NEXTLOCLLM: NEXT LOCATION PREDICTION USING LLMS

W3.1 In the ablation study, only using LLM-enhanced POI achieves the second-best result. It challenges the significance of the proposed methods.

We would like to first clarify that 'only using LLM-enhanced POI' in this ablation study is equivalent to 'not using prompt prefix'. Indeed, ' only using LLM-enhanced POI embedding' performs exceptionally well in the ablation study, second only to the complete model. However, we believe this result actually demonstrates the significance and validates the rationale and unique advantages of our approach.

We argue that one main advantage of LLMs in next location prediction lies in their semantic understanding capabilities. By analyzing POI information, LLMs can better understand the functional attributes of locations, allowing them to infer the human behavioral intentions hidden in the trajectories. This is something that traditional small models struggle to achieve. Even when using only LLM-enhanced POI embeddings, LLMs achieve high performance, which demonstrates that their understanding of regional semantics contributes significantly to the model's performance.

W3.2 In addition, it is weird that using both POI and LORA will lead to worse performance than pure POI.

As for why the introduction of LoRA results in a performance drop, this is a noteworthy issue in the field of spatiotemporal LLMs. A similar phenomenon was observed in [1]，where the study pointed out that QLoRA's low-rank updates could lead to information loss when handling complex semantic features.

For this phenomenon, we hypothesize that LLMs, trained on large-scale textual corpora, have absorbed rich language-specific knowledge, much of which goes beyond the requirements of spatiotemporal sequence analysis. In spatiotemporal prediction tasks, these language-related details may interfere with the model’s ability to capture spatiotemporal patterns effectively, resulting in performance degradation. LoRA fine-tuning, which updates the model’s parameters, might further activate or amplify these irrelevant language-specific features, exacerbating the interference and leading to diminished prediction accuracy. We hope our response answers your questions.

[1]Jin M, Wang S, Ma L, et al. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models[C]//The Twelfth International Conference on Learning Representations.

W4 This work aims to achieve generalization ability by LLM. However, I am wondering whether this ability can be achieved without road graph. It seems that this paper aim to realize zero-shot generalization with purely POI even without trajectory.

We would like to clarify that our study specifically targets the prediction of mobility with semantic patterns (e.g., stay point as locations) rather than road network-constrained vehicle trajectory prediction.

Specifically, our method leverages POI data to represent functional attributes (e.g., commercial, residential or educational areas). This representation allows for cross-city generalization, enabling the model to align functionally similar regions between cities, thereby significantly enhancing performance in zero-shot transfer tasks.

Moreover, in our method, historical and current trajectories remain the core data sources, capturing both short-term and long-term movement patterns of users. These trajectory records provide critical dynamic movement information and serve as the primary basis for predicting the next location. POI information plays a complementary role by enhancing the model’s understanding of regional characteristics, providing semantic support for functional areas. This design allows the model to analyze user movement patterns while leveraging regional functionality to make more informed predictions, articularly in cross-city scenarios.

W5 The formalization is a little bit confusing. TI={(d,t)} and t is time-of-day (0 ≤ t ≤ 23 in hours). Since Taxi data seem to be less than an hour. This formalization seems not complete

In our task, the time feature $t$ is defined as the hour of the day, with a range from 0 to 23 (i.e., time-of-day). For trajectory data with sampling intervals smaller than an hour, all trajectory records falling within the same hour are assigned the same hour-level time feature. This approach is based on the assumption that trajectory records within the same hour generally exhibit similar behavioral patterns and are not significantly affected by minute-level variations. For example, trajectory records at 10:15 and 10:45 are, in most cases, considered to share similar temporal feature. This processing method avoids overly fine-grained time segmentation, thereby reducing the complexity of learning time features while still capturing the primary temporal characteristics of different behavioral patterns.

W2 It seems that the proposed KD-Tree-based prediction contributes, but you did not involve it in the experiments.

We would like to clarify that our contribution lies beyond KD-tree. It lies in introducing a regression-based approach to replace traditional classification methods, providing a more effective solution for zero-shot next location prediction tasks. The KD-tree plays a supportive role to accelerate the post-processing of regression outputs. We will revise our manuscript to make this distinction clearer.

Traditionally, next location prediction tasks are formulated as classification problems, where each possible location ID is treated as a class, and the model predicts a probability distribution over these classes to determine the next location. However, this approach has significant limitations in zero-shot scenarios. Since location ID systems differ completely between cities, classification-based methods fail to transfer effectively.

In contrast, NextLocLLM employs a regression-based approach that directly predicts the spatial coordinates of the next location instead of discrete class labels. This design effectively addresses the limitations of traditional classification methods, removing the need for unified mappings across cities and significantly enhancing transferability. Additionally, regression predicts precise latitude and longitude coordinates, offering superior geographic resolution compared to classification, which is restricted to discrete grids or predefined locations..

To address your concern, we have further conducted experiments to test the contribution of the use of KD-tree when not considering the zero-shot scenario. Specifically, after predicting spatial coordinates (x,y), we replace the KD-tree nearest-neighbor search with a linear layer and softmax function in the output layer, redefining the task as a classification problem that outputs a probability distribution over all possible locations with cross-entropy loss.

The results show that in fully supervised scenarios, KD Tree consistently outperforms the classification-based approach across all three metrics. This reinforces the advantages of the regression-based framework in both zero-shot and fully supervised settings, demonstrating its flexibility and robustness.

In summary, our innovation lies in the regression-based design, with KD-tree enhancing its efficiency. We will revise the manuscript to clarify these points, include the supplementary experimental results, and further detail the role of KD-tree in our framework. Thank you again for your valuable feedback!

W1 why using LLaMA2 and LLaMA3 does not outperform GPT2.

This is a profound question that touches on the core of how LLMs handle spatiotemporal problems. Previous studies have also observed similar phenomena. For instance, [1] found that GPT-2 outperformed LLaMA2, attributing this to GPT-2’s parameter scale being better suited for relatively smaller datasets. Similarly, in [2], experiments demonstrated that GPT-2 achieved better results than LLaMA3 in next location prediction tasks. These findings collectively suggest that smaller-scale LLMs are often better suited for spatiotemporal prediction problems.

Our understanding of this phenomenon is as follows. LLMs contain a vast amount of knowledge from text corpus, while much of the knowledge exceeds the requirements of spatiotemporal sequence analysis. For spatiotemporal prediction tasks, the redundancy of such language-specific information can interfere with the model’s ability to effectively capture spatiotemporal patterns, thereby potentially degrading its performance. However, certain aspects of LLMs—particularly those related to human intent and semantic understanding—can significantly enhance spatiotemporal prediction tasks. This duality necessitates a balance when selecting an LLM: minimizing the interference of redundant language information while retaining the semantic understanding capabilities of LLMs to leverage their strengths in modeling human behavioral patterns.

Based on this understanding, we selected GPT-2 as the core architecture for NextLocLLM, rather than larger-scale models like LLaMA2 or LLaMA3. GPT-2’s smaller parameter size and more streamlined training corpus make it better aligned with the demands of spatiotemporal prediction tasks, while still preserving its core strengths in natural language understanding. Our experimental results support this choice: under identical experimental settings, GPT-2 outperformed LLaMA2 and LLaMA3, demonstrating its suitability for this specific task context. We will provide further related analysis in the revised manuscript. Thank you again for your valuable feedback!

[1] Cheng J, Yang C, Cai W, et al. NuwaTS: Mending Every Incomplete Time Series[J]. arXiv preprint arXiv:2405.15317, 2024.

[2] Wang X, Fang M, Zeng Z, et al. Where would i go next? large language models as human mobility predictors[J]. arXiv preprint arXiv:2308.15197, 2023.

Thank you for your response. While it addressed some of my concerns, due to the following reasons, I will maintain my score.

1. The primary motivation of the paper is to leverage large language models (LLMs) to enhance next-location prediction. However, the experimental results suggest that better-performing LLMs such as LLaMA do not necessarily lead to improved prediction outcomes. This raises questions about the suitability and effectiveness of the proposed framework for the stated task.

2. Although the paper reports state-of-the-art performance in Table 1, it is unclear whether this improvement stems from the proposed framework itself or primarily from the use of the KD tree. This distinction warrants further investigation to validate the novelty and contribution of the framework.

3. The clarity and precision of the manuscript require further refinement. For instance, in Definition 2.4, the authors state that the goal is to predict the next location $loc_{t+1}$. However $loc$ is defined as a tuple $(id,x,y,poi)$. It seems that the prediction target is the $id$ rather than the entire tuple. This ambiguity could lead to misinterpretations and should be clarified.

4. The experiments are conducted on a taxi dataset, which typically involves trips with predefined destinations. I question the necessity of including POI information in this context, as the POIs along the route are merely passed by and may not hold significant relevance to the prediction task. This design choice should be better justified, or its impact should be clearly demonstrated.

Thank you for your timely feedback. We will provide a more detailed explanation here regarding the new issues you raised. We hope this will address your concerns and enable you to reassess our work.

Additional Q1 The primary motivation of the paper is to leverage large language models (LLMs) to enhance next-location prediction. However, the experimental results suggest that better-performing LLMs such as LLaMA do not necessarily lead to improved prediction outcomes. This raises questions about the suitability and effectiveness of the proposed framework for the stated task.

Firstly, from the perspective of experimental results, we compare NextLocLLM using LLaMA2, LLaMA3, and GPT2 as its backbone against other baselines. Regardless of whether in fully-supervised or zero-shot scenarios, NextLocLLM demonstrates advantages. In fully-supervised scenarios, NextLocLLM consistently outperforms traditional small models in most cases. Moreover, no matter which LLM is used as the backbone, NextLocLLM achieves superior performance compared to other LLM-based methods in zero-shot scenarios, where traditional small models are unable to generalize across cities. This indicates that our design framework of NextLocLLM and its integration with LLMs are both effective and meaningful.

Secondly, from the perspective of theoretical analysis, while LLaMA generally outperforms GPT2 in most natural language-related tasks, this does not necessarily imply it is superior in all non-textual tasks. In fact, numerous studies have found that in certain specific non-textual tasks, LLaMA as a backbone performs worse than GPT2. For example, [1][2][3][4][5][6] have observed such phenomena. Among them, [4] suggests that this may be due to the parameter scale of LLaMA being overly redundant for the requirements of time-series tasks. Combined with our analysis in response to W1, we suppose that LLaMA contains more redundant information learned from text corpora compared to GPT2. While such information is beneficial for natural language processing tasks, it may interfere with the model's ability to effectively capture spatiotemporal patterns and trajectory semantic information in spatiotemporal tasks, thereby affecting prediction performance.

The above summarizes our understanding of this issue. We hope that the analysis from both experimental and theoretical perspectives can address your concerns.

[1] Tempo: Prompt-Based Generative Pre-Trained Transformer for Time Series Forecasting

[2] AutoTimes: Autoregressive Time Series Forecasters via Large Language Model

[3] Enhancing Graph Neural Networks with Limited Labeled Data by Actively Distilling Knowledge from Large Language Model

[4] NuwaTS: Mending Every Incomplete Time Series

[5] Where would i go next? large language models as human mobility predictors

[6] Spatial-Temporal Large Language Model for Traffic Prediction

Additional Q2 Although the paper reports state-of-the-art performance in Table 1, it is unclear whether this improvement stems from the proposed framework itself or primarily from the use of the KD tree. This distinction warrants further investigation to validate the novelty and contribution of the framework.

In addition to our explanation of the role of the KD tree in response to W2, we believe that the right-hand section of Table 13 provides further critical evidence to address your concerns. In this part of the experiment, we calculate the average distance between the predicted coordinates from NextLocLLM and the actual coordinates when using different LLMs (e.g., GPT-2, LLaMA2, and LLaMA3) as the backbone. Furthermore, we conduct an additional experiment where the input and output of C-MHSA—the best-performing small model in the fully-supervised scenario—are modified from discrete location IDs to spatial coordinates, consistent with NextLocLLM. The average distance error of its predictions is also calculated. This prediction of spatial coordinates occurs before the KD tree query and directly reflects the effectiveness of the framework itself, independent of the KD tree’s impact.

The experimental results show that regardless of the specific LLM backbone used, the average distance between the predicted and actual positions for NextLocLLM is significantly smaller than that of C-MHSA. Additionally, the prediction errors of NextLocLLM with different backbones are generally less than half the size of a location grid. This indicates that most predicted coordinates fall within the location of the ground truth, meaning the predictions are correct. This demonstrates that the performance improvements of our proposed framework are significant, and the KD tree's role is to mapping the predicted coordinates to the nearest top-k locations. If the model itself has high prediction error, even an efficient KD tree query cannot compensate for the discrepancy between the predicted coordinates and the actual positions.

Therefore, we can confidently conclude that the core source of NextLocLLM's performance improvement lies in our proposed framework design itself, rather than the KD tree. We hope this additional experiment and analysis further address your concerns!

Additional Q3 The clarity and precision of the manuscript require further refinement.

Thank you for your valuable feedback! In response to the issue you raised, we have revised Definition 2.4 to explicitly clarify that the prediction target is the ID of the next location, thereby avoiding any ambiguity that might lead to interpreting the target as the entire tuple. This revision further ensures the clarity and accuracy of the manuscript, minimizing the potential for misunderstanding. We sincerely appreciate your suggestion!

Additional Q4 The experiments are conducted on a taxi dataset, which typically involves trips with predefined destinations. I question the necessity of including POI information in this context, as the POIs along the route are merely passed by and may not hold significant relevance to the prediction task. This design choice should be better justified, or its impact should be clearly demonstrated.

On this point, we will explore the role and importance of POI information in this task from three perspectives: data processing, method design, and its contribution to cross-city tasks.

First, in terms of data preprocessing, we follow the approach adopted in [1][2] to preprocess the trajectory data of the Xi’an and Chengdu datasets, converting them into staypoints to eliminate transit trajectory points and focus on locations where users remain relatively stationary. Specifically, we set two key thresholds: a minimum stay duration of 10 minutes and a distance of more than 200 meters from the previous staypoint. For the Singapore dataset, due to its characteristics as population mobility data, we adjust the stay duration threshold to 1 minute. This preprocessing method effectively filters out transient points and retains only staypoints relevant to the prediction task. These staypoints are then mapped to specific locations and serve as input data for our next location prediction task.

Second, in terms of method design, introducing POI information aims to enhance the model's understanding of the functional attributes of locations, thereby improving its ability to capture the semantic information behind user behavior. Spatial coordinates only reflect the geographical characteristics of a location, while POI information reveals its functional properties, such as whether the location is a commercial area, a residential area, or a transportation hub. Such functional information is crucial for predicting user behavior and preferences. Even in taxi trajectory data, drivers exhibit preferences, such as whether they tend to stay in commercial areas or transportation hubs when not carrying passengers. Incorporating POI information as a supplement to spatial semantics and conducting experiments on vehicle trajectory data is not uncommon in next location prediction tasks. For instance, [1] integrated POI information using an LDA model on the Swiss Green Class Car dataset, and [3] normalized POI counts for experiments on taxi datasets in Shanghai, Chengdu, and San Francisco. These studies demonstrate the effectiveness of POI information in enhancing the model's understanding of the functional properties of vehicle trajectory locations. Compared to these methods, our proposed LLM-enhanced POI embedding further improves the understanding of POI information by capturing its deeper functional properties and providing a unified approach to modeling the semantic functions of locations. This approach not only overcomes the limitations of fixed POI taxonomy but also offers stronger generalization capabilities, particularly excelling in cross-city tasks.

Finally, in cross-city prediction scenarios, POI information is especially important for understanding the spatial semantics of locations. Solely relying on spatial coordinates for modeling has significant limitations in cross-city tasks. For example, in the Xi’an dataset, major commercial areas are concentrated in the city center, while in the Singapore dataset, commercial areas are distributed not only in the city center but also in numerous suburban satellite towns. If modeling relies only on spatial coordinates, it becomes difficult for the model to capture the similarities and differences in the functional attributes of locations in different cities. However, with the introduction of POI information, our model can better align locations with similar functional attributes across different cities, thereby improving generalization ability. To validate this, we conduct additional ablation experiments, and the results shows that removing POI information significantly reduces prediction performance, further demonstrating its critical role in this task.

[1] Context-aware multi-head self-attentional neural network model for next location prediction

[2]Where would i go next? large language models as human mobility predictors

[3] An Efficient Destination Prediction Approach Based on Future Trajectory Prediction and Transition Matrix Optimization

W1 The model parameters and configurations are completely lacking in introduction.

To address your concerns, we present the detailed parameters and configurations for NextLocLLM as mentioned in your comments (see section K):

Basic Training Parameters

• Learning Rate: 0.0002
• Batch Size: 128
• Max Epochs: 30
• Optimizer: Adam

Key Hyperparameters

• LLM Backbone: GPT-2
• Spatial Coordinate Embedding Dimension: 128
• Day & Hour Embedding Dimension: 16
• Duration Embedding Dimension: 16
• POI MLP Dimension: 1024
• LLM Layers: 6

Model Input Configuration

• Historical Trajectory Length: 40 records, capturing long-term behavioral patterns of users.
• Current Trajectory Length: 5 records, capturing short-term behavioral features.

LLM Tokenizer and Embedding Layer

We used the tokenizer and embedding layer consistent with the GPT-2 model to ensure compatibility and uniformity in input processing.

Experimental Hardware

The experiments were conducted on a set of Tesla V100-SXM2-32GB GPUs to support large-scale data training and inference.

We include these detailed parameter and configuration descriptions in the revised manuscript to enhance the clarity of the model presentation and the reproducibility of the experiments. We greatly appreciate the reviewers’ valuable suggestions.

W2 Lack of sensitivity analysis on hyperparameters. Although the authors conducted a complete ablation study demonstrating the key roles of each module, it is still necessary to conduct corresponding sensitivity analysis on key parameters to demonstrate the robustness of the model.

We thank the reviewers for their valuable suggestions. In response, we have further conducted a detailed sensitivity analysis of the key hyperparameters to validate the robustness of the model's performance (see Section Q for details). Below is a summary of the findings:

Parameters like $d_{xy}$, $d_d$, $d_t$, and $d_{poi}$​ exhibit relatively stable effects on the model’s performance. For instance, increasing the spatial coordinate embedding dimension from 64 to 256 steadily enhances performance, but further increases show saturation, indicating that smaller dimensions effectively capture geographic information. Similarly, the time and date embedding dimension shows slight improvements as it grows from 8 to 32, with minimal gains beyond this range, reflecting the robustness of these features.

On the other hand, parameters such as the historical trajectory length ($M$) and the number of LLM layers ($N_{layers​}$) have a more pronounced impact. Increasing the historical trajectory length improves the model’s ability to capture past travel patterns. Moreover, performance improves significantly as the number of layers increases, indicating that deeper architectures enhance the model’s representational capacity. However, beyond 6 layers, the performance stabilizes. Considering the trade-off between efficiency and performance, we chose to use a 6-layer GPT configuration.

This sensitivity analysis underscores the robustness of the model to most parameter settings while identifying key parameters with greater influence on performance. These results and analyses are included in the revised manuscript to provide a more comprehensive evaluation. Thank you again for your valuable feedback!

W1 Innovation Concerns: The core contribution, POI information embedding, appears somewhat incremental since it aligns with previous works' embedding logic but integrates this with LLM-based representation.

In next location prediction, understanding POI information is essential for capturing the functional attributes of locations. This helps the model gain deeper insights into regional characteristics and better capture user behavior semantics. However, traditional methods typically compute POI embeddings directly from numerical vectors representing POI distributions. This approach overlooks the rich semantic information inherent in POI categories, limiting the model's ability to fully understand the functional attributes of locations.

The limitations of traditional POI embedding methods become more pronounced in zero-shot cross-city tasks. These methods assume consistent POI taxonomy between the source and target cities [1][2]. However, in reality, there are significant differences in POI formats and taxonomy across cities. For example, in our study, the POI categories used in the datasets from Xi’an, Singapore, and Japan differ significantly, making it challenging to align POI categories strictly. This inconsistency hampers the generalization ability of traditional models in cross-city tasks, as POI embeddings derived solely from numerical vectors of POI distributions fail to resolve the semantic ambiguities caused by the inconsistency in POI taxonomy systems.

To address this challenge, we propose an LLM-enhanced POI embedding method. This method leverages the natural language understanding capabilities of LLMs to transform the natural language descriptions of POI categories into semantic embedding vectors, thereby capturing the deep functional semantics of POI categories. Unlike traditional methods, this approach does not rely on fixed POI taxonomy but instead operates within a unified semantic space, enabling effective modeling even when POI definitions and classification systems differ across cities. Experimental results show that this method effectively resolves semantic ambiguities in zero-shot scenarios and exhibits significant advantages in cross-city tasks.

[1] Jiping L, An L, Fuhao Z, et al. Hybrid Classification Method for Multi-source POIs Based on Semantic Analysis[J].

[2]Jiang R, Song X, Fan Z, et al. Transfer urban human mobility via poi embedding over multiple cities[J]. ACM Transactions on Data Science, 2021, 2(1): 1-26.

Q1 I hope the author can give an intuitive analysis of why the LLM-based model performs better.

Thanks for your suggestions. Below, we provide a detailed analysis and supporting experiments to explain the strengths of LLMs and the rationale behind our design decisions.

The primary advantage of integrating LLMs into this task lies in their ability to transcend the limitations of small models. Traditional small models typically rely on spatial and temporal patterns learned from large trajectory datasets to predict the next location. While effective in data-rich scenarios, these models often overlook the semantic nature of trajectories—the underlying human behavioral patterns reflected in the real world. In data-scarce or zero-shot scenarios, where statistical patterns alone are insufficient, trajectory semantics become critical. LLMs, with their powerful natural language understanding and reasoning capabilities, excel in capturing the inherent semantics of trajectories, even with limited data. This is the foundational motivation behind using LLMs as the core design of our framework.

To maximize the utility of LLMs, we optimized our model to enhance its semantic understanding of trajectories. First, we employed prompt prefixes to clearly define the prediction task and data, coupled with natural language descriptions of POIs to convey locations' functional attributes. This design leverages the LLM’s ability to interpret natural language and human behavioral patterns, enabling it to go beyond learning spatial and temporal patterns and uncover the deeper semantic and logical relationships within trajectories. This capability is particularly beneficial for zero-shot tasks, where such semantic understanding significantly boosts performance.

To validate our hypothesis, we conducted experiments where we replaced the LLM in our framework with a randomly initialized Transformer which shares the same amount of parameters, and trained it on Xi’an dataset. The results showed that this model performed significantly worse in zero-shot and fully-supervised tasks, with a particularly notable gap in zero-shot scenarios. This demonstrates the importance of LLM’s pre-trained knowledge and reasoning capabilities. The LLM effectively utilizes its internal semantic and logical information to complement the lack of statistical data. Even when the Transformer model was trained on a large amount of data, its performance still fell short of the LLM, further confirming the LLM’s unique strengths in zero-shot and transfer tasks.

In fully supervised scenarios with abundant data, small models may hold certain advantages, such as computational efficiency and faster training times. However, in real-world applications, data scarcity and transfer tasks are common challenges. In these scenarios, the LLM’s ability to understand and reason about the semantics of trajectories gives it a clear edge.

Looking ahead, we believe human trajectories inherently contain spatiotemporal intent and semantics. These deep characteristics offer promising directions for using LLMs in spatiotemporal data mining. For instance, LLMs can be further explored to uncover behavioral patterns, user preferences, and higher-level semantic relationships within trajectories, enabling them to address a broader range of tasks. We deeply appreciate the reviewer’s interest in this topic and look forward to further discussing the potential and future directions of LLMs in spatiotemporal data mining. Thank you for your valuable feedback!

W4 Case Studies in Experiments: The inclusion of case studies could bolster the practical applicability of the framework by demonstrating real-world scenarios.

To address your concerns, we have included a case study based on the Singapore dataset to showcase the real-world application of NextLocLLM and its significant advantages in zero-shot cross-city tasks. For details, please refer to Section P and Figure 8.

In this case study, we evaluated the zero-shot capabilities of NextLocLLM trained on the Xi'an dataset by testing it on the Singapore dataset. We also compared these results with the fully-supervised setting, where the model was both trained and tested on the Singapore dataset. The Singapore dataset was chosen due to its extensive coverage of location IDs and its distinct POI classification system, which differs from that of Xi'an. This characteristic makes our case study more representative of real-world transfer scenarios, where different cities often use varying POI classification and definition systems, posing a challenge for models to generalize across cities.

Specifically, we analyzed a user trajectory from the Singapore dataset with densely sampled points to provide an intuitive demonstration of the model's performance. The trajectory had an average sampling interval of 21 minutes. We compared the following approaches:

• Zero-shot NextLocLLM (using coordinates)
• Zero-shot NextLocLLM (using location IDs)
• Fully-supervised NextLocLLM (using coordinates)
• LLM-Mob prediction

The experimental results clearly highlight the differences in performance across these methods: Zero-shot NextLocLLM (using coordinates) exhibited outstanding predictive capability, generating trajectories highly consistent with the ground truth. This demonstrates NextLocLLM's strong zero-shot generalization in cross-city tasks. In contrast, Zero-shot NextLocLLM (using location IDs) showed significant deviations from the ground truth. This result underscores the limitations of location ID-based models in cross-city tasks, as inconsistent location ID systems across cities fail to convey the semantic information of geographic spaces. LLM-Mob prediction consistently limited its forecasting to locations explicitly mentioned in the prompts, revealing that models purely based on prompts and location IDs struggle to comprehend cross-city spatial information or capture trajectory sequence patterns. This limitation significantly reduces the applicability of LLM-Mob for tasks involving unknown locations, particularly in zero-shot cross-city scenarios.Finally, Fully-supervised NextLocLLM (using coordinates) exhibits the superior performance in a supervised setting, achieving results only marginally better than the zero-shot setting. This further validates the robust generalization capability of NextLocLLM, which maintains near-supervised performance even without target city training data.

Through this case study, we demonstrate the clear advantages of NextLocLLM in cross-city tasks, particularly its remarkable zero-shot generalization capabilities. Wel include a detailed analysis and visualizations of this case study in the revised manuscript.

W5 Minor Errors: Minor errors, such as in Equation 9, should be corrected for accuracy.

Thank you for your valuable suggestions. We have revised our manuscript and corrected the equation.

W2 Contribution of KD-tree: The KD-tree application is commonplace in traffic scenarios, particularly with GNN-based models, and thus may not qualify as a novel contribution.

We would like to clarify that our contribution lies beyond KD-tree. It lies in introducing a regression-based approach to replace traditional classification methods, providing a more effective solution for zero-shot next location prediction tasks. We will revise our manuscript to make this distinction clearer.

Traditionally, next location prediction tasks are formulated as classification problems, where each possible location ID is treated as a class, and the model predicts a probability distribution over these classes to determine the next location. However, this approach has significant limitations in zero-shot scenarios. Since location ID systems differ completely between cities, classification-based methods fail to transfer effectively. Additionally, classification inherently relies on predefined discrete categories, making it incapable of predicting unseen locations.

In contrast, NextLocLLM employs a regression-based approach that directly predicts the spatial coordinates of the next location instead of discrete class labels. This design effectively addresses the limitations of traditional classification methods. The regression approach leverages the continuous nature of spatial coordinates, enabling the prediction of unseen geographic locations and achieving better generalization by learning geographic spatial relationships. On the other hand, classification methods are restricted to predefined categories and fail to generalize in new scenarios. Furthermore, the regression approach eliminates the dependence on location IDs, removing the need for unified mappings across cities and significantly enhancing the applicability of the model in cross-city tasks. Additionally, while classification methods typically predict discrete grids or predefined locations, regression directly predicts precise latitude and longitude coordinates, offering superior geographic resolution.

Within this framework, the KD-tree plays a supportive role to accelerate the post-processing of regression outputs by efficiently mapping the predicted coordinates to the nearest location in the candidate set through spatial search. This process is intended to enhance the inference efficiency of the regression-based design.

In summary, our innovation lies in introducing the regression-based design, which overcomes the core limitations of traditional classification methods and provides a more adaptive and generalizable solution for next location prediction in zero-shot and cross-city tasks. The KD-tree, meanwhile, serves as a tool to enhance the efficiency of this framework. We will clarify this distinction further in the revised manuscript. Thank you for your valuable feedback!

W3 Model Architecture Innovation: There appears to be limited innovation in the model structure. I would appreciate a discussion with the authors to ensure I am not overlooking any novel aspects.

To address your concerns, we further highlight the key innovations of NextLocLLM in terms of its model architecture.

According to [1], the role of LLM in spatiotemporal data mining tasks can be categorized into three paradigms: LLM as enhancer, LLM as predictor, and LLM as agent. NextLocLLM is the first model in the field of next location prediction to integrate LLM as enhancer and LLM as predictor into a unified framework. This novel paradigm tightly combines the semantic understanding capabilities of pre-trained large language models with the geographic prediction task, offering a groundbreaking approach to next location prediction.

By leveraging LLMs in this dual role, NextLocLLM not only addresses the limitations of traditional models in cross-city tasks but also establishes a new direction for the research and application of next location prediction. This integration allows the model to utilize both LLM-enhanced feature embeddings (e.g., POI semantics) and LLM-based regression capabilities for predicting spatial coordinates, enabling a more robust, adaptable, and generalizable solution.

We deeply appreciate your insights and look forward to discussing future directions for next location prediction, particularly how the potential of LLMs can be further explored to enhance not only next location prediction but also other spatiotemporal tasks. Thank you again for your valuable feedback!

[1] Jin M, Zhang Y, Chen W, et al. Position: What Can Large Language Models Tell Us about Time Series Analysis[C] ICML. 2024.

W3 The paper could benefit from further analysis on which specific design modules contribute to the zero-shot capability of NextLocLLM and why it outperforms other LLM-based methods.

Thank you for your valuable feedback! To provide a comprehensive explanation of the model’s performance in zero-shot scenarios, we conducted additional experiments and provided detailed discussions in Section S, focusing on the contributions of key design components to zero-shot capability, particularly spatial coordinate inputs and prompt prefix. The analysis is summarized as follows:

• The Role of Spatial Coordinate Inputs: Experimental results demonstrate that using spatial coordinates significantly improves the model’s performance in zero-shot scenarios compared to location IDs, validating spatial coordinates as a universal representation of geographic information.
• The Importance of prompt prefix: Removing prompt prefix leads to a significant performance drop, highlighting their critical role in helping the model understand the data characteristics and task objectives.

Additionally, regarding why NextLocLLM outperforms other LLM-based models in zero-shot scenarios, we suggest that this advantage is closely tied to the use of spatial coordinates as inputs. Pre-trained LLMs are not inherently proficient in understanding spatial relationships between coordinates [1]. Consequently, previous LLM models relying on prompt engineering often used location IDs to represent spatial information. However, as observed in our case study (see section P), prompt-based models using location IDs can only predict locations explicitly seen in the prompt, failing entirely to generalize to unseen locations. This limitation severely impacts their applicability and accuracy. In contrast, NextLocLLM’s integration of spatial coordinates allows the model to generalize to unseen geographic regions, resulting in significant performance improvements in zero-shot tasks.

[1] Manvi R, Khanna S, Mai G, et al. Geollm: Extracting geospatial knowledge from large language models[J]. ICLR 2024

Q1. How is the frequency of each POI category within a location calculated?

Similar to [2], the frequency of POI categories at each location was calculated through the following steps. First, we mapped different types of POIs to their corresponding geographic grids. This allowed us to identify the specific POI types present within each grid. Next, we counted the frequency of each POI category in every grid, accurately capturing the distribution of POI types across different regions.

To better represent the functional attributes of each area, we aggregated fine-grained POI categories into higher-level functional groups. This aggregation was necessary because individual fine-grained POI categories (e.g., Restaurants, Cafés) often fail to comprehensively reflect the overall functionality of an area. By combining similar or related POI categories, the aggregated functional groups provide a semantically richer and more intuitive representation. For example, a grid containing multiple restaurants and cafés can be classified under the broader functional category of Dining, which directly reflects the primary use of the area without focusing on the specifics of individual POI types. This aggregation process was based on the classification rules outlined in Tables 6 and 7.

Through these steps, we constructed detailed POI frequency distributions for each location, providing semantically enriched functional attributes that are essential for supporting the model’s learning process. We include a detailed explanation of this POI frequency calculation and aggregation process in the revised manuscript. Additionally, this part of the code has already been included in our open-source repository for researchers to reference and use.

[2]Luo G, Ye J, Wang J, et al. Urban functional zone classification based on POI data and machine learning[J]. Sustainability, 2023, 15(5): 4631.

W1 Ablation studies to assess whether using spatial coordinates provides a clear advantage over location IDs is missing.

We sincerely appreciate the reviewers' suggestions regarding the ablation study in our paper. To investigate whether using spatial coordinates provides advantages over location IDs, we conducted comparative experiments (see Section O and the table below), where location IDs were directly used as inputs and outputs for the model. These results were then compared to the settings where spatial coordinates were employed.

The experimental results reveal that in the fully supervised scenario, the model utilizing spatial coordinates outperforms the one using location IDs across all metrics on both the Xi'an and SG datasets. This indicates that spatial coordinates better capture geographic relationships, thereby improving prediction performance in fully supervised settings. The advantage of spatial coordinates becomes even more pronounced in zero-shot scenarios. Models relying on location IDs perform poorly in zero-shot settings, with all metrics nearing 0%. This result further highlights the limitations of using location IDs, particularly in generalizing across cities. The discrepancy in performance indicates that location IDs fail to capture the continuity of geographic space, and the location IDs vary significantly between cities, which constrains the model's ability to generalize across urban regions.

To illustrate the superiority of spatial coordinates over location IDs more intuitively, we also present a case study (see Section P). This case study demonstrates that in zero-shot scenarios, NextlocLLM using spatial coordinates generates predictions that are significantly closer to the actual trajectories. In contrast, the predictions from the location ID-based model deviate substantially from the true trajectories.

Both experiments demonstrate the significant advantages of spatial coordinates over location IDs. Spatial coordinates not only enhance the model's ability to capture geographic relationships but also substantially improve its generalization performance in cross-city scenarios. We include a detailed analysis and discussion of these experimental results in our revised manuscript.

W2 The paper does not include a detailed description of raw trajectory processing. For instance, there is no clarification on determining the length of historical and current trajectories or whether noisy points in the raw trajectory data are filtered out.

Thanks for your suggestion. Regarding the detailed description of data preprocessing, we add comprehensive rules and procedures in Section I, along with Figure 6 as an illustrative diagram to visually demonstrate the effects of noise filtering and data cleaning. Below is a brief summary of the key steps:

• Noise Filtering: Using the trackintel library to generate staypoints, we applied rules based on time intervals (60 minutes), distance thresholds (200 meters), and minimum dwell time (10 minutes or 1 minute for the Singapore dataset) to effectively eliminate noisy data.
• Ensuring Trajectory Completeness: We filtered out trajectories with fewer than 15 days of coverage, removing short-term noise trajectories and ensuring the model learns long-term behavioral patterns.
• Spatial Grid Partitioning and POI Mapping: The urban area was divided into grids of 500m × 500m, and the number of each type of POI in each grid was calculated. This process introduced functional characteristics of regions into the trajectory data, capturing semantic and behavioral patterns.
• Optimizing Historical Trajectory Length: Based on hyperparameter sensitivity analysis, we determined that setting the historical trajectory length to 30–40 records as a default configuration ensures the model's stability and performance.

Thank you for your feedback. We are pleased to know that we have addressed most of your concerns. Regarding your doubts about the novelty of this work, we would like to take this opportunity to summarize our key contributions beyond the LLM-enhanced POI embedding:

1. Replacing Location IDs with Spatial Coordinates for Location Representation

Existing methods typically use numeric identifiers (location IDs) to represent locations. However, these discrete IDs fail to capture geographical relationships between locations and cannot generalize across cities due to the significant differences in location ID systems between cities. We propose using spatial coordinates instead of discrete IDs to represent geographical information.

Our experiments—including comparisons between geographic coordinates and location IDs, ablation studies on zero-shot prediction performance, and case studies—demonstrate the clear advantages of spatial coordinates. These benefits are evident in both fully-supervised tasks and zero-shot tasks, where spatial coordinates significantly improve performance.

2. A Regression-Based Approach for Zero-Shot Next Location Prediction

We introduce a regression-based approach that directly predicts the spatial coordinates of the next location, replacing the traditional classification-based design and offering a more effective solution for zero-shot tasks. Conventionally, next location prediction is treated as a classification problem where each possible location ID is considered a class, and the model predicts the probability distribution over these classes to determine the next location. However, this approach faces significant limitations in zero-shot scenarios:

• Limited transferability: Classification models are bound to city-specific location ID systems, which vary widely and hinder effective cross-city generalization.
• Lack of generalization: Classification relies on predefined classes and cannot predict unseen locations.
• Low geographical resolution: Classification methods typically predict discrete grids or locations, whereas regression provides precise latitude and longitude predictions.

In contrast, our regression-based method directly predicts the spatial coordinates of the next location, eliminating dependence on location IDs and significantly improving the applicability of the model to zero-shot cross-city tasks.

3. First Integration of LLM as Both Enhancer and Predictor in a Unified Framework for Next Location Prediction

Large language models (LLMs) encapsulate extensive knowledge and rules about daily life [1][2], exhibit strong reasoning abilities [2][3], and have achieved remarkable success in spatiotemporal tasks when used as data enhancers, predictors, or agents [4]. Inspired by this, we are the first to integrate LLMs as both enhancers and predictors within a unified framework for next location prediction. This novel paradigm tightly combines the semantic understanding capabilities of pretrained LLMs with geographic prediction tasks, offering a groundbreaking solution for next location prediction.

By leveraging LLMs in dual roles, NextLocLLM not only overcomes the limitations of traditional models in cross-city tasks but also opens new avenues for research and applications in next location prediction. This integration enables the model to utilize both LLM-enhanced feature embeddings (e.g., POI semantics) and LLM-based regression capabilities for spatial coordinate prediction, resulting in a more robust, adaptive, and generalizable solution.

We believe the above three contributions, together with the LLM-enhanced POI embedding, constitute the core innovations of this work. These designs address the limitations of traditional approaches while introducing a novel paradigm for next location prediction research.

Thank you again for your valuable feedback! We look forward to future opportunities for further discussion and collaboration!

[1] Prompt-Based Time Series Forecasting: A New Task and Dataset

[2] Large Language Models Are Zero-Shot Time Series Forecasters

[3] LSTPrompt: Large Language Models as Zero-Shot Time Series Forecasters by Long-Short-Term Prompt

[4] What Can Large Language Models Tell Us about Time Series Analysis