18]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    Pith/arXiv arXiv 2025

  19. [19]

    Huizhong Guo, Zhu Sun, Dongxia Wang, Tianjun Wei, Jinfeng Li, and Jie Zhang

  20. [20]

    Yilun Hao, Yongchao Chen, Yang Zhang, and Chuchu Fan. 2025. Large language models can solve real-world planning rigorously with formal verification tools. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 3434–3483

    2025

  21. [21]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. InProceedings of the International Conference on Learning Representations (ICLR). Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Huang et al

    2022

  22. [22]

    2000.Asymptotics in statistics: Some basic concepts

    Lucien Le Cam and Grace Lo Yang. 2000.Asymptotics in statistics: Some basic concepts. Springer Science & Business Media

    2000

  23. [23]

    2005.Testing statistical hypotheses

    Erich Leo Lehmann and Joseph P Romano. 2005.Testing statistical hypotheses. Springer

    2005

  24. [24]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al

  25. [25]

    Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera

  26. [26]

    Jianhua Lin. 2002. Divergence measures based on the Shannon entropy.IEEE Transactions on Information Theory37, 1 (2002), 145–151

    2002

  27. [27]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

    Pith/arXiv arXiv 2024

  28. [28]

    Xiaohao Liu, Jie Wu, Zhulin Tao, Yunshan Ma, Yinwei Wei, and Tat-seng Chua

  29. [29]

    Mistral AI. 2025. Mistral small 3. Mistral AI News. https://mistral.ai/news/mistral- small-3

    2025

  30. [30]

    Cristina Ioana Muntean, Franco Maria Nardini, Fabrizio Silvestri, and Ranieri Baraglia. 2015. On learning prediction models for tourists paths.ACM Transac- tions on Intelligent Systems and Technology (TIST)7, 1 (2015), 1–34

    2015

  31. [31]

    Hang Ni, Fan Liu, Xinyu Ma, Lixin Su, Shuaiqiang Wang, Dawei Yin, Hui Xiong, and Hao Liu. 2025. TP-RAG: Benchmarking retrieval-augmented large lan- guage model agents for spatiotemporal-aware travel planning.arXiv preprint arXiv:2504.08694(2025)

    arXiv 2025

  32. [32]

    In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Fine-tuning multimodal large language models for product bundling. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 848–858

  33. [33]

    OpenAI. 2024. text-embedding-3-large (Model documentation). https://platform .openai.com/docs/models/text-embedding-3-large. Accessed 2025-12-15

    2024

  34. [34]

    OpenAI. 2025. o4-mini Model (OpenAI API Documentation). https://platform.o penai.com/docs/models/o4-mini Accessed: 2026-01-08

    2025

  35. [35]

    Yincen Qu, Huan Xiao, Feng Li, Hui Zhou, and Xiangying Dai. 2025. TripScore: Benchmarking and rewarding real-world travel planning with fine-grained eval- uation.arXiv preprint arXiv:2510.09011(2025)

    arXiv 2025

  36. [36]

    OpenAI. 2024. GPT-4.1. https://platform.openai.com/docs/models#gpt-4.1. Large language model accessed via OpenAI API

    2024

  37. [37]

    Zijian Shao, Jiancan Wu, Weijian Chen, and Xiang Wang. 2025. Personal travel solver: A preference-driven LLM-solver system for travel planning. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 27622–27642

    2025

  38. [38]

    Roger W Sinnott. 1984. Virtues of the Haversine.Sky and telescope68, 2 (1984), 158

    1984

  39. [39]

    Zhu Sun, Kaidong Feng, Jie Yang, Xinghua Qu, Hui Fang, Yew-Soon Ong, and Wenyuan Liu. 2024. Adaptive in-context learning with large language models for bundle generation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 966–976

    2024

  40. [40]

    Jie-Jing Shao, Bo-Wen Zhang, Xiao-Wen Yang, Baizhi Chen, Si-Yu Han, Wen-Da Wei, Guohao Cai, Zhenhua Dong, Lan-Zhe Guo, and Yu-feng Li. 2024. China- Travel: An open-ended benchmark for language agents in Chinese travel planning. arXiv preprint arXiv:2412.13682(2024)

    Pith/arXiv arXiv 2024

  41. [41]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. 2025. Gemma 3 technical report.arXiv preprint arXiv:2503.19786 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    Llama Team. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  43. [43]

    Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. 2016. Yfcc100m: The new data in multimedia research.Commun. ACM59, 2 (2016), 64–73

    2016

  44. [44]

    Yihong Tang, Zhaokai Wang, Ao Qu, Yihao Yan, Zhaofeng Wu, Dingyi Zhuang, Jushi Kai, Kebing Hou, Xiaotong Guo, Jinhua Zhao, et al. 2024. ItiNera: Integrating spatial optimization with large language models for open-domain urban itinerary planning. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 1413–1432

    2024

  45. [45]

    Xiaoting Wang, Christopher Leckie, Jeffrey Chan, Kwan Hui Lim, and Tharshan Vaithianathan. 2016. Improving personalized trip recommendation by avoiding crowds. InProceedings of the 25th ACM international on Conference on Information and Knowledge Management. 25–34

    2016

  46. [46]

    Tianjun Wei, Huizhong Guo, Yingpeng Du, Zhu Sun, Chen Huang, Dongxia Wang, and Jie Zhang. 2025. Mirroring users: Towards building preference-aligned user simulator with user feedback in recommendation.arXiv preprint arXiv:2508.18142 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    Qianfeng Wen, Yifan Liu, Joshua Zhang, George Saad, Anton Korikov, Yury Sambale, and Scott Sanner. 2024. Elaborative subtopic query reformulation for broad and indirect queries in travel destination recommendation.arXiv preprint arXiv:2410.01598(2024)

    Pith/arXiv arXiv 2024

  48. [48]

    Kaimin Wang, Yuanzhe Shen, Changze Lv, Xiaoqing Zheng, and Xuan-Jing Huang

  49. [49]

    In Findings of the Association for Computational Linguistics: ACL 2025

    Triptailor: A real-world benchmark for personalized travel planning. In Findings of the Association for Computational Linguistics: ACL 2025. 9705–9723

    2025

  50. [50]

    Dongjie Yang, Chengqiang Lu, Qimeng Wang, Xinbei Ma, Yan Gao, Yao Hu, et al

  51. [51]

    Qing Yin, Hui Fang, Zhu Sun, and Yew-Soon Ong. 2023. Understanding diversity in session-based recommendation.ACM Transactions on Information Systems42, 1 (2023), 1–34

    2023

  52. [52]

    Yuguo Yuan and Weimin Zheng. 2024. Your trip, your way: An adaptive tourism recommendation system.Applied Soft Computing154 (2024), 111330

    2024

  53. [53]

    Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024. TravelPlanner: A benchmark for real-world planning with language agents. InProceedings of the 41st International Conference on Machine Learning. 54590–54613

    2024

  54. [54]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  55. [55]

    Zhihan Zhang, Yixin Cao, Chenchen Ye, Yunshan Ma, Lizi Liao, and Tat-Seng Chua. 2024. Analyzing temporal complex events with large language models? A benchmark towards temporal, long context understanding. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024): Bangkok, Thailand, August. 11–16

    2024

  56. [56]

    InProceddings of the 39th Conference on Neural Information Processing Systems

    Wide-horizon thinking and simulation-based evaluation for real-world LLM planning with multifaceted constraints. InProceddings of the 39th Conference on Neural Information Processing Systems

  57. [57]

    Tao Zhe, Rui Liu, Fateme Memar, Xiao Luo, Wei Fan, Xinyue Ye, Zhongren Peng, and Dongjie Wang. 2025. Constraint-aware route recommendation from natural language via hierarchical LLM agents.arXiv preprint arXiv:2510.06078(2025). Acknowledgments This work is partially supported by National Natural Science Foun- dation of China under Grants 71971221 and 6247...

    arXiv 2025

  58. [59]

    Hongyu Zhang, Zhuoxuan Huang, Zixu Jiang, Hua Ma, and Haibin Zhu. 2024. Collaborative route planning of road trips in regional central cities of China: An approach based on E-CARGO model.IEEE Transactions on Computational Social Systems11, 5 (2024), 6347–6365

    2024

  59. [60]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176(2025)

    Pith/arXiv arXiv 2025

  60. [62]

    Xinping Zhao, Xinshuo Hu, Zifei Shan, Shouzheng Huang, Yao Zhou, Xin Zhang, Zetian Sun, Zhenyu Liu, Dongfang Li, Xinyuan Wei, et al. 2025. Kalm-embedding- v2: Superior training techniques and data inspire a versatile embedding model. arXiv preprint arXiv:2506.20923(2025)

    arXiv 2025

  61. [64]

    Furthermore, the prompts used for itinerary modification are shown as box 8-10

    The template of memory module is shown as box 7. Furthermore, the prompts used for itinerary modification are shown as box 8-10. Table 9: Tool box of itinerary perturbation for our V3.2 FM. Function Purpose geo_distance_segments Computes consecutive spatial distances between adjacent POIs and discretizes each segment into 𝑙𝑜𝑤/𝑚𝑒𝑑𝑖𝑢𝑚/ℎ𝑖𝑔ℎ classes for origi...

    arXiv 2018

  62. [66]

    This addition must disrupt the itinerary in terms of at least one of the following aspects: - Spatial distance consistency - Popularity consistency - Category diversity consistency

    Perturbation operation (ADD) Insert exactly ONE new POI from the candidate POIs into the original itinerary at any valid position. This addition must disrupt the itinerary in terms of at least one of the following aspects: - Spatial distance consistency - Popularity consistency - Category diversity consistency

  63. [67]

    categories_raw_before

    Definitions of Consistency and Disruption (3.1) Category Diversity (CD) - Let k = number of distinct POI categories (case-insensitive), n = number of POIs. - Rule: If k == 1, then CD = 0. Else, CD = k / n. - For ADD: categories_before = categories of the original itinerary categories_after = categories_before plus the inserted POI's category CD disruption...

  64. [68]

    Target Intent Count

    Comparison and Intent Matching Input includes a "Target Intent Count": 1|2|3. - Always try to generate a perturbation that yields exactly this number of valid intents. - If it is possible, output that exact number. - If it is not possible, output the actual number of valid intents instead. - If no valid disruption exists, output "UNKNOWN". INTENT SELECTIO...

    2018

  65. [70]

    Perturbed Itinerary

    Output format You MUST return the result in strict JSON format. { "Perturbed Itinerary": [...], "Intents": [...], -- Category Debug -- "categories_raw_before": [...], "categories_set_before": [...], "categories_raw_after": [...], "categories_set_after": [...], "cd_before": <float>, "cd_after": <float>, "cd_disruption": true|false, -- Popularity Debug -- "...

    2018

  66. [72]

    This deletion must disrupt the itinerary in terms of at least one of the following aspects: - Spatial distance consistency - Popularity consistency - Category diversity consistency

    Perturbation operation (DELETE) Remove exactly ONE POI from the original itinerary. This deletion must disrupt the itinerary in terms of at least one of the following aspects: - Spatial distance consistency - Popularity consistency - Category diversity consistency

  67. [73]

    categories_raw_before

    Definitions of Consistency and Disruption (3.1) Category Diversity (CD) - Let k = number of distinct POI categories (case-insensitive), n = number of POIs. - Rule: If k == 1, then CD = 0. Else, CD = k / n. - For DELETE: categories_before = categories of the original itinerary categories_after = categories_before minus the deleted POI's category (if no oth...

  68. [74]

    UNKNOWN". INTENT SELECTION RULE (STRICT): - If cd_disruption = true -> include

    Comparison and Intent Matching Compare the original and perturbed itineraries. From the candidate intents, select only those that are truly supported by the above rules. If none apply, output "UNKNOWN". INTENT SELECTION RULE (STRICT): - If cd_disruption = true -> include "Category diversity disruption" - If popularity_disruption = true -> include "Popular...

  69. [75]

    Validate all numeric values: cd_before and cd_after must be within [0.0, 1.0]

    Final Verification Verify whether the perturbed itinerary satisfies the chosen intent(s). Validate all numeric values: cd_before and cd_after must be within [0.0, 1.0]. popularity_H and spatial_H must be within [0.0, 1.0]. iTIMO: An LLM-empowered Synthesis Dataset for Travel Itinerary Modification Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Bo...

    2018

  70. [76]

    Perturbed Itinerary

    Output format You MUST return the result in strict JSON format. { "Perturbed Itinerary": [...], "Intents": [...], -- Category Debug -- "categories_raw_before": [...], "categories_set_before": [...], "categories_raw_after": [...], "categories_set_after": [...], "cd_before": <float>, "cd_after": <float>, "cd_disruption": true|false, -- Popularity Debug -- "...

    2018

  71. [77]

    Definition of an itinerary An itinerary is a sequence of visiting activities, where each activity is represented as: [POI name, POI category, longitude, latitude, popularity level]

  72. [78]

    Perturbation operation (REPLACE) Replace exactly ONE POI in the original itinerary with ONE new POI from the candidate POIs. This replacement must disrupt the itinerary in terms of at least one of the following aspects: - Spatial distance consistency - Popularity consistency - Category diversity consistency

  73. [79]

    categories_raw_before

    Definitions of Consistency and Disruption (3.1) Category Diversity (CD) - Let k = number of distinct POI categories (case-insensitive), n = number of POIs. - Rule: If k == 1, then CD = 0. Else, CD = k / n. - For REPLACE: categories_before = categories of the original itinerary categories_after = categories_before but with replaced POI's category substitut...

  74. [80]

    Target Intent Count

    Comparison and Intent Matching Input includes a "Target Intent Count": 1|2|3. - Always try to generate a perturbation that yields exactly this number of valid intents. - If it is possible, output that exact number. - If it is not possible, output the actual number of valid intents instead. - If no valid disruption exists, output "UNKNOWN". INTENT SELECTIO...

    2018

  75. [81]

    Validate all numeric values: cd_before and cd_after must be within [0.0, 1.0]

    Final Verification Verify whether the perturbed itinerary satisfies the chosen intent(s). Validate all numeric values: cd_before and cd_after must be within [0.0, 1.0]. popularity_H and spatial_H must be within [0.0, 1.0]. popularity_tau_b and spatial_tau_b must be within [-1.0, 1.0] if numeric. If any value falls outside these ranges: Recompute CD, Helli...

  76. [82]

    Perturbed Itinerary

    Output format You MUST return the result in strict JSON format. { "Perturbed Itinerary": [...], "Intents": [...], -- Category Debug -- "categories_raw_before": [...], "categories_set_before": [...], "categories_raw_after": [...], "categories_set_after": [...], "cd_before": <float>, "cd_after": <float>, "cd_disruption": true|false, -- Popularity Debug -- "...

    2018

  77. [83]

    error":...} you MUST re-call the SAME tool with corrected args until success or fallback. - Case discipline / domains: * Popularity labels MUST be Title Case

    Tool-Calling Contract (HARD RULES - zero-omission) - Zero omission: supply ALL required keys exactly; no renaming; no missing args. - Retry on error: on {"error":...} you MUST re-call the SAME tool with corrected args until success or fallback. - Case discipline / domains: * Popularity labels MUST be Title Case "High","Medium","Low" (domain=["High","Mediu...

  78. [84]

    Prefer inserting a POI whose name is NOT in used_poi (exact string match)

  79. [85]

    Prefer an insertion index i NOT in used_index; if unavoidable, choose among the least-used indices

  80. [86]

    If multiple valid insertions meet the Target Intent Count, select one differing from Memory in BOTH POI and index; if still tied, sample uniformly at random

Showing first 80 references.