Pith. sign in

REVIEW 4 major objections 5 minor 70 references

Think2Go shows that an LLM trained to reason and answer jointly predicts the next place a user will visit more accurately than prior LLM recommenders, and understands semantic location IDs better.

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

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

T0 review · deepseek-v4-flash

2026-08-03 15:41 UTC pith:ERPIIHWO

load-bearing objection Plausible gains from RL reasoning for next-POI, but the paper never says what happens to the self-correct bridge at inference, and the advantage math doesn't cohere; worth refereeing after those are fixed. the 4 major comments →

arxiv 2607.28997 v1 pith:ERPIIHWO submitted 2026-07-31 cs.IR

Think2Go: Generative Next POI Recommendation with LLM Reasoning

Zhuang Zhuang , Shanshan Feng , Hangwei Qian , Mingqi Yang , Heng Qi , Yanming Shen , Baocai Yin This is my paper
classification cs.IR
keywords next POI recommendationLLM reasoningsemantic IDsreinforcement learningadvantage calibrationepistemic uncertaintykernel density estimationgenerative recommendation
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.

Think2Go is a generative next-POI recommender that trains a large language model to think before answering: it produces a reasoning trace, then a semantic-ID prediction, and both are optimized in one loss. The paper's central claim is that this reasoning loop, with two calibrated advantage-weighting mechanisms, lets the model retain users' habitual patterns while exploring new spatial-temporal options better than prior LLM recommenders that only fine-tuned on semantic IDs. If true, the practical payoff is that LLM-based recommenders no longer have to choose between memorization and reasoning; the same model gets both, and it also generalizes better to unseen regions and to semantic-ID category understanding.

Core claim

The paper's central discovery is that a next-POI recommender can be built as a large reasoning model: Think2Go generates a Think step, a Go step containing the candidate semantic ID, then a self-correct bridge and the final answer, with the reasoning tokens trained by policy gradients and the answer tokens by supervised fine-tuning in one combined loss. On three real-world check-in datasets, the model reports Acc@1 of 0.3847, 0.3239, and 0.2583, outperforming all baselines; relative to the strongest prior LLM baseline, its Acc@1 improves by 6.33%, 5.78%, and 7.49%, respectively. Ablation studies attribute the gains to the reasoning tokens, the spatial-temporal epistemic-uncertainty weighting

What carries the argument

The central mechanism is the calibrated advantage in the Calibrated Advantage Policy Optimization (CAPO) objective, L = L_SFT + alpha * L_CAPO. The Think-to-Go tokens are scored by a rule-based reward, and the advantage is the centered reward multiplied by three calibrated coefficients: spatial-temporal epistemic uncertainty (STEU), which uses kernel density estimates to assess how well the query time and location fit the user's history, giving larger updates to uncertain prompts; a difficulty-aware reward gap (DRG), which rescales rewards by their distance from the group maximum so easy and hard groups with identical centered distributions no longer receive identical updates; and a token-le

Load-bearing premise

The load-bearing premise is that the self-correct bridge text, which appears during training but not in the inference prompt, is a harmless training-only anchor rather than a cue the model exploits; if the model exploits it, the evaluated inference does not match training and the reported gains could be an artifact.

What would settle it

Train the same model with the self-correct bridge removed from the objective, or alternatively add it at inference, and compare Acc@1 on the three datasets; if the gap to the strongest LLM baseline shrinks to near zero, the Think2Go gain is a train/inference mismatch artifact. A second check is to rerun the zero-shot semantic-ID category probe with and without the bridge to see whether the semantic-understanding gain survives.

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

If this is right

  • If the central claim holds, LLM-based next-POI recommenders can gain roughly 6-7.5 percent in Acc@1 over the strongest prior LLM baseline by adding a reasoning loop, without handcrafted feature engineering.
  • Reasoning training improves the model's grasp of semantic IDs: a zero-shot category-prediction probe jumps from 12.41 percent to 30.24 percent accuracy, suggesting the IDs carry learnable meaning that RL brings out.
  • The two calibrations prevent entropy collapse during training, so exploration persists; the reported curves show higher entropy, reward, and accuracy throughout than a GRPO-style baseline.
  • Unified SFT plus RL with compact semantic IDs is more efficient than text-prompt reinforcement fine-tuning, with roughly 56 percent less training time, 64 percent less test time, and 64 percent less memory on one dataset.
  • The method transfers better across cities: out-of-domain experiments show smaller degradation than prior LLM recommenders, pointing to improved generalization to unseen POIs and regions.

Where Pith is reading between the lines

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

  • Editorial inference: STEU and DRG form a generic recipe for sparse generative recommendation: they shape policy updates by prompt difficulty and model competence, so the same advantage-calibration idea could transfer to next-item or next-product prediction tasks.
  • Editorial inference: The self-correct bridge is the paper's fragile point: it appears during training but not in the inference prompt, so part of the gain could be a train/inference mismatch unless the model truly learns to produce final answers independently of that cue.
  • Editorial inference: The zero-shot semantic-ID probe suggests a testable path: semantic IDs could serve as a shared vocabulary for transferring mobility knowledge across cities, if the ID quantization is aligned between source and target regions.
  • Editorial inference: The experiments use an 8-billion-parameter base model with low-rank adapters, so the natural next question is whether the unified objective scales to larger models or compresses to smaller edge-deployable ones, which the paper does not address.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Think2Go, a generative next-POI recommendation framework that jointly optimizes supervised fine-tuning (SFT) and reinforcement-learning-based reasoning through a calibrated policy optimization objective (CAPO). The model generates a Think→Go reasoning trace, followed by a Self-Correct bridge and Answer tokens; SFT is applied to Answer tokens and policy-gradient RL to reasoning tokens. Two advantage-calibration mechanisms are introduced: Spatial-Temporal Epistemic Uncertainty (STEU), estimated via kernel density methods, and Difficulty-aware Reward Gap (DRG). Experiments on NYC, TKY, and CA report consistent Acc@1/MRR gains over traditional and LLM baselines; additional analyses address SID-category understanding, cross-domain transfer, and training efficiency.

Significance. If the reported gains are reproducible, the paper makes a useful contribution by showing that integrated reasoning and SFT can improve generative next-POI recommendation and semantic-ID understanding, with code availability and a reasonably complete set of ablations. The cross-domain and SID-classification experiments are constructive extensions. However, the central empirical claim currently rests on an incompletely specified training/inference protocol and on an advantage formula that is not internally consistent, so the significance cannot be fully assessed without revision.

major comments (4)
  1. [§4.1, Appendix C.1, Tables 7 and 9] The unified objective trains Answer tokens after a 'Self-correct' bridge whose content ('Unfortunately/Fortunately...') depends on ground-truth correctness. The inference protocol in Table 7 and the case study in Table 9 contain no such bridge, and no mechanism for choosing it at inference is specified. The appendix's assertion that training and inference must remain consistent is not operationalized. Since the SFT loss is computed on P(answer | prompt, think, go, correctness-cue), evaluating without the cue is a train/inference mismatch. Please either include the bridge in inference (with a stated selection rule), retrain without it, or provide an ablation demonstrating that the answer distribution is insensitive to the bridge.
  2. [§4.2.1, Eq. (5), Eq. (10)] The CAPO objective is not coherent as written. Eq. (5) defines the advantage as \hat{A}_i = c_i^t (A_i + e_i^t), with c_i^t never defined and e_i^t the token entropy. Eq. (10) then defines the final advantage as \hat{A}_i^t = exp(γ p m) A_i^t, with no entropy term and no c_i^t. The DAPO loss in Eq. (1) uses the Eq. (10) version. The token-level entropy bonus, which is ablated in Table 3, therefore never appears in the formal objective. Please define c_i^t, reconcile the two equations, and state explicitly where the entropy bonus enters the loss.
  3. [§5.4, Tables 2–6] All performance numbers are point estimates; no standard deviations, confidence intervals, or significance tests are reported. The headline improvements over GNPR are 6.33%, 5.78%, and 7.49% on Acc@1 from a single checkpoint. Without repeated runs or a significance test, the 'consistently outperforms' claim is not statistically supported. Additionally, Table 2 reports only Acc@1 for LLM4POI and GNPR, so the comparison on Acc@5/Acc@10/MRR is incomplete for the strongest LLM baselines.
  4. [§4.2.3–4.2.4, Table 6] STEU and DRG each involve free parameters (β, σ, γ, G) and are motivated by the failure modes they are then shown to fix on the same datasets. Only γ and G are subjected to sensitivity analysis (Table 6); β and σ are not swept. The ablations in Table 3 are helpful, but without sensitivity analysis for β and σ, and with no held-out hyperparameter selection described, the risk remains that the reported gains depend on dataset-specific tuning.
minor comments (5)
  1. [§5.2, Table 2] PLSPL appears in Table 2 but is not listed or described in the baselines in Section 5.2. Please either add its description or remove it from the table.
  2. [Table 3] The explanation for 'w/o STEU' says it 'assigns higher advantage weights to harder and less predictable targets,' which describes the effect of including STEU, not its removal. Please clarify the intended meaning.
  3. [Eq. (4)–(5)] The importance ratio r_t^{(i)}(θ) and the reward r_i are both denoted with r, which is confusing. Please use distinct symbols for the importance ratio and the reward.
  4. [Appendix E, Table 10] The memory usage reported (632GB for Refine-POI) seems inconsistent with the described hardware setup of five L40 GPUs (48GB each). Please clarify the hardware configuration used for the efficiency comparison.
  5. [§1] The claim of being 'the first generative next POI recommendation framework based on LLM reasoning' should be softened, since Refine-POI is cited as prior reasoning-based work and the novelty is more accurately in unifying SFT and RL with calibrated advantages.

Circularity Check

0 steps flagged

No significant circularity: main results are held-out empirical comparisons; CAPO is a training objective and self-citations are not load-bearing.

full rationale

The central claim (Table 2) is an empirical comparison on held-out test splits against external baselines; nothing in the derivation makes the test prediction equal to a fitted input. The CAPO advantage in Eq. (10)-(11) is a training-time loss used to update the policy, not a quantity renamed as a prediction; STEU and DRG are validation-tuned heuristics, and tuning/ablation on the same benchmarks is standard model selection rather than a construction-level reduction. The paper's self-citations [60]-[63] concern the authors' earlier POI work and are not used to justify the Think2Go architecture or the CAPO derivation. I explicitly considered the flagged train/inference issue in Sec. 4.1 vs App. C.1: the 'Self-correct' bridge ('Unfortunately (Fortunately), the predicted answer is incorrect (correct)...') conditions SFT answer tokens on a correctness cue that is unavailable at inference, and App. C.1 asserts 'training and inference stages must remain consistent' without showing how the bridge is chosen for inference. That is a validity/consistency risk, not a circularity by construction: it does not reduce the reported test metric to the paper's inputs or to a fitted parameter. Under the hard requirement to exhibit a specific Eq.-to-Eq. or parameter-to-prediction reduction, no circular step can be produced.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The paper's contribution is empirical and heuristic. It relies on inherited semantic IDs, domain assumptions about RL-based reasoning, and two ad hoc calibration mechanisms whose hyperparameters are largely unreported or tuned on the evaluation datasets.

free parameters (6)
  • gamma (γ) = 0.5
    Scaling factor for the calibration coefficient in the advantage (Eq. 10); tuned on validation, reported in Appendix B.
  • group size G = 4
    Number of RL rollouts per prompt; tuned in Appendix B (Table 6).
  • STEU combination weight beta (β) = not reported
    Linear combination of temporal and spatial uncertainty in Section 4.2.3; no value given in the paper.
  • KDE bandwidth sigma (σ) = not reported
    Bandwidth of Gaussian kernels in Eqs. 7-8; controls KDE smoothness and is never specified.
  • SFT/RL loss balance alpha (α) = not reported
    Coefficient in L = L_SFT + α L_CAPO (Section 4.1); no value given.
  • Reward thresholds and unit reward = 0.5 partial / 1.0 full; 300-token length threshold; 3/L per unit
    Progressive reward design in Section 4.2.2; hand-set constants not justified by data.
axioms (5)
  • domain assumption Semantic IDs trained by RQ-VAE encode geographic/functional semantics (inherited from GNPR).
    The entire framework assumes SIDs are meaningful to the LLM; this is taken from prior work and not re-derived.
  • domain assumption GRPO/DAPO-style RL improves LLM reasoning on structured prediction tasks.
    The motivation for using RL-based policy optimization is taken from the broader reasoning-model literature.
  • ad hoc to paper KDE of historical check-ins measures prompt-level 'epistemic uncertainty'.
    STEU assumes spatial-temporal proximity to history is a valid difficulty/uncertainty proxy; no independent validation.
  • ad hoc to paper R/R_max scaling (DRG) reflects intrinsic prediction difficulty.
    The reward-gap coefficient m = R/R_max is asserted to distinguish easy from hard instances; this is a heuristic.
  • domain assumption The 80/10/10 chronological split with last-check-in as target is a valid evaluation protocol.
    Standard in the field, but the filtering (only users/POIs in training) limits out-of-domain conclusions.
invented entities (1)
  • Self-correct token/prompt bridge no independent evidence
    purpose: Unify SFT and RL by conditioning answer tokens on a self-correction phrase during training.
    No independent evidence that this preserves performance at inference; the train/inference conditioning mismatch is not analyzed.

pith-pipeline@v1.3.0-daily-deepseek · 19032 in / 14830 out tokens · 154726 ms · 2026-08-03T15:41:24.469366+00:00 · methodology

0 comments
read the original abstract

Next Point-of-Interest (POI) recommendation task focuses on mining user behavioral preference patterns from historical check-ins to provide personalized suggestions for the next destination. Existing methods primarily rely on shallow contextual information and handcrafted feature interactions to predict the next POI. However, the inherent sparsity and complexity of user mobility patterns limit the computational capacity of non-reasoning models to capture deep intent, while large language models (LLMs) perform suboptimally because they lack a deep understanding of semantic IDs (SIDs) when SIDs are trained separately. To address these limitations, we propose Think2Go, a novel generative next POI recommendation framework, which enhances the model's comprehension of SID representations and explores diverse spatial-temporal patterns via test-time computational scaling. We unify supervised fine-tuning (SFT) and reinforcement learning (RL)-based reasoning within a single architecture, enabling joint optimization of memorization and adaptive reasoning to better retain user behavior patterns while exploring diverse user preferences. To further calibrate policy optimization in adaptive reasoning, we propose two advantage weighting mechanisms that integrate (1) prompt epistemic uncertainty, estimated via kernel density methods to assess the spatial-temporal periodic pattern alignment between queries and user history, promoting increased exploration under high epistemic uncertainty; and (2) reward-informed advantage scaling, captured by normalizing rewards against their maxima to adapt update magnitudes, thereby improving training stability and mitigating overfitting to noisy signals. This joint calibration forms an implicit curriculum learning strategy, delivering fine-grained, instance-aware policy updates that prevent entropy collapse and support robust exploration.

Figures

Figures reproduced from arXiv: 2607.28997 by Baocai Yin, Hangwei Qian, Heng Qi, Mingqi Yang, Shanshan Feng, Yanming Shen, Zhuang Zhuang.

Figure 2
Figure 2. Figure 2: The overall framework of Think2Go, which performs iterative policy optimization over the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Advantage Estimation Analysis. We compare Refine [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

70 extracted references · 20 linked inside Pith

  1. [1]

    Chen Cheng, Haiqin Yang, Michael R Lyu, and Irwin King. 2013. Where you like to go next: Successive point-of-interest recommendation.. InIJCAI, Vol. 13. 2605–2611

    2013

  2. [2]

    Eunjoon Cho, Seth A Myers, and Jure Leskovec. 2011. Friendship and mobility: user movement in location-based social networks. InProceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. 1082–1090

    2011

  3. [3]

    Kai Lai Chung. 1967. Markov chains.Springer-Verlag, New York(1967)

    1967

  4. [4]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Jie Feng, Yong Li, Chao Zhang, Funing Sun, Fanchao Meng, Ang Guo, and Depeng Jin. 2018. Deepmove: Predicting human mobility with attentional recurrent networks. InProceedings of the 2018 world wide web conference. 1459–1468

    2018

  6. [6]

    Shanshan Feng, Xutao Li, Yifeng Zeng, Gao Cong, and Yeow Meng Chee. 2015. Personalized ranking metric embedding for next new poi recommendation. In IJCAI’15 Proceedings of the 24th International Conference on Artificial Intelligence. ACM, 2069–2075

    2015

  7. [7]

    Shanshan Feng, Feiyu Meng, Lisi Chen, Shuo Shang, and Yew Soon Ong. 2024. Rotan: A rotation-based temporal attention network for time-specific next poi recommendation. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 759–770

    2024

  8. [8]

    Alex Graves. 2012. Long short-term memory.Supervised sequence labelling with recurrent neural networks(2012), 37–45

    2012

  9. [9]

    Xuegang Hao, Ming Zhang, Alex Li, Xiangyu Qian, Zhi Ma, Yanlong Zang, Shijie Yang, Zhongxuan Han, Xiaolong Ma, Jinguang Liu, et al. 2025. OxygenREC: An Instruction-Following Generative Framework for E-commerce Recommendation. arXiv preprint arXiv:2512.22386(2025)

    arXiv 2025

  10. [10]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.Neural computation9, 8 (1997), 1735–1780

    1997

  11. [11]

    Zeyu Huang, Tianhao Cheng, Zihan Qiu, Zili Wang, Yinghui Xu, Edoardo M Ponti, and Ivan Titov. 2025. Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling.arXiv preprint arXiv:2507.01679(2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Yiheng Jiang, Yuanbo Xu, Yongjian Yang, Funing Yang, Pengyang Wang, and Chaozhuo Li. 2025. Auto Encoding Neural Process for Multi-interest Recommen- dation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 11879–11887

    2025

  13. [13]

    Yiheng Jiang, Yuanbo Xu, Yongjian Yang, Funing Yang, Pengyang Wang, Chaozhuo Li, Fuzhen Zhuang, and Hui Xiong. 2024. TriMLP: A foundational MLP-like architecture for sequential recommendation.ACM Transactions on Information Systems42, 6 (2024), 1–34

    2024

  14. [14]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516(2025)

    Pith/arXiv arXiv 2025

  15. [15]

    Chen Li, Guoyan Huang, Zhu Sun, Lu Zhang, Shanshan Feng, and Guanfeng Liu

  16. [16]

    Peibo Li, Shuang Ao, Hao Xue, Yang Song, Maarten de Rijke, Johan Barthélemy, Tomasz Bednarz, and Flora D Salim. 2025. Refine-POI: Reinforcement Fine-Tuned Large Language Models for Next Point-of-Interest Recommendation.arXiv preprint arXiv:2506.21599(2025)

    arXiv 2025

  17. [17]

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

  18. [18]

    Songwei Li, Jie Feng, Jiawei Chi, Xinyuan Hu, Xiaomeng Zhao, and Fengli Xu

  19. [19]

    Siyu Li, Toan Tran, Haowen Lin, John Krumm, Cyrus Shahabi, Lingyi Zhao, Khurram Shafique, and Li Xiong. 2024. Geo-llama: Leveraging llms for human mobility trajectory generation with spatiotemporal constraints.arXiv preprint arXiv:2408.13918(2024)

    Pith/arXiv arXiv 2024

  20. [20]

    Xuchuan Li, Fei Huang, Jianrong Lv, Zhixiong Xiao, Guolong Li, and Yang Yue

  21. [21]

    arXiv preprint arXiv:2408.12832(2024)

    Limp: Large language model enhanced intent-aware mobility prediction. arXiv preprint arXiv:2408.12832(2024). KDD ’26, August 09–13, 2026, Jeju Island, Republic of Korea Zhuang Zhuang et al

    Pith/arXiv arXiv 2024

  22. [22]

    Shuai Liu, Ning Cao, Yile Chen, Yue Jiang, and Gao Cong. 2024. nextlocllm: next location prediction using LLMs.arXiv preprint arXiv:2410.09129(2024)

    arXiv 2024

  23. [23]

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025. Understanding r1-zero-like training: A critical perspective.arXiv preprint arXiv:2503.20783(2025)

    Pith/arXiv arXiv 2025

  24. [24]

    Be more real: Travel diary generation using llm agents and individual profiles.arXiv preprint arXiv:2407.18932(2024)

    Pith/arXiv arXiv 2024

  25. [25]

    Mingyang Liu, Gabriele Farina, and Asuman Ozdaglar. 2025. UFT: Unifying Supervised and Reinforcement Fine-Tuning.arXiv preprint arXiv:2505.16984 (2025)

    arXiv 2025

  26. [26]

    Jing Long, Liang Qu, Guanhua Ye, Tong Chen, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2024. Unleashing the power of large language models for group poi recommendations.arXiv preprint arXiv:2411.13415(2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Yan Luo, Haoyi Duan, Ye Liu, and Fu-Lai Chung. 2023. Timestamps as prompts for geography-aware location recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1697–1706

    2023

  28. [28]

    Zhao Liu, Wei Liu, Huajie Zhu, Jianxing Yu, Jian Yin, Wang-Chien Lee, and Shun Wang. 2025. Geography-Aware Large Language Models for Next POI Recommendation.arXiv preprint arXiv:2505.13526(2025)

    Pith/arXiv arXiv 2025

  29. [29]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al

  30. [30]

    Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

    arXiv 2025

  31. [31]

    2024.Learning to Reason with LLMs

    OpenAI. 2024.Learning to Reason with LLMs. https://openai.com/index/learning- to-reason-with-llms/ Accessed: 2025-07-31

    2024

  32. [32]

    Hongliang Qiao, Shanshan Feng, Xutao Li, Huiwei Lin, Han Hu, Wei Wei, and Yunming Ye. 2023. RotDiff: A hyperbolic rotation representation model for information diffusion prediction. InProceedings of the 32nd ACM international conference on information and knowledge management. 2065–2074

    2023

  33. [33]

    Yingtao Luo, Qiang Liu, and Zhaocheng Liu. 2021. Stan: Spatio-temporal attention network for next location recommendation. InProceedings of the web conference

    2021

  34. [34]

    Xiaoting Lyu, Yufei Han, Wei Wang, Hangwei Qian, Ivor Tsang, and Xiangliang Zhang. 2024. Cross-context backdoor attacks against graph prompt learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2094–2105

    2024

  35. [35]

    Wesley Mathew, Ruben Raposo, and Bruno Martins. 2012. Predicting future locations with hidden Markov models. InProceedings of the 2012 ACM conference on ubiquitous computing. 911–918

    2012

  36. [36]

    Kangrui Ruan, Xinyang Wang, and Xuan Di. 2024. From twitter to reasoner: Understand mobility travel modes and sentiment using large language models. In2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 454–459

    2024

  37. [37]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    Pith/arXiv arXiv 2024

  38. [38]

    Hongliang Qiao, Shanshan Feng, Min Zhou, WenTao Li, and Fan Li. 2025. Hyper- bolic Multi-semantic Transition for Next POI Recommendation. InCompanion Proceedings of the ACM on Web Conference 2025. 1830–1837

    2025

  39. [39]

    Xuan Rao, Lisi Chen, Yong Liu, Shuo Shang, Bin Yao, and Peng Han. 2022. Graph- flashback network for next location recommendation. InProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 1463–1471

    2022

  40. [40]

    Xuan Rao, Shuo Shang, Lisi Chen, Renhe Jiang, and Peng Han. 2025. Disentangled and personalized representation learning for next point-of-interest recommenda- tion. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. 7697–7705

    2025

  41. [41]

    Zhenglin Wan, Anjun Gao, Xingrui Yu, Pingfu Chao, Jun Song, and Maohao Ran

  42. [42]

    Dongsheng Wang, Yuxi Huang, Shen Gao, Yifan Wang, Chengrui Huang, and Shuo Shang. 2025. Generative Next POI Recommendation with Semantic ID. arXiv preprint arXiv:2506.01375(2025)

    Pith/arXiv arXiv 2025

  43. [43]

    Zhengke Sun, Hangwei Qian, and Ivor Tsang. 2025. Exploring the Effectiveness and Interpretability of Texts in LLM-based Time Series Models.arXiv preprint arXiv:2504.08808(2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Yihong Tang, Zhaokai Wang, Ao Qu, Yihao Yan, Zhaofeng Wu, Dingyi Zhuang, Jushi Kai, Kebing Hou, Xiaotong Guo, Han Zheng, et al. 2024. ItiNera: Integrating spatial optimization with large language models for open-domain urban itinerary planning.arXiv preprint arXiv:2402.07204(2024)

    Pith/arXiv arXiv 2024

  45. [45]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  46. [46]

    Xinfeng Wang, Fumiyo Fukumoto, Jin Cui, Yoshimi Suzuki, Jiyi Li, and Dongjin Yu. 2023. Eedn: Enhanced encoder-decoder network with local and global context learning for poi recommendation. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 383–392

    2023

  47. [47]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    POI Recommendation via Multi-Objective Adversarial Imitation Learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 12676– 12684

  48. [48]

    Wilson Wongso, Hao Xue, and Flora D Salim. 2024. Genup: Generative user profilers as in-context learners for next poi recommender systems.arXiv preprint arXiv:2410.20643(2024)

    Pith/arXiv arXiv 2024

  49. [49]

    En Wang, Yiheng Jiang, Yuanbo Xu, Liang Wang, and Yongjian Yang. 2022. Spatial- temporal interval aware sequential POI recommendation. In2022 IEEE 38th international conference on data engineering (ICDE). IEEE, 2086–2098

    2022

  50. [50]

    Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, et al. 2025. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reasoning.arXiv preprint arXiv:2506.01939(2025)

    Pith/arXiv arXiv 2025

  51. [51]

    Xinglei Wang, Meng Fang, Zichao Zeng, and Tao Cheng. 2023. Where would i go next? large language models as human mobility predictors.arXiv preprint arXiv:2308.15197(2023)

    Pith/arXiv arXiv 2023

  52. [52]

    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

  53. [53]

    Zhipeng Wei, Kuo Cai, Junda She, Jie Chen, Minghao Chen, Yang Zeng, Qiang Luo, Wencong Zeng, Ruiming Tang, Kun Gai, et al. 2025. Oneloc: Geo-aware generative recommender systems for local life service.arXiv preprint arXiv:2508.14646(2025)

    Pith/arXiv arXiv 2025

  54. [54]

    Runyang You, Yongqi Li, Xinyu Lin, Xin Zhang, Wenjie Wang, Wenjie Li, and Liqiang Nie. 2025. Towards Large Recommender Models with Reasoning.arXiv preprint arXiv:2505.16994(2025)

    arXiv 2025

  55. [55]

    Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024. Travelplanner: A benchmark for real-world planning with language agents.arXiv preprint arXiv:2402.01622(2024)

    Pith/arXiv arXiv 2024

  56. [56]

    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 retrieval. 403–412

    2023

  57. [57]

    Dingqi Yang, Benjamin Fankhauser, Paolo Rosso, and Philippe Cudre-Mauroux

  58. [58]

    Keyu Zhao, Fengli Xu, and Yong Li. 2025. Reason-to-Recommend: Using Interaction-of-Thought Reasoning to Enhance LLM Recommendation.arXiv preprint arXiv:2506.05069(2025)

    Pith/arXiv arXiv 2025

  59. [59]

    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

  60. [60]

    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

  61. [61]

    Zhuang Zhuang, Lu Wang, Zhenliang Hao, Heng Qi, Yanming Shen, and Baocai Yin. 2021. Traffic knowledge graph based trajectory destination prediction. In CCF Conference on Big Data. Springer, 194–215

    2021

  62. [62]

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. 2025. Dapo: An open- source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476 (2025)

    Pith/arXiv arXiv 2025

  63. [63]

    Xie Yu, Jingyuan Wang, Yifan Yang, Qian Huang, and Ke Qu. 2025. BIGCity: A universal spatiotemporal model for unified trajectory and traffic state data analysis. In2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE, 4455–4469

    2025

  64. [64]

    Zhaoqi Zhang, Haolei Pei, Jun Guo, Tianyu Wang, Yufei Feng, Hui Sun, Shaowei Liu, and Aixin Sun. 2025. OneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial Recommender.arXiv preprint arXiv:2510.26104(2025)

    arXiv 2025

  65. [67]

    Zhuang Zhuang, Lingbo Liu, Heng Qi, Yanming Shen, and Baocai Yin. 2024. CMAAN: Cross-Modal Aggregation Attention Network for Next POI Recommen- dation.IEEE Transactions on Computational Social Systems(2024)

    2024

  66. [69]

    Zhuang Zhuang, Tianxin Wei, Lingbo Liu, Heng Qi, Yanming Shen, and Baocai Yin. 2024. TAU: trajectory data augmentation with uncertainty for next POI recommendation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 22565–22573

    2024

  67. [70]

    Zhuang Zhuang, Haitao Yuan, Shanshan Feng, Heng Qi, Yanming Shen, and Baocai Yin. 2025. MGSTDN: Multi-Granularity Spatial-Temporal Diffusion Net- work for Next POI Recommendation. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 4560–4570. Think2Go: Generative Next POI Recommendation with LLM Reasoning KDD ’2...

    2025

  68. [2020]

    InPro- ceedings of the twenty-ninth international joint conference on artificial intelligence

    Location prediction over sparse user mobility traces using rnns. InPro- ceedings of the twenty-ninth international joint conference on artificial intelligence. 2184–2190

  69. [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

  70. [2025]

    PCDe: A personalized conversational debiasing framework for next POI recommendation with uncertain check-ins.Neural Networks188 (2025), 107443

    2025