Pith. sign in

REVIEW 5 major objections 4 minor 67 references

CoMaPOI: A Collaborative Multi-Agent Framework for Next POI Prediction Bridging the Gap Between Trajectory and Language

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Three collaborating LLM agents outperform 11 baselines at next-POI prediction by translating trajectories into language and constraining the candidate set.

desk verdict CoMaPOI is a competent, reproducible LLM multi-agent system for next-POI prediction with genuine but modest gains; the 5-10% headline and the theoretical proof are the real weak spots. read the letter →

arxiv 2505.23837 v1 pith:IXSPLTL3 submitted 2025-05-28 cs.CL cs.IR

Lin Zhong , Lingzhi Wang , Xu Yang , Qing Liao This is my paper
classification cs.CLcs.IR
keywords NextPOIpredictionLargelanguagemodelsMulti-agentcollaborationCandidatesetoptimizationReversereasoningfine-tuningSpatiotemporaltrajectoryunderstandingRetrieval-augmentedgenerationLocation-basedsocialnetworks
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

This paper claims that the two obstacles to using large language models for next point-of-interest (POI) prediction—weak comprehension of numeric spatiotemporal check-ins and an enormous unconstrained candidate space—can both be overcome by a collaborative multi-agent design. CoMaPOI assigns one agent (Profiler) to translate trajectories into language profiles and mobility patterns, a second (Forecaster) to retrieve and shrink the candidate POI set, and a third (Predictor) to make the final choice. On the NYC, TKY, and CA benchmarks it reports state-of-the-art results, beating all 11 baselines and improving the best baseline by roughly 2 to 12 percent per metric. A sympathetic reader would take the paper's contribution to be a template for making LLMs usable on spatiotemporal ranking tasks without modifying the LLM architecture.

What carries the argument

The load-bearing machinery is the three-agent decomposition. The Profiler runs statistical tools (frequency, category, time, location) over historical and current trajectories and asks the LLM to write a long-term user profile and a short-term mobility pattern in natural language. The Forecaster embeds POI descriptions in a vector database, retrieves a raw candidate set by cosine similarity, then prompts the LLM to re-rank and shrink it against the profile and pattern, producing long-term and short-term candidate sets. The Predictor receives trajectory, profile, pattern, and both candidate sets and returns a POI. Fine-tuning uses Reverse Reasoning Fine-Tuning (RRF): the true next POI $p^*$ is used to construct the 'ideal' profile, pattern, and candidate sets, and each agent is trained to reproduce these from trajectory-only inputs. The paper also derives a sufficient condition, Eq. 17, under which supplying a candidate set with hit rate $\mathbb{P}(p^*\in C)$ lowers expected error below global search, and reports measured hit rates that satisfy it.

What would settle it

Train the Profiler and Forecaster with RRF labels produced without access to the true next POI (for example, using a strong baseline's top prediction instead), then compare HR@10 on the CA dataset; if the gap between CoMaPOI and the w/o-RRF ablation collapses, the original RRF gains come from leaked answer information rather than transferable reasoning.

Watch

Extended reading notes

Core claim

The paper's central claim is that next-POI prediction becomes state-of-the-art when a single LLM is replaced by three specialized LLM agents whose outputs feed one another. The Profiler converts raw coordinates, timestamps, and categories into statistical language descriptions, giving the model a handle on spatiotemporal distributions it cannot read directly from numbers. The Forecaster uses semantic retrieval to build an initial candidate set and then, guided by the Profiler's profile and pattern, re-ranks and narrows it so the true next POI is much more likely to appear near the top. The Predictor fuses trajectory, profile, pattern, and both candidate sets to emit the final POI. The reported experiments show consistent wins over all baselines on all three datasets, and the paper derives a hit-rate inequality (Eq. 17) as a sufficient condition for candidate-set guidance to reduce prediction error, with measured values satisfying it on every dataset.

Load-bearing premise

The training scheme assumes that the ideal profiles, mobility patterns, and candidate sets built from the true next POI can be reconstructed from trajectory data alone when that true POI is unknown at test time.

Editorial extensions

If this is right

  • The same three-agent pattern—language translation, candidate constraint, final prediction—can be applied to any LLM-based ranking task with a huge output space, not just POIs.
  • The framework is architecture-independent in the paper's experiments: swapping in different open LLMs and even much smaller models preserves most of the gain, so the approach scales down to resource-constrained deployment.
  • The Forecaster's candidate sets are directly useful as an interpretable intermediate: they show which POIs were considered and why, making the final prediction easier to audit.
  • The hit-rate inequality gives practitioners a concrete diagnostic: measure $\mathbb{P}(p^*\in C)$ and the conditional success probabilities, and check whether the candidate set is actually helping before trusting it.
  • RRF is indispensable in the ablations: removing it causes the largest single drop, so the fine-tuning strategy, not only the architecture, carries the result.

Reading between the lines

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

  • Editorial inference: the paper's candidate-set argument is agnostic to how the candidate set is built, so the same inequality could be used to evaluate candidate generators from classical recommender systems, not just LLM-based retrieval.
  • Editorial inference: because RRF builds its supervision from the ground-truth next POI, part of the measured gain may reflect the agents learning to recognize artifacts of those labels; a cleaner test would construct RRF targets from a held-out predictor's top choices and compare.
  • Editorial inference: the multi-agent pipeline costs multiple LLM calls per prediction; the authors' single-agent RRF variant suggests a cost–accuracy frontier that a deployment-oriented follow-up could map explicitly.
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

5 major / 4 minor

Summary. CoMaPOI is a multi-agent LLM framework for next-POI prediction. Three specialized agents (Profiler, Forecaster, Predictor) convert numeric trajectories into language, retrieve and refine a candidate set, and make the final prediction. The paper introduces Reverse Reasoning Fine-Tuning (RRF), which constructs oracle-style training labels (ideal profiles, mobility patterns, candidate sets that always contain the ground-truth next POI p*) by backward reasoning from p*, then fine-tunes the agents to reproduce those labels from trajectory data alone. Experiments on NYC, TKY, and CA claim consistent improvements over 11 baselines, and Section 4.3 claims a theoretical proof that candidate-set optimization reduces prediction error.

Significance. If the empirical gains are genuine and robust, CoMaPOI is a practically useful contribution to LLM-based POI prediction: it demonstrates a modular, tool-augmented multi-agent design that can be applied across different LLM backbones, and it makes code publicly available. However, the significance is conditioned on two concerns: the RRF training strategy may introduce label leakage that inflates test performance, and the 'theoretical proof' in Section 4.3 is an algebraic restatement rather than a substantive theorem. The paper's central claim of state-of-the-art performance therefore needs additional evidence to rule out shortcut learning and to establish that the gains transfer to realistic inference settings.

major comments (5)
  1. [Section 4.2, Eqs. (10)-(11); Table 3] The RRF strategy constructs training candidate sets that always contain the true next POI p* and reverse-engineers user profiles and mobility patterns to be maximally relevant to p*. The agents are then fine-tuned to reproduce these oracle-derived labels from trajectory data. At inference, p* is unknown and the candidate set may miss the answer; Table 3 shows that the out-of-candidate success rate is only 0.31% (NYC), 0.33% (CA), and 13.17% (TKY), meaning the framework almost never recovers the label when the candidate set lacks it. This is exactly the signature of a model that has learned to exploit the guaranteed presence of p* in the training candidate sets rather than robust spatiotemporal reasoning. The paper should directly test this by, e.g., training without the guarantee that p* is in the candidate set, or by evaluating inference where p* is deliberately excluded from the candidate set, to show that the reported gains are not an artifact of this leakage. This is load-bearing because the central claim is that CoMaPOI outperforms baselines by genuinely improved reasoning.
  2. [Section 4.3, Eq. (17)] The 'theoretical proof' that candidate-set provision reduces prediction error is tautological. Equation (17) is obtained by rearranging the total-probability decomposition in Eqs. (13)-(14); it is a necessary condition that must hold for the candidate-based error to be smaller, but it does not establish that this condition actually holds in general. The verification in Table 3 uses the same test-set predictions that produced the reported results, which is circular. The authors should either remove the claim of a theoretical proof or replace it with a substantive analysis, such as deriving conditions under which Eq. (17) is expected to hold based on candidate-set hit rates and posterior calibration, rather than merely restating the inequality.
  3. [Abstract and Table 1] The abstract states that CoMaPOI improves 'all metrics by 5% to 10%' compared to the best baselines, but Table 1 shows TKY HR@5 improves by 3.81% and TKY MRR by only 2.05% — both below the advertised 5% range. The improvement is also not uniform across datasets or metrics. The quantitative claim should be corrected to match the reported numbers.
  4. [Section 5.5, Figure 3(a)] The candidate-set size K=25 is selected based on the test-set performance curve in Figure 3(a). Choosing a hyperparameter on the test set leaks test information into model selection and can overestimate performance. The authors should either select K on a validation split or explicitly report this as an optimistic bias. Since the candidate-set size is central to the method, this is more than a minor methodological detail.
  5. [Section 5.2, Table 1] No error bars, confidence intervals, or significance tests are reported for any baseline or for CoMaPOI. Given that the central claim is 'state-of-the-art' performance and that the margins on some metrics (e.g., TKY MRR: 31.82 vs. 31.18) are small, the authors should at least report variance across repeated runs or a statistical comparison to the strongest baseline. The reader cannot assess whether the improvements are reliable.
minor comments (4)
  1. [Section 4.2, Eqs. (10)-(11)] The notation in the arg max expressions appears inconsistent: Eq. (10) writes the maximization over M on the left-hand side but the score uses C_H; Eq. (11) writes over C but the score uses M_u. Please clarify the intended optimization variables.
  2. [Section 5.3, Figure 2] The text describes 'CoMaPOI (RRF)' as a single-agent variant, but the name suggests it is the framework with RRF fine-tuning, which is the default configuration elsewhere. Please rename or clearly disambiguate this variant to avoid confusion.
  3. [References] Reference [4] is garbled with repeated author names and ellipses; this should be fixed before publication.
  4. [Section 5.4, Table 2] The 'w/o RRF' ablation removes the RRF fine-tuning strategy, but it is not clear whether the agents are then completely unfine-tuned or fine-tuned with a different objective. Clarify what 'w/o RRF' means operationally, especially because the 'only SFT' row appears to describe an alternative fine-tuning procedure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA claim rests on external benchmark comparisons, and the oracle-derived RRF labels are a training-time supervision construction rather than a circular derivation of the reported results.

full rationale

The paper's central claim is empirical: CoMaPOI outperforms 11 external baselines on three fixed benchmark datasets (NYC, TKY, CA). Those numbers are compared against independent methods and are not derived from CoMaPOI's own equations, so the main claim is externally grounded. The RRF strategy (Section 4.2, Eqs. 10-12) uses the true next POI p* to construct ideal profiles, mobility patterns, and candidate sets for fine-tuning, then trains the agents to reproduce those targets from trajectory inputs. This is a standard supervised-learning construction with oracle-derived labels; it is not a case where a test-time prediction is supplied by the input it claims to predict. At inference the pipeline must generate profiles and candidate sets without access to p*, and the hit rates and posterior statistics in Table 3 are measured on those generated outputs. The theoretical analysis in Section 4.3 (Eqs. 13-17) is an algebraic rearrangement of the total-probability decomposition of the error difference; it yields a necessary condition for the candidate-set approach to help, but it is not used to manufacture the reported results. The claim that this 'theoretically proves' the method reduces error is overstated and is better read as a soundness/correctness concern than as circularity. The few self-citations (e.g., Refs. [31], [45], [60]) appear only in related-work context and are not load-bearing for the framework's design or its empirical conclusions.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The system rests on two domain assumptions: that text embeddings capture transition likelihoods for retrieval, and that RRF labels built with access to the answer can be learned without that access. The only mathematical axiom is standard probability calculus. No new physical entities are proposed.

free parameters (2)
  • Candidate set size K_C = 25
    Chosen from the performance curve in Figure 3(a) on the benchmark data; used for Forecaster retrieval and final candidate sets, and not derived from theory or a separate validation split.
  • Current trajectory length L = 30
    The last 30 check-ins are fixed as the current trajectory/test window (Section 5.1) to control token budget; this is a hand-chosen design decision that affects results.
assumptions (4)
  • domain assumption Semantic similarity in a general text-embedding space is a useful proxy for POI transition likelihood (Eq. 6).
    The Forecaster's initial retrieval assumes POIs embedded from coordinate and category descriptions will have high cosine similarity for likely next visits; the paper validates this empirically but does not establish it independently.
  • domain assumption Reverse-engineered RRF labels (Eqs. 10-11) are learnable from trajectory data without access to the true next POI at inference.
    RRF creates fine-tuning labels using p* as input; the approach assumes the Profiler and Forecaster can reproduce these labels from T only, which is the key transfer assumption highlighted in the report.
  • standard math The law of total probability and algebraic rearrangement used in Eqs. 14-17 are valid.
    This is standard probability; the issue is not correctness of the algebra but the interpretation of it as a proof of the method's benefit.
  • domain assumption Large language models can reason effectively over natural-language statistical descriptions of trajectories.
    The Profiler design assumes that converting coordinates, timestamps, and frequencies into text improves spatiotemporal reasoning; the ablation results are consistent with this but do not prove it for all LLMs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoMaPOI: A Collaborative Multi-Agent Framework for Next POI Prediction Bridging the Gap Between Trajectory and Language." pith.science (2026). https://pith.science/paper/IXSPLTL3

@misc{pith2026250523837,
  author       = {Pith},
  title        = {Pith review of: CoMaPOI: A Collaborative Multi-Agent Framework for Next POI Prediction Bridging the Gap Between Trajectory and Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IXSPLTL3}},
  note         = {Machine review of arXiv:2505.23837}
}
read the original abstract

Large Language Models (LLMs) offer new opportunities for the next Point-Of-Interest (POI) prediction task, leveraging their capabilities in semantic understanding of POI trajectories. However, previous LLM-based methods, which are superficially adapted to next POI prediction, largely overlook critical challenges associated with applying LLMs to this task. Specifically, LLMs encounter two critical challenges: (1) a lack of intrinsic understanding of numeric spatiotemporal data, which hinders accurate modeling of users' spatiotemporal distributions and preferences; and (2) an excessively large and unconstrained candidate POI space, which often results in random or irrelevant predictions. To address these issues, we propose a Collaborative Multi Agent Framework for Next POI Prediction, named CoMaPOI. Through the close interaction of three specialized agents (Profiler, Forecaster, and Predictor), CoMaPOI collaboratively addresses the two critical challenges. The Profiler agent is responsible for converting numeric data into language descriptions, enhancing semantic understanding. The Forecaster agent focuses on dynamically constraining and refining the candidate POI space. The Predictor agent integrates this information to generate high-precision predictions. Extensive experiments on three benchmark datasets (NYC, TKY, and CA) demonstrate that CoMaPOI achieves state of the art performance, improving all metrics by 5% to 10% compared to SOTA baselines. This work pioneers the investigation of challenges associated with applying LLMs to complex spatiotemporal tasks by leveraging tailored collaborative agents.

Figures

Figures reproduced from arXiv: 2505.23837 by the authors.

Figure 1
Figure 1. The overview of the CoMaPOI framework, consisting of three specialized agents (Profiler, Forecaster, and Predictor) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Performance Evaluation of CoMaPOI Across Differ [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Impact of Candidate Size and Visualization of Can [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 35 canonical work pages

  1. [1]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wang Wenjie, Fuli Feng, and Xiangnan He. 2023. Large language models for recommendation: progresses and future directions. InProceedings of the Annual International ACM SIGIR Conference on SIGIR ’25, July 13–18, 2025, Padua, Italy Lin Zhong, Lingzhi Wang, Xu Yang, and Qing Liao Research and Development in Information Re...

    work page arXiv 2023

  2. [2]

    Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2024. ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=FQepisCUWu

    2024

  3. [3]

    Jiuhai Chen and Jonas Mueller. 2024. Quantifying uncertainty in answers from any language model and enhancing their trustworthiness. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 5186–5200

    work page 2024

  4. [4]

    2024202420242024

    Florin Florin CuconasuFlorin Cuconasu, , , , Giovanni Giovanni TrappoliniGio- vanni Trappolini, , , , Federico Federico SicilianoFederico Siciliano, , , , Simone Simone FiliceSimone Filice, , , , Cesare CampagnanoCesare Campagnano, , , , Yoelle MaarekYoelle Maarek, , , , Nicola Nicola TonellottoNicola Tonellotto, , , , andand Fabrizio SilvestriFabrizio Si...

    work page

  5. [5]

    Qiang Cui, Chenrui Zhang, Yafeng Zhang, Jinpeng Wang, and Mingchen Cai. 2021. ST-PIL: Spatial-Temporal Periodic Interest Learning for Next Point-of-Interest Recommendation. InProceedings of the 30th ACM International Conference on Information & Knowledge Management. 2960–2964. https://doi.org/10.1145/ 3459637.3482189

    arXiv 2021

  6. [6]

    Yizhou Dang, Enneng Yang, Guibing Guo, Linying Jiang, Xingwei Wang, Xiaoxiao Xu, Qinghui Sun, and Hong Liu. 2023. Uniform sequence better: Time interval aware data augmentation for sequential recommendation. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 4225–4232

    2023

  7. [7]

    2023.Unsloth

    Michael Han Daniel Han and Unsloth team. 2023.Unsloth. http://github.com/ unslothai/unsloth

    2023

  8. [8]

    Shanshan Feng, Haoming Lyu, Fan Li, Zhu Sun, and Caishun Chen. 2024. Where to move next: Zero-shot generalization of llms for next poi recommendation. In 2024 IEEE Conference on Artificial Intelligence (CAI). IEEE, 1530–1535

    work page 2024

Show all 67 references
  1. [9]

    Binzong Geng, Zhaoxin Huan, Xiaolu Zhang, Yong He, Liang Zhang, Fajie Yuan, Jun Zhou, and Linjian Mo. 2024. Breaking the length barrier: Llm-enhanced CTR prediction in long textual user behaviors. InProceedings of the 47th International ACM SIGIR Conference on Research and Dev...

    2024

  2. [10]

    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. InInternational Conference on Learning Representations. https: //openreview.net/forum?id=nZeVKeeFYf9

    2022

  3. [11]

    Liwei Huang, Yutao Ma, Shibo Wang, and Yanbo Liu. 2019. An attention-based spatiotemporal lstm network for next poi recommendation.IEEE Transactions on Services Computing14, 6 (2019), 1585–1597

    2019

  4. [12]

    Tianhao Huang, Xuan Pan, Xiangrui Cai, Ying Zhang, and Xiaojie Yuan. 2024. Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 8535–8543

    2024

  5. [13]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thoma...

    2023 arXiv

  6. [14]

    Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. InICLR. https://openreview.net/forum?id=Unb5CVPtae

    2024

  7. [15]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

    2018

  8. [16]

    Hyebin Kim and Sugie Lee. 2024. POI GPT: Extracting POI information from social media text data.The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences48 (2024), 113–118

    2024

  9. [17]

    Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large language models meet collaborative filtering: An efficient all-round llm-based recommender system. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Da...

    2024

  10. [18]

    Máté Kolat, Tamás Tettamanti, Tamás Bécsi, and Domokos Esztergár-Kiss. 2023. On the relationship between the activity at point of interests and road traffic. Communications in Transportation Research3 (2023), 100102. https://doi.org/10. 1016/j.commtr.2023.100102

    2023

  11. [19]

    Siqi Lai, Zhao Xu, Weijia Zhang, Hao Liu, and Hui Xiong. 2023. Large Lan- guage Models as Traffic Signal Control Agents: Capacity and Opportunity.CoRR abs/2312.16044 (2023). https://doi.org/10.48550/arXiv.2312.16044

    2023 doi arXiv

  12. [20]

    Yantong Lai, Yijun Su, Lingwei Wei, Gaode Chen, Tianci Wang, and Daren Zha

  13. [21]

    Yantong Lai, Yijun Su, Lingwei Wei, Tianqi He, Haitao Wang, Gaode Chen, Daren Zha, Qiang Liu, and Xingxing Wang. 2024. Disentangled contrastive hypergraph learning for next POI recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research and Develop...

    2024

  14. [22]

    Peibo Li, Maarten de Rijke, Hao Xue, Shuang Ao, Yang Song, and Flora D Salim

  15. [23]

    Zihao Li, Aixin Sun, and Chenliang Li. 2023. Diffurec: A diffusion model for sequential recommendation.ACM Transactions on Information Systems42, 3 (2023), 1–28

    2023

  16. [24]

    Bo Liu, Jun Zeng, Junhao Wen, Min Gao, and Wei Zhou. 2024. CBRec: A causal way balancing multidimensional attraction effect in POI recommendations. Knowledge-Based Systems305 (2024), 112607

    2024

  17. [25]

    Lei Liu, Shuo Yu, Runze Wang, Zhenxun Ma, and Yanming Shen. 2024. How can large language models understand spatial-temporal data?arXiv preprint arXiv:2401.14192(2024)

    2024 arXiv

  18. [26]

    Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. 2025. St-llm: Large language models are effective temporal learners. InEuropean Conference on Computer Vision. Springer, 1–18

    2025

  19. [27]

    Yuanjie Lyu, Zhiyu Li, Simin Niu, Feiyu Xiong, Bo Tang, Wenjin Wang, Hao Wu, Huanyong Liu, Tong Xu, and Enhong Chen. 2024. Crud-rag: A comprehensive chinese benchmark for retrieval-augmented generation of large language models. ACM Transactions on Information Systems(2024)

    2024

  20. [28]

    Nishanth Nakshatri, Siyi Liu, Sihao Chen, Dan Roth, Dan Goldwasser, and Daniel Hopkins. 2023. Using LLM for improving key event discovery: Temporal-guided news stream clustering with event summaries. InFindings of the Association for Computational Linguistics: EMNLP 2023. 4162–4173

    2023

  21. [29]

    NetEase Youdao

    Inc. NetEase Youdao. 2023. BCEmbedding: Bilingual and crosslingual embedding for RAG. https://github.com/netease-youdao/BCEmbedding

    2023

  22. [30]

    Zijie Pan, Yushan Jiang, Sahil Garg, Anderson Schneider, Yuriy Nevmyvaka, and Dongjin Song. 2024. S2IP-LLM: Semantic space informed prompt learning with LLM for time series forecasting.arXiv preprint arXiv:2403.05798(2024)

    2024 arXiv

  23. [31]

    Jianyang Qin, Yan Jia, Binxing Fang, and Qing Liao. 2025. TaylorS: A Multi-Order Expansion Structure for Urban Spatio-Temporal Forecasting.IEEE Transactions on Knowledge and Data Engineering(2025)

    2025

  24. [32]

    Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. 2022. Contrastive learning for representation degeneration problem in sequential recommendation. InProceedings of the fifteenth ACM international conference on web search and data mining. 813–823

    2022

  25. [33]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  26. [34]

    Pablo Sánchez and Alejandro Bellogín. 2022. Point-of-Interest Recommender Sys- tems Based on Location-Based Social Networks: A Survey from an Experimental Perspective. 54, 11s, Article 223 (2022), 37 pages. https://doi.org/10.1145/3510409

    2022 doi

  27. [35]

    Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi, Marcin Cho- chowski, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro, Jan Kautz, and Pavlo Molchanov. 2024. Llm pruning and distillation in practice: The minitron approach.arXiv preprint arXiv:2408.11796(2024)

    2024 arXiv

  28. [36]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  29. [37]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. Idgenrec: Llm-recsys alignment with textual id learning. InProceed- ings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364

    2024

  30. [38]

    Hongjin Tao, Jun Zeng, Ziwei Wang, Min Gao, and Junhao Wen. 2023. Next POI recommendation based on spatial and temporal disentanglement representation. In2023 IEEE International Conference on Web Services (ICWS). IEEE, 84–90

    2023

  31. [39]

    Alicia Y Tsai, Adam Kraft, Long Jin, Chenwei Cai, Anahita Hosseini, Taibai Xu, Zemin Zhang, Lichan Hong, Ed H Chi, and Xinyang Yi. 2024. Leveraging LLM Reasoning Enhances Personalized Recommender Systems.arXiv preprint arXiv:2408.00802(2024)

    2024 arXiv

  32. [40]

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec

  33. [41]

    Kuo Wang, LingBo Liu, Yang Liu, GuanBin Li, Fan Zhou, and Liang Lin. 2023. Urban regional function guided traffic flow prediction.Information Sciences634 CoMaPOI: A Collaborative Multi-Agent Framework for Next POI Prediction Bridging the Gap Between Trajectory and Language SIG...

    2023 doi

  34. [42]

    Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, and Yangqiu Song. 2024. Rethinking the Bounds of LLM Reasoning: Are Multi-Agent Discussions the Key? arXiv preprint arXiv:2402.18272(2024)

    2024 arXiv

  35. [43]

    Tianxing Wang and Can Wang. 2024. Embracing LLMs for Point-of-Interest Recommendations.IEEE Intelligent Systems39, 1 (2024), 56–59

    2024

  36. [44]

    Xinfeng Wang, Jin Cui, Yoshimi Suzuki, and Fumiyo Fukumoto. 2024. RDRec: Rationale distillation for LLM-based recommendation.arXiv preprint arXiv:2405.10587(2024)

    2024 arXiv

  37. [45]

    Ziwei Wang, Jun Zeng, Lin Zhong, Ling Liu, Min Gao, and Junhao Wen. 2024. DSDRec: Next POI recommendation using deep semantic extraction and diffusion model.Information Sciences(2024), 121004

    2024

  38. [46]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Jun- feng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large language models with graph augmentation for recommendation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 806–815

    2024

  39. [47]

    Xiaodong Yan, Tengwei Song, Yifeng Jiao, Jianshan He, Jiaotuan Wang, Ruopeng Li, and Wei Chu. 2023. Spatio-temporal hypergraph learning for next POI rec- ommendation. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrie...

    2023

  40. [48]

    Dingqi Yang, Daqing Zhang, Vincent W Zheng, and Zhiyong Yu. 2014. Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs.IEEE Transactions on Systems, Man, and Cybernetics: Systems45, 1 (2014), 129–142

    2014

  41. [49]

    Song Yang, Jiamou Liu, and Kaiqi Zhao. 2022. GETNext: trajectory flow map enhanced transformer for next POI recommendation. InProceedings of the 45th International ACM SIGIR Conference on research and development in information retrieval. 1144–1153

    2022

  42. [50]

    Yuxuan Yang, Siyuan Zhou, He Weng, Dongjing Wang, Xin Zhang, Dongjin Yu, and Shuiguang Deng. 2024. Siamese learning based on graph differential equation for next-POI recommendation.Applied Soft Computing150 (2024), 111086

    2024

  43. [51]

    Yaowen Ye, Lianghao Xia, and Chao Huang. 2023. Graph masked autoencoder for sequential recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 321–330

    2023

  44. [52]

    Feiyu Yin, Yong Liu, Zhiqi Shen, Lisi Chen, Shuo Shang, and Peng Han. 2023. Next POI recommendation with dynamic graph and explicit dependency. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 4827–4834

    2023

  45. [53]

    Xinli Yu, Zheng Chen, and Yanbin Lu. 2023. Harnessing LLMs for temporal data-a study on explainable financial time series forecasting. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track. 739–753

    2023

  46. [54]

    Chenhan Yuan, Qianqian Xie, Jimin Huang, and Sophia Ananiadou. 2024. Back to the future: Towards explainable temporal reasoning with large language models. InProceedings of the ACM on Web Conference 2024. 1963–1974

    2024

  47. [55]

    Quan Yuan, Gao Cong, Zongyang Ma, Aixin Sun, and Nadia Magnenat Thalmann

  48. [56]

    Jun Zeng, Hongjin Tao, Haoran Tang, Junhao Wen, and Min Gao. 2025. Global and local hypergraph learning method with semantic enhancement for POI recommendation.Information Processing & Management62, 1 (2025), 103868

    2025

  49. [57]

    Jun Zeng, Yizhu Zhao, Ziwei Wang, Hongjin Tao, Min Gao, and Junhao Wen

  50. [58]

    Pengpeng Zhao, Anjing Luo, Yanchi Liu, Jiajie Xu, Zhixu Li, Fuzhen Zhuang, Victor S Sheng, and Xiaofang Zhou. 2020. Where to go next: A spatio-temporal gated network for next poi recommendation.IEEE Transactions on Knowledge and Data Engineering34, 5 (2020), 2512–2524

    2020

  51. [59]

    Yuyue Zhao, Jiancan Wu, Xiang Wang, Wei Tang, Dingxian Wang, and Maarten de Rijke. 2024. Let me do it for you: Towards llm empowered recommendation via tool learning. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrie...

    2024

  52. [60]

    Lin Zhong, Jun Zeng, Ziwei Wang, Wei Zhou, and Junhao Wen. 2024. SCFL: Spatio-temporal consistency federated learning for next POI recommendation. Information Processing & Management61, 6 (2024), 103852

    2024

  53. [61]

    Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. 2023. Toolqa: A dataset for llm question answering with external tools.Advances in Neural Information Processing Systems36 (2023), 50117–50143

    2023

  54. [63]

    LGSA: A next POI prediction method by using local and global interest with spatiotemporal awareness.Expert Systems with Applications227 (2023), 120291

    2023

  55. [2013]

    InProceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval

    Time-aware point-of-interest recommendation. InProceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval. 363–372

  56. [2019]

    InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management(Beijing, China)(CIKM ’19)

    BERT4Rec: Sequential recommendation with bidirectional encoder repre- sentations from transformer. InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management(Beijing, China)(CIKM ’19). ACM, New York, NY, USA, 1441–1450. https://doi.org/10....

  57. [2020]

    https://github.com/huggingface/ trl

    TRL: Transformer reinforcement learning. https://github.com/huggingface/ trl

  58. [2023]

    InInternational Conference on Database Systems for Advanced Applications

    Multi-view spatial-temporal enhanced hypergraph network for next poi recommendation. InInternational Conference on Database Systems for Advanced Applications. Springer, 237–252

  59. [2024]

    In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Large language models for next point-of-interest recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1463–1472

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.