Pith. sign in

REVIEW 4 major objections 3 minor 2 cited by

RETAIL: Towards Real-world Travel Planning for Large Language Models

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper argues that real-world travel planning by LLMs is far harder than current benchmarks suggest, and that the missing pieces are implicit-query clarification, environmental awareness, plan revision, and all-in-one plan detail.

desk verdict A useful new benchmark for LLM travel planning, but the headline pass rates conflate hidden-slot recovery with plan quality. read the letter →

arxiv 2508.15335 v1 pith:AK6QGJRY submitted 2025-08-21 cs.AI

Bin Deng , Yizhe Feng , Zeming Liu , Qing Wei , Xiangrong Zhu , Shuai Chen , Yuanfang Guo , Yunhong Wang This is my paper
classification cs.AI
keywords travelplanninglargelanguagemodelsbenchmarkdatasetimplicitqueriesdecision-makingsupportmulti-agentframeworkenvironmentalawarenessplanevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that LLM-based travel planning fails on real-world trips because the problem has been set up too simply: prior benchmarks give users explicit, complete queries, ignore weather and other dynamic conditions, and settle for bare POI lists. The authors build RETAIL, a 10,182-dialogue dataset in which queries are often implicit, plans must be revised, and final itineraries include tickets, room types, reviews, and costs, grounded in a knowledge base of over 60,000 points of interest across 24 Chinese cities. They score generated plans against 14 constraints drawn from commonsense and user preferences. On this test, the strongest existing model passes only 1.0% of plans, while the authors' TGMA framework—topic-guided dialogue management plus three specialized agents—raises the pass rate to 2.72% in the main setting. The practical upshot: real-world travel planning by LLMs is still far from viable, but decomposing the task into guided clarification and staged plan generation is a measurable step forward.

What carries the argument

The load-bearing mechanism is the RETAIL evaluation protocol combined with the TGMA planning pipeline. RETAIL supplies the real-world substrate—24 cities, 60,279 POIs, weather and transport data, and 10,182 dialogues covering explicit and implicit queries, each with or without revision—and defines feasibility as satisfying 14 constraints simultaneously, with novel checks for spatial-temporal consistency (e.g., 30-minute minimum activity gaps, valid POIs, hotels matched to preferences). TGMA is the countermeasure: a topic-guided interaction logic selects among eight dialogue states across five knowledge domains to keep clarification natural, while three agents (intent detection, overall plann

What would settle it

Take a set of plans that pass all 14 constraints and a matched set that fail exactly one, anonymize them, and ask a panel of real travelers or professional trip planners which they would actually use. If the failed plans are chosen as often as the passed ones, the benchmark's pass rates are artifacts of the thresholds. A quicker check: relax the 30-minute time interval and the 4-activity minimum, rerun Table 4, and see whether the 2.72% vs 1.0% gap and the ranking of methods survive.

Watch

Extended reading notes

Core claim

The paper's central claim is that real-world travel planning remains extremely challenging for LLMs, and that the barrier is as much the benchmark's definition of the task as the models themselves. RETAIL operationalizes the missing realism: implicit queries that require clarification, environmental factors such as weather and transportation that can trigger plan revision, and all-in-one plans that carry concrete details like ticket prices, room types, and user reviews. The evaluation reduces 'plan quality' to 14 checkable constraints—ten commonsense (city coverage, activity repetition, 30-minute time intervals, accommodation each night, daily schedule, return journey, POI existence, locatio

Load-bearing premise

The headline pass rates assume that the 14 hand-designed constraints—such as a 30-minute minimum gap between activities and a minimum of 4 activities per city per day—are the right definition of a feasible, high-quality travel plan; if real travelers would accept plans that fail these thresholds, or reject plans that pass them, the 1.0% and 2.72% numbers stop being evidence about real-world readiness.

Editorial extensions

If this is right

  • If the 1.0% vs 2.72% gap holds, task decomposition—guided clarification plus staged plan generation—is a productive direction for real-world travel planning.
  • RETAIL gives the field a benchmark where implicit queries, weather, transport, revisions, and rich POI details are all present, so future systems can be measured on the full planning pipeline instead of itinerary text alone.
  • Constraint-level analysis shows time interval and POI validation dominate commonsense failures, while hotel type and required sites dominate preference failures, identifying concrete targets: better activity scheduling, POI verification, and accommodation compliance.
  • Near-zero macro pass rates for baselines imply that models satisfy constraints on average but not consistently across cases, making consistency, not average quality, the key bottleneck.
  • Because plans must include real ticket prices, room types, and reviews, the benchmark raises the bar for what counts as a usable travel plan in an automated system.

Reading between the lines

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

  • If the 14 constraints were calibrated against actual traveler preferences—say, by having real users choose between plans that pass and plans that fail exactly one check—the headline pass rates might move substantially; the paper does not report such calibration.
  • The dataset is tied to one country and one commercial data source, so the 1% and 2.72% figures may not generalize to other languages, POI ecosystems, or booking conventions; an international replication would test that.
  • The appendix result that one-shot prompting outperforms chain-of-thought for the best configuration suggests prompting format and framework modules may be confounded; ablating those separately would isolate TGMA's true contribution.
  • Because many constraints involve judgment (how 'nearby' restaurants are, what counts as a must-visit site), measuring annotator agreement on constraint application, not just on dialogue quality, would strengthen the reliability of the pass rates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper introduces RETAIL, a Chinese tourism dialogue and all-in-one travel-planning dataset with 10,182 dialogues, and TGMA, a topic-guided multi-agent framework. Implicit queries are generated by removing fields from explicit queries; plans are evaluated by 9 commonsense and 4 user-preference constraints. Experiments report that the best baseline reaches 1.0% final pass rate while TGMA reaches 2.72%, supporting the claim that real-world travel planning is extremely challenging.

Significance. The contribution is potentially significant. RETAIL is larger and more realistic than previous travel-planning datasets: it includes decision-making support, revision cases, weather/transportation data, and detailed POI information over 24 cities. The multi-agent framework is a reasonable design, and the ablations (Tables 13–14) show that both topic guidance and the intent/overall-plan agents contribute. The code/data release is promised, which would aid reproducibility. However, the headline quantitative claims rest on an evaluation protocol whose construct validity needs to be demonstrated; the current manuscript does not yet support the strength of the abstract's conclusions.

major comments (4)
  1. [3.2.1, 5.2, Table 5] The headline comparison (1.0% vs 2.72%) is computed against hidden-slot ground truth. Section 3.2.1 constructs implicit queries by removing fields from a 12-field intention slot; Table 5 then scores plans against Budget, Hotel Type, Required Sites, and Excluded Sites, i.e., exactly those hidden fields. A plan that clarifies a different but equally reasonable budget or hotel type is counted as failed. The human evaluation in Appendix A.2 compares DMS vs no-DMS but never tests whether preference-failing plans are acceptable to a real user. This conflates exact hidden-slot recovery with plan quality and directly affects the central 'extremely challenging' conclusion. Please (i) have human raters judge preference-failing plans that otherwise pass commonsense constraints, and/or (ii) report sensitivity to relaxed thresholds (e.g., Time Interval 20/45 min, Activity Count 3/5, Budget ±20%).
  2. [Table 2 vs Appendix C.2 Table 22] Dataset statistics are internally inconsistent. Section 3.4/Table 2 state 10,182 total dialogues (6,000/2,182/2,000 train/val/test). Appendix C.2 Table 22 lists Total column values summing to 10,000, and each row's Total does not equal the sum of its Training/Val/Test entries (e.g., Single-turn: 3,500 vs 2,100+830+700=3,630; Multi-turn: 2,500 vs 1,500+1,167+500=3,167). The split sums actually add to 10,182, so the Total column appears to be erroneous, but as written the dataset description is contradictory. This must be corrected since dataset scale is a core contribution.
  3. [Section 5.4 Table 4; A.4 Tables 11-12] The final pass-rate claims are not reproducible from the tables. The main result for TGMA+Qwen2.5-7B-Instruct(SFT) is 2.72% in Table 4, but Appendix A.4 reports 2.99% under CoT and 4.06% under one-shot for the same configuration. No prompting setting is specified for Table 4. Please state the exact evaluation protocol and reconcile the numbers.
  4. [Section 5.4 Table 4] The key comparisons involve very small counts: 1.0% is about 20/2,000, 2.72% about 54/2,000. No confidence intervals or significance tests are reported. With counts this small, the gap between Qwen-Plus (1.00%) and TGMA (2.72%) may be within sampling noise. Report Wilson intervals or bootstrap CIs, and ideally per-category breakdowns and failure decomposition.
minor comments (3)
  1. [Section 6 RQ1] The sentence '14 types of constraints (9 commonsense constraints and 4 user preference constraints)' adds to 13. Table 5 lists 9 commonsense and 4 preference metrics. Correct the count.
  2. [Appendix A.5.2] The opening sentence says 'As shown in Table 16', but the ablation results are in Table 14. The cross-reference is wrong.
  3. [Abstract / Table 4] The phrase 'strongest existing model' is not tied to a specific baseline in Table 4. Qwen-Plus has the highest baseline pass rate (1.00%), while gpt-4o has only 0.10%. Please name the model and state the evaluation protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the benchmark results are empirical measurements and the minor self-citations are not load-bearing.

full rationale

The paper's central claims are empirical: RETAIL is a newly constructed dataset, and the 1.0% vs. 2.72% final pass rates are measured outcomes on that dataset under the stated constraint metrics. No equation in the paper defines a predicted quantity in terms of a fitted parameter, and no derivation reduces the final pass rate to the dataset construction process. The construction of implicit queries by removing fields from explicit queries (Section 3.2.1) and then evaluating against those hidden fields (Budget, Hotel Type, Required Sites, Excluded Sites in Table 5) is a benchmark-design and construct-validity concern, not circularity: the pass rate is not forced by construction, since models can fail or succeed on many independent constraints. The paper's use of self-citations is limited to methodological references: Section 3.3 and Section 5.2 cite Shi et al. (2023) and Liu et al. (2022) for human evaluation practice and standard dialogue metrics (BLEU, ROUGE, METEOR). These citations are not load-bearing for the main claim; the metrics are standard external tools, and the dataset quality score 0.85 is the authors' own measurement, not a value imported from the cited papers. There is no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result as a new contribution. The Limitations section concerns computational resources and context length, which are practical constraints rather than circular dependencies. Overall, the derivation chain is self-contained with respect to the dataset, framework, and evaluation; any concerns about whether the hand-designed constraints reflect real-world traveler preferences belong under correctness risk, not circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The ledger shows that the central numbers depend on several hand-chosen thresholds and on the faithfulness of GPT-4o-simulated data. These are not fitted to an external ground truth, so they are domain assumptions entered without independent evidence.

free parameters (5)
  • Time Interval threshold = 30 minutes
    In Table 5, the Time Interval constraint requires a 30-minute minimum gap between activities; this is a hand-chosen number that directly affects the pass rate.
  • Activity Count minimum = 4 activities per city daily
    Table 5 defines an Activity Count constraint of at least 4 activities per city per day; this is an arbitrary threshold used to score plans.
  • Nearby POI linkage count = 3-5 restaurants and hotels per attraction
    Section 3.1 states each attraction is linked to 3-5 nearby restaurants and hotels, constraining the plan generation search space.
  • Destination and attraction counts = 2-4 destination cities, 8-10 key attractions
    Section 3.2.1 uses these ranges to construct explicit queries; the ranges are chosen by the authors and affect dataset difficulty.
  • Intention slot size = 12 fields
    Section 3.2.1 removes fields from a 12-field intention slot to create implicit queries; the number and composition of fields is a design choice.
assumptions (3)
  • domain assumption GPT-4o-simulated dialogues are representative of real-world tourist-agent interactions
    The dataset construction in Section 3.2.2 uses GPT-4o to simulate all tourist and agent utterances. The entire 'real-world' claim rests on this synthetic dialogue being a faithful proxy for human behavior, which is not validated against real user logs.
  • domain assumption The constraint set in Table 5 defines plan quality
    The final pass rate is the conjunction of 13 hand-designed constraints (30-minute gaps, 4 activities/day, budget, hotel type, etc.). The paper assumes these are both necessary and sufficient for a practical travel plan, with no human study linking constraint satisfaction to user satisfaction.
  • domain assumption The knowledge base (Meituan POIs, 2345 weather) is accurate and complete for the test period
    Sections 3.1 and 3.3 describe web-crawled data with manual spot checks, but the full knowledge base is not released, so POI existence, prices, and weather are taken on faith, which affects POI Validation and weather-dependent revision cases.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RETAIL: Towards Real-world Travel Planning for Large Language Models." pith.science (2026). https://pith.science/paper/AK6QGJRY

@misc{pith2026250815335,
  author       = {Pith},
  title        = {Pith review of: RETAIL: Towards Real-world Travel Planning for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AK6QGJRY}},
  note         = {Machine review of arXiv:2508.15335}
}
read the original abstract

Although large language models have enhanced automated travel planning abilities, current systems remain misaligned with real-world scenarios. First, they assume users provide explicit queries, while in reality requirements are often implicit. Second, existing solutions ignore diverse environmental factors and user preferences, limiting the feasibility of plans. Third, systems can only generate plans with basic POI arrangements, failing to provide all-in-one plans with rich details. To mitigate these challenges, we construct a novel dataset \textbf{RETAIL}, which supports decision-making for implicit queries while covering explicit queries, both with and without revision needs. It also enables environmental awareness to ensure plan feasibility under real-world scenarios, while incorporating detailed POI information for all-in-one travel plans. Furthermore, we propose a topic-guided multi-agent framework, termed TGMA. Our experiments reveal that even the strongest existing model achieves merely a 1.0% pass rate, indicating real-world travel planning remains extremely challenging. In contrast, TGMA demonstrates substantially improved performance 2.72%, offering promising directions for real-world travel planning.

Figures

Figures reproduced from arXiv: 2508.15335 by the authors.

Figure 1
Figure 1. Comparison between TravelPlanner (Xie et al., 2024) and RETAIL, highlighting our UGC-enriched decision-making support for implicit queries, environmental awareness, and all-in-one travel plans with rich details. etc, ensuring practical feasibility and mitigating the third challenge of lacking comprehensive travel planning. As shown on the right side of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of TGMA framework: the Topic-Guided Interaction Logic handles user queries through [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Impact of Different Constraints on Pass Rates. (a) CC: City Coverage, AR: Activity Repetition, TI: Time [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Geographic distribution of the 24 cities in our [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

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

    cs.CL 2026-07 conditional novelty 6.5 of 10

    On 800 jointly constrained trip tasks with a deterministic scorer and achievable gold, the best of 15 LLM agents fully solves only 46.2% of feasible plans, with unstated persona needs as the universal bottleneck.

  2. Interface fluctuations for $1$D stochastic Allen-Cahn equation -- singular regime

    math.PR 2025-08 unverdicted novelty 6.0 of 10

    For the 1D stochastic Allen-Cahn equation with half-derivative spacetime white noise, small noise and long-time scaling make the interface location move like an approximate diffusion, with two divergences canceling.

Reference graph

Works this paper leans on

32 extracted references · 19 canonical work pages · cited by 2 Pith papers

  1. [2]

    CoRR, abs/2406.10196

    TRIP-PAL: travel planning with guarantees by com- bining large language models and automated plan- ners. CoRR, abs/2406.10196. DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingx- uan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Da...

    arXiv

  2. [3]

    CoRR, abs/2412.19437

    Deepseek-v3 technical report. CoRR, abs/2412.19437. Pierluca Ferraro and Giuseppe Lo Re

    arXiv

  3. [4]

    arXiv preprint arXiv:2406.12793

    Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793. Feiran Huang, Jie Xu, and Jian Weng

    arXiv

  4. [6]

    Towards Full Delegation: Designing Ideal Agentic Behaviors for Travel Planning

    Towards full delegation: Designing ideal agentic behaviors for travel planning. CoRR, abs/2411.13904. Da JU, Song Jiang, Andrew Cohen, Aaron Foss, Sasha Mitts, Arman Zharmagambetov, Brandon Amos, Xian Li, Justine T. Kao, Maryam Fazel-Zarandi, and Yuandong Tian

    work page Pith review arXiv

  5. [7]

    CoRR, abs/2410.16456

    To the globe (TTG): towards language-driven guaranteed travel planning. CoRR, abs/2410.16456. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- taka Matsuo, and Yusuke Iwasawa

    arXiv

  6. [10]

    CoRR, abs/2410.14940

    Baichuan alignment technical report. CoRR, abs/2410.14940. Zeming Liu, Jun Xu, Zeyang Lei, Haifeng Wang, Zheng- Yu Niu, and Hua Wu

    arXiv

  7. [11]

    In Proceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 1024–1034

    Where to go for the holi- days: Towards mixed-type dialogs for clarification of user goals. In Proceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 1024–1034. Association for Computational Linguistics. Joel Pinho Lucas, Nuno Luz, María N. Moreno...

    work page 2022

  8. [12]

    CoRR, abs/2303.08774

    GPT-4 technical report. CoRR, abs/2303.08774. OpenAI

    arXiv

Show all 32 references
  1. [13]

    https://openai.com/ index/hello-gpt-4o/

    Hello gpt-4o. https://openai.com/ index/hello-gpt-4o/. Accessed: 2025-01-15. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu

    2025

  2. [15]

    In The Twelfth International Con- ference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11,

    Toolllm: Fa- cilitating large language models to master 16000+ real-world apis. In The Twelfth International Con- ference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11,

    2024

  3. [16]

    In Advances in Neural Information Pro- cessing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16,

    Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Pro- cessing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16,

    2023

  4. [18]

    CoRR, abs/2412.13682

    Chinatravel: A real-world benchmark for language agents in chinese travel planning. CoRR, abs/2412.13682. Xiaoming Shi, Zeming Liu, Chuan Wang, Haitao Leng, Kui Xue, Xiaofan Zhang, and Shaoting Zhang

    arXiv

  5. [19]

    In Proceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8145–8157

    Midmed: Towards mixed-type dialogues for medi- cal consultation. In Proceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8145–8157. Associa- tion for Computational L...

    2023

  6. [20]

    In Advances in Neural Information Pro- cessing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16,

    Re- flexion: language agents with verbal reinforcement learning. In Advances in Neural Information Pro- cessing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16,

    2023

  7. [21]

    In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: EMNLP 2024 - Industry Track, Miami, Florida, USA, Novem- ber 12-16, 2024 , pages 486–514

    Personal large language model agents: A case study on tailored travel planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: EMNLP 2024 - Industry Track, Miami, Florida, USA, Novem- ber 12-16, 2024 , pages 486–514. Association for ...

    2024

  8. [22]

    In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing: EMNLP 2024 - Industry Track, Miami, Florida, USA, November 12-16, 2024, pages 1413–1432

    Itinera: Integrating spatial optimiza- tion with large language models for open-domain urban itinerary planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing: EMNLP 2024 - Industry Track, Miami, Florida, USA, November 12-16, 2024,...

    2024

  9. [23]

    CoRR, abs/2302.13971

    Llama: Open and efficient foundation language models. CoRR, abs/2302.13971. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V . Le, and Denny Zhou

    arXiv

  10. [24]

    Chain-of-thought prompting elicits reasoning in large language models. In Ad- vances in Neural Information Processing Systems 35: Annual Conference on Neural Information Process- ing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9,

    2022

  11. [25]

    CoRR, abs/2405.18208

    A human-like reasoning framework for multi-phases planning task with large language models. CoRR, abs/2405.18208. Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su

    arXiv

  12. [26]

    In Forty-first International Conference on Machine Learning, ICML 2024, Vi- enna, Austria, July 21-27,

    Travelplanner: A benchmark for real-world planning with language agents. In Forty-first International Conference on Machine Learning, ICML 2024, Vi- enna, Austria, July 21-27,

    2024

  13. [27]

    In Proceedings of the 32nd ACM International Con- ference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 6969–6978

    P-RAG: progressive retrieval augmented generation for planning on embodied everyday task. In Proceedings of the 32nd ACM International Con- ference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 6969–6978. ACM. An Yang, Baosong Yang...

    2024

  14. [28]

    CoRR, abs/2412.15115

    Qwen2.5 technical report. CoRR, abs/2412.15115. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao

    arXiv

  15. [29]

    In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  16. [30]

    Scientific Programming, 2022(1):5495822

    Analysis and study on intelli- gent tourism route planning scheme based on weighted mining algorithm. Scientific Programming, 2022(1):5495822. Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dong- sheng Li, and Deqing Yang

    2022

  17. [31]

    CoRR, abs/2406.14228

    Evoagent: To- wards automatic multi-agent generation via evolu- tionary algorithms. CoRR, abs/2406.14228. Xiyuan Zhang, Ranak Roy Chowdhury, Rajesh K Gupta, and Jingbo Shang. 2024a. Large language models for time series: A survey. arXiv preprint arXiv:2402.01801. Xuan Zhang, Y...

    2024 arXiv

  18. [32]

    useful for refining initial ideas,

    and our benchmark, while those in brown are unique to our benchmark. As for automated metrics, we employed the same evaluation metrics used in Travel Plan Evalu- ation 5.4 to assess plans generated with and without assisted decision-making processes. The results in Table 6 dem...

    1998

  19. [2002]

    In Proceedings of the 40th Annual Meeting of the Association for Compu- tational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311–318

    Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Compu- tational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311–318. ACL. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, ...

    2002

  20. [2007]

    In Proceed- ings of the Second Workshop on Statistical Machine Translation, WMT@ACL 2007, Prague, Czech Re- public, June 23, 2007, pages 228–231

    METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In Proceed- ings of the Second Workshop on Statistical Machine Translation, WMT@ACL 2007, Prague, Czech Re- public, June 23, 2007, pages 228–231. Association for Computational L...

    2007

  21. [2008]

    In 20th IEEE Inter- national Conference on Tools with Artificial Intelli- gence (ICTAI 2008), November 3-5, 2008, Dayton, Ohio, USA, Volume 2, pages 89–96

    e-tourism: A tourist recommenda- tion and planning application. In 20th IEEE Inter- national Conference on Tools with Artificial Intelli- gence (ICTAI 2008), November 3-5, 2008, Dayton, Ohio, USA, Volume 2, pages 89–96. IEEE Computer Society. Jie-Jing Shao, Xiao-Wen Yang, Bo-W...

    2008

  22. [2022]

    Large lan- guage models are zero-shot reasoners. In Advances in Neural Information Processing Systems 35: An- nual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9,

    2022

  23. [2023]

    CoRR, abs/2310.06825

    Mistral 7b. CoRR, abs/2310.06825. Song Jiang, Da JU, Andrew Cohen, Sasha Mitts, Aaron Foss, Justine T. Kao, Xian Li, and Yuandong Tian

    arXiv

  24. [2024]

    CoRR, abs/2409.08069

    Travelagent: An AI as- sistant for personalized travel planning. CoRR, abs/2409.08069. Gang Chen, Sai Wu, Jingbo Zhou, and Anthony K. H. Tung

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.