Pith. sign in

REVIEW 2 major objections 5 minor 124 references

Even the best LLM agent fully solves only 46% of jointly feasible trip-planning tasks; unstated traveler needs are the universal bottleneck.

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

2026-07-30 15:11 UTC pith:WGQIRQ4X

load-bearing objection Solid, reproducible joint-feasibility travel-agent benchmark; the 46% ceiling and D1 bottleneck are real inside the sandbox, but “unstated needs” is oversold relative to what D1 actually measures. the 2 major comments →

arxiv 2607.26977 v1 pith:WGQIRQ4X submitted 2026-07-29 cs.CL

TREK: A Travel Reasoning and Evaluation Kit for LLM Agents in Complex Trip Planning

Jinhu Qi , Wentao Zhang , Siu Man Ng , Feiyang Xu , Yanyu Chen , Yaoman Li , Irwin King This is my paper
classification cs.CL
keywords Travel PlanningTool-useImplicit-needRESTful APILLM AgentFeasible itinerary synthesisDeterministic evaluationTyped infeasibility
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.

A usable trip plan is not a loose suggestion: every booked flight, hotel, and attraction must exist, the day must be physically traversable, the total must fit the budget, and the plan must serve needs the traveler only partly states. TREK is built to measure that joint requirement as one artifact. It gives agents a production-style API sandbox over a large synthetic knowledge base, 800 multi-constraint tasks (including typed impossible ones), a fully deterministic scorer with no LLM judge, and a human-verified gold plan that scores a perfect 1.0—so any shortfall is the agent’s, not the scorer’s. Across 15 agents the strongest fully-feasible rate on solvable tasks is 46.2%, the median is 6.6%, and the floor is 0%. Satisfying unstated persona needs is the dimension almost no model clears at scale, even when the rest of the plan is largely right.

Core claim

On TREK’s feasible tasks, even the strongest evaluated agent produces a plan that is jointly constraint-correct, hallucination-free, spatio-temporally executable, budget-valid, and responsive to unstated persona needs on only 46.2% of cases (median 6.6%, floor 0.0% across 15 agents). Implicit-need satisfaction is the universal bottleneck: it remains the last wall for the frontier model and a top failure for every agent tested.

What carries the argument

Feasible itinerary synthesis, scored by the task-perfect rate: a single plan must pass every applicable one of nine deterministic dimensions (and hard gates) at once. The evaluator is rule-based with no LLM judge; gold references demonstrably score 1.0; 267 tasks are typed-infeasible (route/entity/budget) so correct refusal is first-class.

Load-bearing premise

That success on a synthetic, internally consistent travel sandbox—with unstated needs reduced to facility checklist matches against stated persona labels—certifies the real deployable skill of building executable, traveler-responsive itineraries.

What would settle it

If a frontier tool-using agent, under the same TREK harness and scorer, reaches near-100% task-perfect on the 533 feasible tasks—including implicit-need cells—and human raters judge those plans as realistic and followable at rates matching the gold, the claimed capability gap collapses.

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

If this is right

  • Joint feasibility, not single-axis tool success, becomes the right unit for certifying travel (and similar multi-constraint) agents.
  • Implicit persona needs must be treated as a first-class, hard requirement rather than soft preference flavor text.
  • Typed refusal of impossible requests (route/entity/budget) is a scored skill, not an afterthought.
  • Deterministic no-judge scorers paired with achievable gold make remaining gaps attributable to agents, not rubrics.
  • More deliberation or token spend does not automatically buy more joint feasibility under strict tool schemas.

Where Pith is reading between the lines

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

  • The same joint-feasibility template—synthetic consistent world, typed impossibility, deterministic conjunction scoring—could stress-test agents in logistics, clinical scheduling, or multi-leg procurement.
  • If unstated needs stay the rising bottleneck as models improve elsewhere, progress may require explicit preference-to-resource mapping modules rather than longer chain-of-thought alone.
  • A hidden-persona split (infer the traveler from indirect language, score the same facility checklist) would test whether the current D1 gap understates the real personalization problem.

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

2 major / 5 minor

Summary. TREK is a benchmark for joint feasible itinerary synthesis by tool-using LLM agents. It provides 800 multi-constraint travel tasks (533 feasible, 267 typed-infeasible) over a synthetic, internally consistent KB of 212,530 records across 375 cities and 13 personas, accessed via validated RESTful APIs. Scoring is fully deterministic across nine dimensions with no LLM judge; each task ships a human-verified gold that scores 1.0 under the same evaluator. Evaluating 15 agents, the authors report a top task-perfect rate of 46.2% on solvable tasks (median 6.6%, floor 0.0%), with implicit-need satisfaction (D1) as the universal bottleneck and B3 as a planner/non-planner watershed; reasoning and extra tokens do not clearly help. The release includes dataset, sandbox, evaluator, and agent code.

Significance. If the measurements hold under the stated operationalization, TREK is a meaningful advance over TravelPlanner, ChinaTravel, and concurrent travel benchmarks by coupling (i) a purely rule-based joint-feasibility scorer, (ii) a gold proven to hit the ceiling on all 800 tasks, (iii) typed route/entity/budget infeasibility as scored behavior, and (iv) a production-style tool sandbox with an efficiency axis. The 15-model study is carefully instrumented (per-dimension failure tables, aggregation robustness, truncation diagnostics, shared B3 travel-time helper). These are the right ingredients for a reproducible agent benchmark: bit-reproducible scores, an attainable ceiling, and falsifiable per-dimension failure modes. The work is significant for tool-agent evaluation even if the strongest interpretive claim about “unstated needs” needs tightening.

major comments (2)
  1. [Abstract; §1; §4.2 D1; §5.3; App. E.3] The headline scientific claim that “satisfying travelers’ unstated persona needs” is the universal bottleneck overreaches what D1 measures. Queries name persona keywords explicitly (e.g., “for elderly travelers”); the system prompt (§D.4) further instructs agents that these labels are requirements, to derive amenities, and to use amenity/facility filters; D1 then scores deterministic facility set-intersection (plus fixed star/rating tests) against KB fields (§4.2, App. D.3, A.10). Appendix E.3 correctly concedes D1 does not test persona inference from indirect language. The residual ~53.7% D1 failure for GPT-5.6 is still important—it is multi-cell stated-persona→facility booking under joint constraints—but abstract, intro, RQ2, and conclusion should rename and reframe D1 (e.g., persona-conditioned facility grounding) so the bottleneck claim matches the instrument. Without that edit, the
  2. [§5.1; §5.3; App. D.4 system prompt] Relatedly, the agent harness scaffolds the very capability D1 is said to isolate: §D.4 tells models which traveler types imply concrete booking needs and that plans are scored on matching amenities. That makes the frontier D1 failure more striking as a grounding/execution gap, but it weakens any claim that agents fail at discovering partly spoken needs. Please either (a) report an ablation with a minimal prompt that does not enumerate the persona→facility mapping recipe, or (b) explicitly scope the finding as failure under a scaffold that already reveals the mapping problem. Option (b) is acceptable if the framing in major comment 1 is fixed; option (a) would strengthen the paper.
minor comments (5)
  1. [§2.1; Table 1] Table 1 and §2 positioning vs. ChinaTravel/TravelBench/TravelEval is clear and useful; consider adding one sentence on whether any concurrent benchmark has since added typed infeasibility or achievable gold, to keep the “to our knowledge” claim durable.
  2. [§4.2; App. D.3] Free scoring parameters (D3 β=4, B2 60-minute dwell cap, B3 surface/air constants, efficiency γ and token budget B) are documented but not sensitivity-tested beyond the task-perfect 0.95 check. A short appendix sweep on β and the B2 cap would reassure readers that rankings are not knife-edge.
  3. [Figure 4; Table 15] Figure 4 heatmap and Table 15 are excellent; ensure color/print accessibility and that “top-two failure” is defined in the caption (rank by failure rate among applicable dimensions).
  4. [§6; Appendix F] Appendix F already handles synthetic-world, single-run, and Claude geo-block limits well. Move a one-sentence external-validity caveat into the main conclusion so readers who skip the appendix see the scope bound next to the 46.2% headline.
  5. [Abstract; §3.2] Minor consistency: abstract says needs are “only partly stated” while templates inject explicit persona keywords—align wording with the revised D1 framing.

Circularity Check

0 steps flagged

No significant circularity: TREK is a transparent benchmark whose gold/labels are perfect by construction by design, while the headline agent scores are independent empirical measurements.

full rationale

TREK does not claim a first-principles derivation or a fitted natural-law prediction. Its load-bearing empirical claim—that frontier agents reach only ~46.2% task-perfect on solvable tasks, with D1 as the residual bottleneck—is an external measurement of third-party models under a fixed harness and a deterministic scorer. Feasibility labels and gold itineraries scoring 1.0 are openly engineered to be correct/perfect against that same scorer (§3.3–3.5, §4.1), which is standard achievable-ceiling benchmark design, not a hidden reduction of a predicted quantity to its fit. Budget bands, typed infeasibility, B3’s shared travel-time model, and D1’s facility set-intersection are author-specified world rules disclosed in the paper; agents are scored against them, not used to redefine them. There is no self-citation uniqueness theorem, no parameter fit renamed as prediction, and no ansatz smuggled in via overlapping-author prior work that forces the result. Framing tension around calling D1 “unstated” needs (personas are named in queries) is a construct-validity issue, not circularity of the derivation chain. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

TREK’s claims rest on a constructed evaluation world and scoring conventions rather than free physical constants. Load-bearing choices include synthetic internal consistency as sufficient for certification, author-compiled persona→facility maps as the meaning of ‘unstated needs,’ fixed travel-time and cost models shared by tools and scorer, budget-band generators, and conjunctive/geometric aggregation into task-perfect rate. These define the benchmark; agent performance is then measured inside them.

free parameters (6)
  • Budget slack bands (tight u∈[1.02,1.10], loose u∈[1.35,1.80]; infeasible u∈[0.55,0.92]) = means/ranges reported; near-even 267 tight / 266 loose
    Hand-chosen multipliers relative to C_min/C_floor set how hard budget adherence is and create the 89 budget-infeasible tasks.
  • D3 overspend penalty β=4 = β=4
    Exponential budget score shape is a design choice that maps 25%/50% overspend to 0.37/0.14.
  • B2 dwell-time cap of 60 minutes = 60 minutes
    Required visit length is min(KB duration, 60m); chosen so gold is not failed on 26% of timed visits.
  • B3 travel-time model constants = R=6371 km; surface/air constants as in Eq. (5)
    Surface 60 km/h + 15 min buffer with 45 min floor; air 700 km/h + 180 min overhead—canonical model shared with the helper tool.
  • Efficiency penalties (γ=10 per surplus call, C_max=15, token budget B) = γ=150/C_max=10; cap 15 billable calls
    Tool-call and token overrun scoring is parameterized; held out of the correctness headline but reported.
  • Geometric-mean floor ε for category product = small ε (unspecified numeric)
    Small floor keeps Overall defined when a category is zero; aggregation choice affects composite though TP-feas is primary.
axioms (6)
  • domain assumption A synthetic, internally consistent KB is a stronger basis for certifying executability than a scraped drifting corpus.
    Stated as deliberate design in §3.1; underpins exact ground truth and achievable gold.
  • ad hoc to paper Unstated persona needs are adequately operationalized as deterministic nonempty intersection of fixed facility sets (plus fixed star/rating tests for luxury/foodie) on booked KB resources.
    D1 definition §4.2 and Appendix D.3/E; excludes hidden-persona inference.
  • ad hoc to paper Only logically impossible persona combinations (party-size contradictions; fast-paced budget vs luxury) are excluded; other co-occurrence is jointly satisfiable.
    §3.2 and Appendix B.2 after retiring a 19-pair stereotype-prone conflict list.
  • domain assumption A deployable plan requires simultaneous satisfaction of constraint, truthfulness, executability, and correct refusal—aggregated conjunctively (task-perfect) and via geometric mean of categories.
    §4.3–4.4; rejects compensatory arithmetic averaging for the headline.
  • domain assumption Identifier-level KB match (not mutable price/rating equality) is the right hallucination test; attribute fidelity is deferred to budget/other checks.
    D0-src specification in Appendix D.3.
  • domain assumption Standard tool-agent evaluation practices (temperature 0, fixed harness, provider token counts as cost proxy) suffice for comparative claims in RQ1–RQ3.
    §5.1 and Appendix F limitations on single-run and cost axis.
invented entities (4)
  • TREK task suite + synthetic multi-domain travel KB (212,530 records, 375 cities) no independent evidence
    purpose: Provide controlled ground truth for joint itinerary feasibility and typed infeasibility.
    Core benchmark artifact; not a physical entity but a new evaluation world.
  • Thirteen traveler personas with fixed facility/amenity/service requirement tables independent evidence
    purpose: Make implicit-need satisfaction deterministically scorable (D1).
    Compiled from cited industry sources in Appendix E, but the exact mapping is a TREK construct.
  • Nine-dimension deterministic evaluator and task-perfect metric with hard gates no independent evidence
    purpose: Certify joint feasibility without an LLM judge and with an achievable ceiling.
    Scoring system invented for the paper; gold panel provides partial external realism check.
  • Typed infeasibility labels (entity/route/budget) as first-class scored behavior (D4) no independent evidence
    purpose: Reward correctly diagnosed refusal rather than untyped failure.
    Evaluation construct beyond broad capability-boundary tags in related work.

pith-pipeline@v1.2.0-daily-grok45 · 50948 in / 4321 out tokens · 97601 ms · 2026-07-30T15:11:13.094751+00:00 · methodology

0 comments
read the original abstract

Travel planning is a demanding stress test for tool-using LLM agents: a usable itinerary is a single artifact that must be right along many axes at once - every flight, hotel, and attraction must exist and be bookable, the days must be physically traversable, the total must clear a budget, and the plan must serve a traveler whose needs are only partly stated. Existing agent benchmarks reward these properties one at a time and grade the final output with soft or LLM-judged rubrics, which cannot certify that a returned plan is executable and are neither reproducible nor auditable. We introduce TREK (Travel Reasoning and Evaluation Kit), a benchmark for feasible itinerary synthesis: producing a single plan that is jointly constraint-correct, hallucination-free, spatio-temporally executable, budget-valid, and responsive to the traveler's unstated persona needs. TREK comprises 800 multi-constraint tasks - 533 feasible and 267 provably infeasible with typed route/entity/budget causes - over a synthetic, internally consistent knowledge base of 212,530 records across 375 cities and 13 personas, served through a production-style tool sandbox of validated RESTful APIs. Every task is scored by a fully deterministic, rule-based evaluator with no LLM judge and ships a human-verified gold reference that scores a perfect 1.0 under that same evaluator, so the ceiling is demonstrably achievable and every remaining gap is an agent limitation rather than scorer strictness. Evaluating 15 LLM agents across nine constraint dimensions, we find that even the strongest (GPT-5.6) produces a fully-feasible plan on only 46.2% of solvable tasks, with a median of 6.6% and a floor of 0.0%; satisfying travelers' unstated needs emerges as the universal bottleneck, unsolved even at the frontier. We release the dataset, tool sandbox, deterministic evaluator, and agent code as a fully reproducible benchmark.

Figures

Figures reproduced from arXiv: 2607.26977 by Feiyang Xu, Irwin King, Jinhu Qi, Siu Man Ng, Wentao Zhang, Yanyu Chen, Yaoman Li.

Figure 1
Figure 1. Figure 1: Examples of feasible and infeasible queries in TREK. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TREK. A deterministic pipeline (left) builds a synthetic, internally consistent knowledge base, populates [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-model use of the free compute_travel_time helper: the fraction of the 800 tasks on which the agent queried the exact door-to-door travel-time model that B3 scores against. All 15 models used it; the mean is 54.6% (dashed). Usage did not guarantee feasibility—Qwen3-Next queried it most yet fails B3 most—so B3 failures are a sched￾uling gap, not an unqueryable rule. complementary: D0-key captures fine-gr… view at source ↗
Figure 4
Figure 4. Figure 4: Failure decomposition: per-dimension failure rate (% of applicable tasks scoring [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cost vs. quality across the [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

124 extracted references · 2 canonical work pages

  1. [1]

    AAA. 2017. Foodie Travelers Are Embracing the Culinary Travel Trend. https://newsroom.aaa.com/2017/04/foodie-travelers-embracing-culinary- travel-trend/. AAA Newsroom travel release; Accessed 2026-02-09

    2017

  2. [2]

    AAA Foundation for Traffic Safety. 2020. Older Drivers and Advanced Driver Assistance Systems. https://aaafoundation.org/older-drivers-and-advanced- driver-assistance-systems/. Research brief; Accessed 2026-02-09

    2020

  3. [3]

    Amazon AGI. 2025. The Amazon Nova Family of Models: Technical Report and Model Card. arXiv:2506.12103 [cs.CL]

    Pith/arXiv arXiv 2025

  4. [4]

    Amazon Web Services. 2025. Introducing Amazon Nova 2 Lite, a Fast, Cost- Effective Reasoning Model. https://aws.amazon.com/blogs/aws/introducing- amazon-nova-2-lite-a-fast-cost-effective-reasoning-model/. AWS News Blog release announcement

    2025

  5. [5]

    Hannah Aston. 2024. Travel Trends: Luxury Culinary Tourism. https://discover.hotelbeds.com/resources/insight/2024-travel-trends-luxury- culinary-tourism. Accessed 2026-02-09

    2024

  6. [6]

    Avis Rent A Car. n.d.. Child Safety Seats — Avis Car Rental. https://www.avis. com/en/products-and-services/products/childsafetyseats. Accessed 2026-02-09

    2026

  7. [7]

    Srinivas Billa and Xiaonan Jing. 2025. TravelBench: Exploring LLM Perfor- mance in Low-Resource Domains.arXiv preprint arXiv:2510.02719(2025). arXiv:2510.02719 https://arxiv.org/abs/2510.02719

    arXiv 2025

  8. [8]

    Jiangxi Network Broadcasting and Television Station. 2020. Publishing the 2020 Post-COVID Self-Drive Tourism Report: Consumer Trends Analysis. https://cn. chinadaily.com.cn/a/202010/09/WS5f8020fca3101e7ce972844c.html. Published on China Daily Chinese edition; Accessed 2026-02-09

    2020

  9. [9]

    Soumyabrata Chaudhuri, Pranav Purkar, Ritwik Raghav, Shubhojit Mallick, Manish Gupta, Abhik Jana, and Shreya Ghosh. 2025. TripCraft: A Benchmark for Spatio-Temporally Fine Grained Travel Planning. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics. ar...

    Pith/arXiv arXiv 2025

  10. [10]

    Junle Chen, Wei Chen, Yehong Xu, Zhengjun Huang, Yuqian Wu, Zhoujin Tian, Kai Wang, Lei Wang, and Xiaofang Zhou. 2026. Trip+: Benchmarking Agents in Personalized Interactive Travel Planning.arXiv preprint arXiv:2606.21169(2026). arXiv:2606.21169 https://arxiv.org/abs/2606.21169

    Pith/arXiv arXiv 2026

  11. [11]

    Weiyi Chen, Shuaixiong Wang, Ziyun Gao, Kaichun Hu, Wangze Ni, Shimin Di, Chen Jason Zhang, and Lei Chen. 2026. TravelEval: A Comprehensive Bench- marking Framework for Evaluating LLM-Powered Travel Planning Agents. In Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM. arXiv:2606.01046 doi:10.1145/3770855.3817533

    Pith/arXiv arXiv 2026

  12. [12]

    Xiang Cheng, Yulan Hu, Xiangwen Zhang, Lu Xu, Lide Tan, Zheng Pan, Xin Li, and Yong Liu. 2026. Beyond Itinerary Planning—A Real-World Benchmark for Multi-Turn and Tool-Using Travel Tasks. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 29200–29251...

    Pith/arXiv arXiv 2026

  13. [13]

    Xiang Cheng, Yulan Hu, Lulu Zheng, Zheng Pan, Xin Li, and Yong Liu. 2026. GroupTravelBench: Benchmarking LLM Agents on Multi-Person Travel Plan- ning.arXiv preprint arXiv:2605.25200(2026). arXiv:2605.25200 https://arxiv.org/ abs/2605.25200

    Pith/arXiv arXiv 2026

  14. [14]

    Comcast Business Community Editorial Team. [n. d.]. The “Why” Behind WiFi: How Deploying WiFi is Good for Business. https: //business.comcast.com/community/browse-all/details/the-why-behind- wifi-how-deploying-wifi-is-good-for-business. Accessed 2026-02-09

    2026

  15. [15]

    Condor Ferries. 2025. Family Travel Statistics 2025. https://www.condorferries. co.uk/family-travel-statistics. Accessed 2026-02-09

    2025

  16. [16]

    Condor Ferries. 2025. Pet Travel Statistics 2025. https://www.condorferries.co. uk/pet-travel-statistics. Accessed 2026-02-09

    2025

  17. [17]

    Condor Ferries. 2025. Solo Travel Statistics 2025. https://www.condorferries.co. uk/solo-travel-statistics. Accessed 2026-02-09

    2025

  18. [18]

    Denise Curtin. 2018. THIS hotel is now offering the world’s first ever ‘Instagram Butler’. https://her.ie/life/hotel-now-offering-worlds-first-ever-instagram- butler-370134. Her.ie lifestyle article; Accessed 2026-02-09

    2018

  19. [19]

    CyberPublicity. 2025. Nightlife Enthusiasts. https://www.cyberpublicity.com/ programmatic-advertising/hobbies-passions/nightlife-enthusiasts/. Accessed 2026-02-09

    2025

  20. [20]

    Renfei Dang, Zhening Li, Shujian Huang, and Jiajun Chen. 2025. The First Impression Problem: Internal Bias Triggers Overthinking in Reasoning Models. arXiv:2505.16448 [cs.AI] https://arxiv.org/abs/2505.16448

    arXiv 2025

  21. [21]

    Tomas de la Rosa, Sriram Gopalakrishnan, Alberto Pozanco, Zhen Zeng, and Daniel Borrajo. 2024. TRIP-PAL: Travel Planning with Guarantees by Com- bining Large Language Models and Automated Planners.arXiv preprint arXiv:2406.10196(2024). arXiv:2406.10196 https://arxiv.org/abs/2406.10196

    Pith/arXiv arXiv 2024

  22. [22]

    DeepSeek-AI. 2025. DeepSeek-V3.2. https://huggingface.co/deepseek-ai/ DeepSeek-V3.2. Model card

    2025

  23. [23]

    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. Association for Computa- tional Linguistics. arXiv:2508.15335 doi:10.18653/v1/2025.emnlp...

    Pith/arXiv arXiv 2025

  24. [24]

    Ben Duhig. 2022. McKinsey reveals 7 emerging traveller archetypes. https://www.linkedin.com/posts/benduhig_7-emerging-traveller-archetypes- all-travel-activity-7211269404188176384-JYBG. LinkedIn post. Accessed 2026-02-09

    2022

  25. [25]

    Enterprise Rent-A-Car. n.d.. What is the Enterprise Pet Policy? https://www.enterprise.com/en/car-rental-faqs/us-general/car-rental- pet-friendly-policy.html. Accessed 2026-02-09

    2026

  26. [26]

    Expedia Group. 2022. New research: 97% of honeymoon plans were thwarted by the pandemic, leading to the rise of the ’mega-moon’. https://www.expediagroup.com/media/media-details/2022/New-research-97- of-honeymoon-plans-were-thwarted-by-the-pandemic-leading-to-the-rise- of-the-mega-moon/default.aspx. Accessed 2026-02-09

    2022

  27. [27]

    Food Inspiration Magazine. 2023. Trendwatch – Food tourism. https://www. foodinspirationmagazine.com/39-food-tourism/trendwatch-food-tourism. Ac- cessed 2026-02-09

    2023

  28. [28]

    Four Seasons Hotels and Resorts. 2024. Hotels With Michelin Star Restaurants. https://www.fourseasons.com/magazine/taste/michelin-starred- restaurants/. Accessed 2026-02-09

    2024

  29. [29]

    Four Seasons Hotels and Resorts. 2024. A Meal to Remember: Luxury Dining with Four Seasons. https://www.fourseasons.com/magazine/taste/michelin- starred-restaurants/. Accessed 2026-02-09

    2024

  30. [30]

    GLM-5 Team, Z.ai. 2026. GLM-5: From Vibe Coding to Agentic Engineering. arXiv:2602.15763 [cs.CL]

    Pith/arXiv arXiv 2026

  31. [31]

    Google DeepMind. 2026. Gemma 4: Byte for Byte, the Most Capable Open Models. https://blog.google/innovation-and-ai/technology/developers-tools/ gemma-4/. Release announcement; Gemma 4 31B Dense

    2026

  32. [32]

    Atharva Gundawar, Mudit Verma, Lin Guan, Karthik Valmeekam, Siddhant Bhambri, and Subbarao Kambhampati. 2024. Robust Planning with LLM-Modulo Framework: Case Study in Travel Planning.arXiv preprint arXiv:2405.20625 (2024). arXiv:2405.20625 https://arxiv.org/abs/2405.20625

    Pith/arXiv arXiv 2024

  33. [33]

    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

  34. [34]

    Yilun Hao, Yongchao Chen, Yang Zhang, and Chuchu Fan. 2025. Large Lan- guage Models Can Solve Real-World Planning Rigorously with Formal Verifica- tion Tools. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (Volume 1: Long Papers). Association for C...

    Pith/arXiv arXiv 2025

  35. [35]

    Hertz. 2024. Exploring Hertz’s Premium Emergency Roadside Assistance Ser- vice. https://www.hertz.com/us/en/blog/resources/hertz-premium-roadside- assistance. Accessed 2026-02-09

    2024

  36. [36]

    Hertz. n.d.. Child Car Seats — Value-Added Services. https://www.hertz.com/us/ en/products-and-services/value-added-services/united-states/child-car-seats. Accessed 2026-02-09

    2026

  37. [37]

    Angelina Holt. 2023. A Road Trip Guide for Your Hotel Guests: Accommo- dation Essentials. https://www.hoteliga.com/en/blog/a-road-trip-guide-for- your-hotel-guests-accommodation-essentials. Hoteliga blog article. Accessed 2026-02-09

    2023

  38. [38]

    Tzu-Heng Huang, Harit Vishwakarma, and Frederic Sala. 2025. Time To Im- peach LLM-as-a-Judge: Programs are the Future of Evaluation.arXiv preprint arXiv:2506.10403(2025). arXiv:2506.10403 https://arxiv.org/abs/2506.10403

    Pith/arXiv arXiv 2025

  39. [39]

    Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R

    Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R. Narasimhan. 2024. SWE-bench: Can Language Models Resolve Real-World GitHub Issues?. InInternational Conference on Learning Rep- resentations. arXiv:2310.06770 https://openreview.net/forum?id=VTF8yNQM66

    Pith/arXiv arXiv 2024

  40. [40]

    Kao, Maryam Fazel-Zarandi, and Yuandong Tian

    Da Ju, Song Jiang, Andrew Cohen, Aaron Foss, Sasha Mitts, Arman Zharmagam- betov, Brandon Amos, Xian Li, Justine T. Kao, Maryam Fazel-Zarandi, and Yuandong Tian. 2024. To the Globe (TTG): Towards Language-Driven Guaran- teed Travel Planning. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. As...

    Pith/arXiv arXiv 2024

  41. [41]

    Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Kaya Stechly, Mudit Verma, Siddhant Bhambri, Lucas Paul Saldyt, and Anil B. Murthy. 2024. Position: LLMs Can’t Plan, But Can Help Planning in LLM-Modulo Frameworks. In Proceedings of the 41st International Conference on Machine Learning, Vol. 235. PMLR, 22895–22907. arXiv:2402.01817 https://proceedings.ml...

    Pith/arXiv arXiv 2024

  42. [42]

    Priyanshu Karmakar, Soumyabrata Chaudhuri, Shubhojit Mallick, Manish Gupta, Abhik Jana, and Shreya Ghosh. 2026. TripTide: A Benchmark for Adaptive Travel Planning under Disruptions. InFindings of the Association for Computational Linguistics: ACL 2026. Association for Computational Linguistics, 40269–40292. arXiv:2510.21329 doi:10.18653/v1/2026.findings-a...

    arXiv 2026

  43. [43]

    Katapult. 2023. Is now the time to improve theme park food and dining experi- ences? https://www.katapult.co.uk/is-now-the-time-to-improve-theme-park- food-and-dining-experiences. Accessed 2026-02-09

    2023

  44. [44]

    Kimi Team. 2026. Kimi K2.5: Visual Agentic Intelligence. arXiv:2602.02276 [cs.CL]

    Pith/arXiv arXiv 2026

  45. [45]

    Cheeun Lee and Eun Hak Lee. 2024. Evaluation of Urban Nightlife Attractiveness for Millennials and Generation Z.Cities149 (2024), 104934. doi:10.1016/j.cities. 2024.104934 Accessed 2026-02-09

    arXiv 2024

  46. [46]

    Lee, Gina Cui, Jungkeun Kim, Yuri Seo, and Hyunji Chon

    Jacob C. Lee, Gina Cui, Jungkeun Kim, Yuri Seo, and Hyunji Chon. 2021. Photo Taking Paradox: Contrasting Effects of Photo Taking on Travel Satisfaction and Revisit Intention. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3978205. SSRN Electronic Journal. Accessed 2026-02-09

    2021

  47. [47]

    Sarah Lee. 2025. Accessible Travel for All. https://www.numberanalytics.com/ blog/accessible-travel-guide. Accessed 2026-02-09

    2025

  48. [48]

    Charlie Leocha. 2025. Top Hotel Amenities That Travelers Really Want When Choosing Accommodations. https://www.travelersunited.org/top-hotel- amenities-travelers-really-want/. Accessed 2026-02-04

    2025

  49. [49]

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. 2023. API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguisti...

    doi:10.18653/v1/2023.emnlp-main.187 2023

  50. [50]

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. 2025. AgentBench: Evaluating LLMs as Agents. arXiv:2308.03688 [cs.AI] https://arxiv.org/abs...

    Pith/arXiv arXiv 2025

  51. [51]

    LOVU Travel. 2025. 7 Hotel Amenities That Make Couples Book Direct. https: //business.lovu.travel/7-hotel-amenities-that-make-couples-book-direct. Ac- cessed 2026-02-09

    2025

  52. [52]

    Travel Maestro. 2012. No-man’s Land: the Rising Trend of Women-Only Hotel Floors. https://www.covingtontravel.com/2012/11/no-mans-land-the-rising- trend-of-women-only-hotel-floors/. Covington Travel blog article; Accessed 2026-02-09

    2012

  53. [53]

    Meta AI. 2025. The Llama 4 Herd: The Beginning of a New Era of Na- tively Multimodal AI Innovation. https://ai.meta.com/blog/llama-4-multimodal- intelligence/. Release announcement; Llama 4 Scout and Maverick

    2025

  54. [54]

    Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2024. GAIA: A Benchmark for General AI Assistants. In International Conference on Learning Representations. arXiv:2311.12983 https: //openreview.net/forum?id=fibxvahvs3

    Pith/arXiv arXiv 2024

  55. [55]

    Mistral AI. 2025. Introducing Mistral 3. https://mistral.ai/news/mistral-3/. Release announcement; introduces Mistral Large 3

    2025

  56. [56]

    MMGY Global. 2022. Portrait of Travelers with Disabilities: Mobility & Ac- cessibility. https://www.mmgyglobal.com/news/portrait-of-travelers-with- disabilities/. MMGY Global news release; Accessed 2026-02-09

    2022

  57. [57]

    Moonshot AI. 2025. Kimi K2 Thinking. https://huggingface.co/moonshotai/Kimi- K2-Thinking. Model card

    2025

  58. [58]

    National Civil Rights Museum. n.d.. Plan Your Visit — National Civil Rights Museum. https://civilrightsmuseum.org/visit/. Accessed 2026-02-09

    2026

  59. [59]

    National Park Service. 2025. Socioeconomic Monitoring Visitor Sur- veys. https://www.nps.gov/subjects/socialscience/socioeconomic-monitoring- visitor-surveys.htm. Accessed 2026-02-09

    2025

  60. [60]

    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. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 12392–12418. ar...

    arXiv 2025

  61. [61]

    nmu. 2022. How To Make Travel a Kid-Friendly Experience at Your Ho- tel. https://www.thesolutionsdesk.com/how-to-make-travel-a-kid-friendly- experience-at-your-hotel/. TheSolutionsDesk (Guest Supply blog). Accessed 2026-02-09

    2022

  62. [62]

    OpenAI, :, Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex Iftimie, Alex Karpenko, Alex Tachard Passos, Alexander Neitz, Alexan- der Prokofiev, Alexander Wei, Allison Tam, Ally Bennett, Ananya Kumar, An- dre Saraiva, Andrea Vallone, Andrew Duberstein, Andrew Ko...

    Pith/arXiv arXiv 2024

  63. [63]

    OpenAI. 2025. gpt-oss-120b & gpt-oss-20b Model Card. arXiv:2508.10925 [cs.CL]

    Pith/arXiv arXiv 2025

  64. [64]

    OpenAI. 2026. GPT-5.6: Frontier Intelligence that Scales with Your Ambition. https://openai.com/index/gpt-5-6/. Model release; GPT-5.6 Sol (frontier tier) via the Responses API

    2026

  65. [65]

    OurAirports. 2026. Open Airport Data. https://ourairports.com/data/. Public domain dataset containing worldwide airport information. Accessed 2026-02-04

    2026

  66. [66]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe

  67. [67]

    Srikant Panda, Amit Agarwal, and Hitesh Laxmichand Patel. 2025. AccessEval: Benchmarking Disability Bias in Large Language Models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Associ- ation for Computational Linguistics. arXiv:2509.22703 doi:10.18653/v1/2025. emnlp-main.1653

    arXiv 2025

  68. [68]

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2023. ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs. arXiv:2307.16789 [cs.AI] https://arxiv.org/a...

    Pith/arXiv arXiv 2023

  69. [69]

    Qwen Team, Alibaba. 2025. Qwen3-Next-80B-A3B-Instruct. https://huggingface. co/Qwen/Qwen3-Next-80B-A3B-Instruct. Model card

    2025

  70. [70]

    Reddit. 2025. LPT: In touristy areas, tourist trap restaurants may have high online ratings from clueless tourists. https://www.reddit.com/r/LifeProTips/ comments/1kgu263/lpt_in_touristy_areas_tourist_trap_restaurants/. Accessed 2026-02-09

    2025

  71. [71]

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics. arXiv:2304.11406 doi:10.18653/v1/2024.acl-long.399

    Pith/arXiv arXiv 2024

  72. [72]

    Sawgrass Marketing. 2023. Seven Travel Personas You Need to Know for Niche Hospitality Marketing. https://www.sawgrassmktg.com/blog/seven-travel- personas-you-need-to-know-for-niche-hospitality-marketing. Accessed 2026- 02-09

    2023

  73. [73]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessí, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: language models can teach themselves to use tools. InProceedings TREK: A Travel Reasoning and Evaluation Kit for LLM Agents in Complex Trip Planning KDD ’27, August 2027, of the 37th Internation...

    2023

  74. [74]

    Jie-Jing Shao, Bo-Wen Zhang, Xiao-Wen Yang, Baizhi Chen, Si-Yu Han, Jinghao Pang, Wen-Da Wei, Guohao Cai, Zhenhua Dong, Lan-Zhe Guo, and Yu-Feng Li

  75. [75]

    Zijian Shao, Jiancan Wu, Weijian Chen, and Xiang Wang. 2025. Personal Travel Solver: A Preference-Driven LLM-Solver System for Travel Planning. InPro- ceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers). Association for Computational Linguistics. doi:10.18653/v1/2025.acl-long.1339

    work page doi:10.18653/v1/2025.acl-long.1339 2025

  76. [76]

    Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, Weicheng Ma, and Soroush Vosoughi. 2024. Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge.arXiv preprint arXiv:2406.07791(2024). arXiv:2406.07791 https: //arxiv.org/abs/2406.07791

    arXiv 2024

  77. [77]

    SiteMinder. 2024. Hotel amenities: Examples and ideas list. https: //www.siteminder.com/r/trends-advice/hotel-management-tips-ideas/hotel- amenities-property-services/. Accessed 2026-02-09

    2024

  78. [78]

    Matylda Siwek, Anna Kolasińska, Krzysztof Wrześniewski, and Mag- dalena Zmuda Palka. 2022. Services and Amenities Offered by City Hotels within Family Tourism as One of the Factors Guaranteeing Satisfactory Leisure Time.International Journal of Environmental Research and Public Health19, 14 (2022), 8321. doi:10.3390/ijerph19148321 Accessed 2026-02-09

    work page doi:10.3390/ijerph19148321 2022

  79. [79]

    Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. 2024. Large Language Models are Inconsistent and Biased Evaluators.arXiv preprint arXiv:2405.01724 (2024). arXiv:2405.01724 https://arxiv.org/abs/2405.01724

    Pith/arXiv arXiv 2024

  80. [80]

    Megan Sullivan. 2013. Pet-Friendly Hotels Prove Profitable. https:// lodgingmagazine.com/profiting-from-pets/. Accessed 2026-02-09

    2013

Showing first 80 references.