Dynamic Cross-Modal Alignment for Robust Semantic Location Prediction

Amirul Rahman; Liu Jing

REVIEW 4 major objections 5 minor 14 references

Dynamic Cross-Modal Alignment for Robust Semantic Location Prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read CoVLA: a two-stage vision-language alignment that lifts semantic location prediction to 87.5% accuracy, 2.3 points above the previous best.

desk verdict An unfinished draft: the context graph never appears in the equations, the promised robustness analysis is missing, and the SOTA claim is unverifiable without code or error bars. read the letter →

arxiv 2412.09870 v1 pith:NI7KW3HL submitted 2024-12-13 cs.CV

Liu Jing , Amirul Rahman This is my paper · ORCID

classification cs.CV

keywords semanticlocationpredictionmultimodallearningvision-languagemodelscross-modalalignmentfusionsocialmediacontextualambiguityattentionmechanism

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 tries to establish that semantic location prediction from social media posts improves when a vision-language model is adapted in two stages: first align image regions with relevant text tokens, then weight text and image contributions dynamically. It reports that its CoVLA framework reaches 87.5% accuracy and 87.3% F1-score on a benchmark, beating the best previous method by 2.3 and 2.5 points respectively. The authors argue that the Contextual Alignment Module resolves contextual ambiguity and the Cross-modal Fusion Module resolves modality imbalance. If the result holds, it would make location-aware services more reliable without raw GPS coordinates.

What carries the argument

The carrying object is a pair of small modules wrapped around a pretrained vision-language encoder. CAM computes a cosine similarity matrix between text-token and image-region features, turns it into softmax attention weights, aggregates image features, and concatenates the result with text through a ReLU layer; CMF computes sigmoid gates from the aligned features and takes a weighted sum of text and aligned visual features, feeding the result to a softmax classifier. A knowledge-distillation loss keeps the fused representation close to the original pretrained model's features. The paper says CAM is guided by a precomputed context graph derived from social media metadata, although the graph itself is not defined or used explicitly in the module equations.

What would settle it

Train or re-implement CoVLA on the same 10,000-post dataset and its 70/15/15 split, first with the context graph removed from CAM and then with the graph built only from hashtags and tags that are stripped of any location-name tokens, and compare test accuracy against the reported 87.5%. If the graph-free model matches the 84.3% no-CAM ablation or the accuracy collapses when location words are removed from the tags, the context graph, not the attention equations, is carrying the result.

Watch

Extended reading notes

Core claim

CoVLA is a discriminative classifier that takes a post's text and image, aligns visual regions to text tokens through cosine-similarity attention (CAM), fuses the aligned visual features with text through sigmoid-gated weighting (CMF), and predicts a semantic location label with a softmax head. The central quantitative discovery is that this two-stage recipe outperforms the strongest published baseline on the benchmark dataset, with accuracy, precision, recall, and F1 all above 87%, versus 85.2% accuracy and 84.9% F1 for the best baseline. Ablations attribute roughly 3.2 accuracy points to CAM and 4.0 points to CMF relative to the full model. The paper also claims that human raters judge CoVLA predictions as more accurate and more contextually coherent than the baseline's.

Load-bearing premise

The entire reported gain depends on a precomputed context graph built from social media metadata that supposedly guides the alignment module; if that graph leaks the semantic location label through hashtags or user-provided tags, or if it is not actually used, the 2.3-point improvement over the baseline is not explained by the method as described.

Editorial extensions

If this is right

If CoVLA's numbers reproduce, it sets a new state of the art on this benchmark, so semantic location prediction is no longer bottlenecked by single-pass multimodal fusion.
Because the method needs no GPS coordinates, it extends to privacy-sensitive settings where location must be inferred from text and images alone.
The 10.8 ms per-sample inference time suggests the two-module architecture can run in real time on social media streams.
The human-evaluation scores imply that predictions are not just correct on a label but contextually coherent, which matters for downstream personalized services.

Reading between the lines

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

Beyond the paper: if the context graph is truly built only from metadata, the same pretrain-then-gate recipe should transfer to other multimodal classification tasks, such as scene type, event, or activity recognition, where text-image ambiguity is common.
Beyond the paper: the CMF gating makes a testable prediction—on posts where the caption names the location but the image is generic, the text gate should dominate; visualizing the two gates on such samples would directly test the modality-discrepancy explanation.
Beyond the paper: the error analysis identifies conflicting modalities as the main failure mode, so a natural extension is to train with deliberate cross-modal disagreement (modality dropout or adversarial perturbation) to push accuracy beyond 87.5%.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this paper reports a 2.3% accuracy / 2.5% F1 gain over SG-MFT on a 10k-sample semantic location prediction benchmark, but the manuscript does not give the reader enough to verify that number. The context graph mentioned in the introduction as guiding the CAM never shows up in the method section, the robustness analysis advertised in the abstract is absent, Table 1 has no error bars or run counts, and there is no code or data. The central claim is currently unverifiable.

What is worth keeping: the task is real, the two challenges they name (contextual ambiguity and modality discrepancy) are genuine, and the overall framing of adapting an LVLM in two stages is reasonable. The CAM and CMF equations, though standard, are clearly written, and the category-wise analysis in Table 4 plus the human evaluation would be useful if the underlying numbers were trustworthy. The related work is adequately positioned against SG-MFT and other recent location-prediction papers.

The soft spots are not minor. First, the context graph is load-bearing. The intro says it is a precomputed graph derived from hashtags and user tags, and that it guides CAM. But Eqs. 2-5 are plain cosine cross-attention plus a concatenation MLP; there is no graph term, no graph loss, no construction details. So either the graph does something that is never described, or it does nothing and the main claimed novelty is not actually part of the model. Second, Section 4.5 promises robustness analysis and dataset-size studies, but only delivers category-wise, efficiency, and error analyses. The abstract's claim about robust performance under noisy conditions is simply not backed by any experiment. Third, ablations remove entire CAM or CMF modules but do not touch the context graph, and lambda in Eq. 11 is not analyzed. Fourth, no code or data means the 2.3% gap could easily be seed or split variance.

I agree with the stress-test note: the evaluation is not reproducible, and the claimed improvement cannot be attributed to the proposed mechanism. This is not a case where a solid paper has one weak section; the paper's central contribution is underspecified and the evidence for its headline result is missing. That said, the paper is clearly written and the task has applied value, so with a major rewrite that defines and actually uses the context graph, shows the robustness experiments, and releases code and data, it could become a decent applied contribution. As it stands, I would not send this to reviewers. I'd return it for major revision or reject with a clear path to resubmission.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes CoVLA, a discriminative framework for semantic location prediction from multimodal social media posts. The method combines a Contextual Alignment Module (CAM) for cross-modal feature alignment and a Cross-modal Fusion Module (CMF) for dynamic text-image fusion, trained with a hybrid loss that includes knowledge distillation. Experiments on a 10,000-sample benchmark dataset are reported, claiming state-of-the-art results (87.5% accuracy and 87.3% F1, versus 85.2% and 84.9% for the SG-MFT baseline), along with ablations, human evaluation, category-wise analysis, efficiency comparison, and an error analysis. The abstract and conclusion further claim robustness to noisy inputs and scalability with dataset size.

Significance. If the reported results are reproducible and the claimed contextual-graph guidance is genuinely part of the method, CoVLA would represent a meaningful advance for semantic location prediction from social media. The paper deserves credit for formulating the task clearly, reporting multiple evaluation angles (ablations, human evaluation, category-wise and efficiency analyses), and making falsifiable performance claims. However, the significance is substantially weakened by (i) the context graph—its central novelty—being entirely absent from the formal method description, (ii) the absence of the promised robustness and dataset-size analyses, and (iii) the lack of error bars, code, or data, which makes the proposed state-of-the-art claim unverifiable in its current form.

major comments (4)

[§1, §3.2 (Eqs. 2–5)] The precomputed context graph is presented as the key mechanism guiding CAM ('The CAM is guided by a precomputed context graph...,' Section 1), yet the entire CAM formulation in Eqs. (2)–(5) is a standard cosine cross-attention with a concatenation MLP. No graph term, graph loss, graph construction procedure, or even a graph notation appears in the method or experiments. The method as formally specified therefore does not implement the paper's stated novelty, and the reported gains cannot be causally attributed to the contextual-alignment contribution.
[§4.5 (and abstract, §5)] The abstract promises 'robustness analysis shows that CoVLA maintains high performance under noisy conditions,' and Section 4.5 announces analysis of robustness and dataset size. However, Sections 4.5.1–4.5.3 contain only category-wise performance, efficiency, and error analyses; no noisy-input experiment and no dataset-size scaling study are presented. The title's 'Robust' and the abstract's robustness claim are therefore unsupported by the reported evidence.
[Table 1 and §4.1] The central SOTA claim rests on a single table of point estimates with no error bars, no number of independent runs, no significance tests, and no description of seed or split variability. The dataset is described only as 10,000 samples with a 70/15/15 split; no source, preprocessing, class distribution, or comparison of splits with the Zhang et al. baseline is provided. Without these details, the 2.3% accuracy and 2.5% F1 gaps could easily fall within run-to-run or split variance, and the claimed improvement is not reproducible from the manuscript.
[§1 vs. §3.4 (Eqs. 9–11)] Section 1 promises a 'two-stage training process' with a contextual pretraining stage and a fine-tuning stage, but Section 3.4 describes only a single hybrid loss L = L_CE + λ L_KD with no distinction between stages, no pretraining schedule, and no separate data or objectives for the two stages. Additionally, the hyperparameter λ and the definition of h_pretrained in Eq. (10) are left unspecified. This makes the training procedure incomplete and unreproducible.

minor comments (5)

[§3.4, Eq. (10)] Please specify which layer and which pretrained LVLM produce h_pretrained; 'before adaptation' is ambiguous.
[§3–§4] The backbone LVLM used for CoVLA is never named; the paper only says 'pretrained LVLM.' Giving the exact model and feature dimensions is essential for reproducibility.
[§3.3, Eq. (6)] The attention weights α_T and α_V are computed independently with sigmoids, so they do not by construction balance the two modalities; the 'dynamic balance' claim needs clarification or a normalization step.
[Table 3] The human evaluation reports agreement across five raters but provides no inter-rater agreement metric (e.g., Cohen's kappa), making the score comparison difficult to interpret.
[References] Several references contain broken spacing in author names (e.g., 'Y ucheng Zhou,' 'Y a-Qi Yu'); this should be corrected for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported gains are empirical benchmark results, and the flagged omissions (unused context graph, missing robustness experiment) are evidence gaps, not constructional circularity.

full rationale

The paper's central claim is an empirical benchmark improvement (Table 1: CoVLA 87.5% vs SG-MFT 85.2% accuracy). Nothing in the derivation chain makes that gap follow by construction. CAM is defined by standard cosine cross-attention and a concatenation MLP (Eqs. 2-5), CMF by sigmoid-gated weighted summation (Eqs. 6-7), and the prediction is a softmax classifier head (Eq. 8). Training on a 70/15/15 split with cross-entropy and a distillation regularizer is ordinary supervised learning, not a fitted parameter renamed as a prediction. The 'precomputed context graph' mentioned in Section 1 is indeed absent from Eqs. 2-5, so its intended role is unstated; however, the paper says it is 'derived from social media metadata (e.g., hashtags, user-provided tags)', and there is no equation or passage asserting that the graph is built from the target semantic location labels. A label-leaking graph would be a validity and reproducibility concern, but the manuscript nowhere exhibits that construction, and speculation about leakage is outside the circularity test. Similarly, the abstract and conclusion claim robustness under noisy conditions, yet Sections 4.5.1-4.5.3 show category-wise, efficiency, and error analyses rather than a noise-injection experiment; that is missing support for an empirical claim, not circularity. The reference list is external prior work with no self-citation chain on which the result depends. No load-bearing step reduces by the paper's own equations to its own inputs, so the honest finding is no significant circularity.

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

The paper's central result rests on an unreleased dataset, an unnamed backbone, an undefined context graph, and a hyperparameter λ that is never reported. These are all unverified inputs.

free parameters (3)

lambda (LKD weight)
Controls the balance between cross-entropy and knowledge distillation loss (Eq. 11); value is never reported, yet it directly affects the final accuracy.
Backbone LVLM choice
The pretrained vision-language model is never named; results likely depend on this choice.
Context graph construction parameters
The context graph mentioned in Section 1 is never defined, so any parameters used to build it are free and unspecified.

assumptions (3)

domain assumption The benchmark dataset from Zhang et al. is correctly labeled and the 70/15/15 split is representative.
All reported numbers depend on this dataset; no dataset description or label quality analysis is given.
domain assumption Cross-modal attention (Eq. 2-4) and sigmoid-gated fusion (Eq. 6-7) are sufficient to capture contextual ambiguity.
The paper asserts these modules address the stated challenges but provides no theoretical or empirical justification beyond the benchmark numbers.
ad hoc to paper The precomputed context graph exists and does not leak label information.
Introduced in Section 1, never defined or used in the method equations; if it leaks labels, the improvement is circular.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dynamic Cross-Modal Alignment for Robust Semantic Location Prediction." pith.science (2026). https://pith.science/paper/NI7KW3HL

@misc{pith2026241209870,
  author       = {Pith},
  title        = {Pith review of: Dynamic Cross-Modal Alignment for Robust Semantic Location Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NI7KW3HL}},
  note         = {Machine review of arXiv:2412.09870}
}

read the original abstract

Semantic location prediction from multimodal social media posts is a critical task with applications in personalized services and human mobility analysis. This paper introduces \textit{Contextualized Vision-Language Alignment (CoVLA)}, a discriminative framework designed to address the challenges of contextual ambiguity and modality discrepancy inherent in this task. CoVLA leverages a Contextual Alignment Module (CAM) to enhance cross-modal feature alignment and a Cross-modal Fusion Module (CMF) to dynamically integrate textual and visual information. Extensive experiments on a benchmark dataset demonstrate that CoVLA significantly outperforms state-of-the-art methods, achieving improvements of 2.3\% in accuracy and 2.5\% in F1-score. Ablation studies validate the contributions of CAM and CMF, while human evaluations highlight the contextual relevance of the predictions. Additionally, robustness analysis shows that CoVLA maintains high performance under noisy conditions, making it a reliable solution for real-world applications. These results underscore the potential of CoVLA in advancing semantic location prediction research.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 5 canonical work pages

[1]

Simila rity guided multimodal fusion transformer for semantic location prediction in social med ia

Zhizhen Zhang, Ning Wang, Haojie Li, and Zhihui Wang. Simila rity guided multimodal fusion transformer for semantic location prediction in social med ia. CoRR, abs/2405.05760,

arXiv
[5]

URL https://doi.org/10.48550/arXiv.2405.10292

doi: 10.48550/ARXIV .2405.10292. URL https://doi.org/10.48550/arXiv.2405.10292. Florian Bordes, Richard Y uanzhe Pang, Anurag Ajay, Alexander C. Li, Adrien Bardes, Suzanne Petryk, Oscar Ma˜ nas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, M ark Ibrahim, Melissa Hall, Y un- yang Xiong, Jonathan Lebensold, Candace Ross, Srihari Jaya kumar, Chuan Guo, Diane ...

arXiv doi:10.48550/arxiv.2405.10292
[6]

URL https://doi.org/10.48550/arXiv.2405.17247

doi: 10.48550/ARXIV .2405.17247. URL https://doi.org/10.48550/arXiv.2405.17247. Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu , Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Ping Luo, Y u Qiao, and Jif eng Dai. Vision- llm v2: An end-to-end generalist multimodal large language model for hundreds of vision- language tasks. CoRR, abs/...

arXiv doi:10.48550/arxiv.2405.17247
[7]

URL https://doi.org/10.48550/arXiv.2406.08394

doi: 10.48550/ARXIV .2406.08394. URL https://doi.org/10.48550/arXiv.2406.08394. Y a-Qi Y u, Minghui Liao, Jiwen Zhang, and Jihao Wu. Texthawk2 : A large vision-language model excels in bilingual OCR and grounding with 16x fewer tokens. CoRR, abs/2410.05261,

arXiv doi:10.48550/arxiv.2406.08394
[8]

URLhttps://doi.org/10.48550/arXiv.2410.05261

doi: 10.48550/ARXIV .2410.05261. URLhttps://doi.org/10.48550/arXiv.2410.05261. Y ucheng Zhou, Wei Tao, and Wenqiang Zhang. Triple sequence generative adversarial nets for unsuper- vised image captioning. In ICASSP 2021-2021 IEEE International Conference on Acousti cs, Speech and Signal Processing (ICASSP) , pages 7598–7602. IEEE,

arXiv doi:10.48550/arxiv.2410.05261 2021
[9]

Sketch storytelling

Y ucheng Zhou. Sketch storytelling. In ICASSP 2022-2022 IEEE International Conference on Acousti cs, Speech and Signal Processing (ICASSP) , pages 4748–4752. IEEE,

work page 2022
[10]

Towards effective next POI prediction: Spatial and semantic augmentation wit h remote sensing data

Nan Jiang, Haitao Y uan, Jianing Si, Minxiao Chen, and Shangg uang Wang. Towards effective next POI prediction: Spatial and semantic augmentation wit h remote sensing data. In 40th IEEE International Conference on Data Engineering, ICDE 20 24, Utrecht, The Netherlands, May 13-16, 2024 , pages 5061–5074. IEEE,

work page 2024
[11]

URL https://doi.org/10.1109/ICDE60146.2024.00104

doi: 10.1109/ICDE60146.202 4.00104. URL https://doi.org/10.1109/ICDE60146.2024.00104. 8 Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. Large la nguage models are zero-shot next location predictors. CoRR, abs/2405.20962,

arXiv 2024

Show all 14 references

[12]

URL https://doi.org/10.48550/arXiv.2405.20962

doi: 10.48550/ARXIV .2405.20962. URL https://doi.org/10.48550/arXiv.2405.20962. Shuai Liu, Ning Cao, Yile Chen, Y ue Jiang, and Gao Cong. nextl ocllm: next location pre- diction using llms. CoRR, abs/2410.09129,

doi arXiv
[13]

URL https://doi.org/10.48550/arXiv.2410.09129

doi: 10.48550/ARXIV .2410.09129. URL https://doi.org/10.48550/arXiv.2410.09129. Jinze Bai, Shuai Bai, Y unfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Y ang Fan, Wenbin Ge, Y u Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609,

doi
[2021]

Y ucheng Zhou and Guodong Long

URL http://proceedings.mlr.press/v139/radford21a.html. Y ucheng Zhou and Guodong Long. Improving cross-modal align ment for text-guided image inpainting. In Proceedings of the 17th Conference of the European Chapter o f the Association for Computational Linguistics, pages 3445...

2023
[2022]

Fine-tuning large vision-language models as decision-making agents via reinforcement learning

Y uexiang Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Shengbang Ton g, Yifei Zhou, Alane Suhr, Saining Xie, Y ann LeCun, Yi Ma, and Sergey Levine. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. CoRR, abs/2405.10292,

arXiv
[2023]

Thread of thought unraveling chaotic contexts

Y ucheng Zhou, Xiubo Geng, Tao Shen, Chongyang Tao, Guodong Long, Jian-Guang Lou, and Jianbing Shen. Thread of thought unraveling chaotic contexts. arXiv preprint arXiv:2311.08734,

arXiv
[2024]

URLhttps://doi.org/10.48550/arXiv.2405.05760

doi: 10.48550/ARXIV .2405.05760. URLhttps://doi.org/10.48550/arXiv.2405.05760. Y ucheng Zhou, Zhi Rao, Jun Wan, and Jianbing Shen. Rethinkin g visual dependency in long-context reasoning for large vision-language models. arXiv preprint arXiv:2410.19732, 2024a. Y ucheng Zhou, X...

doi arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Simila rity guided multimodal fusion transformer for semantic location prediction in social med ia

Zhizhen Zhang, Ning Wang, Haojie Li, and Zhihui Wang. Simila rity guided multimodal fusion transformer for semantic location prediction in social med ia. CoRR, abs/2405.05760,

arXiv

[2] [5]

URL https://doi.org/10.48550/arXiv.2405.10292

doi: 10.48550/ARXIV .2405.10292. URL https://doi.org/10.48550/arXiv.2405.10292. Florian Bordes, Richard Y uanzhe Pang, Anurag Ajay, Alexander C. Li, Adrien Bardes, Suzanne Petryk, Oscar Ma˜ nas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, M ark Ibrahim, Melissa Hall, Y un- yang Xiong, Jonathan Lebensold, Candace Ross, Srihari Jaya kumar, Chuan Guo, Diane ...

arXiv doi:10.48550/arxiv.2405.10292

[3] [6]

URL https://doi.org/10.48550/arXiv.2405.17247

doi: 10.48550/ARXIV .2405.17247. URL https://doi.org/10.48550/arXiv.2405.17247. Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu , Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Ping Luo, Y u Qiao, and Jif eng Dai. Vision- llm v2: An end-to-end generalist multimodal large language model for hundreds of vision- language tasks. CoRR, abs/...

arXiv doi:10.48550/arxiv.2405.17247

[4] [7]

URL https://doi.org/10.48550/arXiv.2406.08394

doi: 10.48550/ARXIV .2406.08394. URL https://doi.org/10.48550/arXiv.2406.08394. Y a-Qi Y u, Minghui Liao, Jiwen Zhang, and Jihao Wu. Texthawk2 : A large vision-language model excels in bilingual OCR and grounding with 16x fewer tokens. CoRR, abs/2410.05261,

arXiv doi:10.48550/arxiv.2406.08394

[5] [8]

URLhttps://doi.org/10.48550/arXiv.2410.05261

doi: 10.48550/ARXIV .2410.05261. URLhttps://doi.org/10.48550/arXiv.2410.05261. Y ucheng Zhou, Wei Tao, and Wenqiang Zhang. Triple sequence generative adversarial nets for unsuper- vised image captioning. In ICASSP 2021-2021 IEEE International Conference on Acousti cs, Speech and Signal Processing (ICASSP) , pages 7598–7602. IEEE,

arXiv doi:10.48550/arxiv.2410.05261 2021

[6] [9]

Sketch storytelling

Y ucheng Zhou. Sketch storytelling. In ICASSP 2022-2022 IEEE International Conference on Acousti cs, Speech and Signal Processing (ICASSP) , pages 4748–4752. IEEE,

work page 2022

[7] [10]

Towards effective next POI prediction: Spatial and semantic augmentation wit h remote sensing data

Nan Jiang, Haitao Y uan, Jianing Si, Minxiao Chen, and Shangg uang Wang. Towards effective next POI prediction: Spatial and semantic augmentation wit h remote sensing data. In 40th IEEE International Conference on Data Engineering, ICDE 20 24, Utrecht, The Netherlands, May 13-16, 2024 , pages 5061–5074. IEEE,

work page 2024

[8] [11]

URL https://doi.org/10.1109/ICDE60146.2024.00104

doi: 10.1109/ICDE60146.202 4.00104. URL https://doi.org/10.1109/ICDE60146.2024.00104. 8 Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. Large la nguage models are zero-shot next location predictors. CoRR, abs/2405.20962,

arXiv 2024