Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models

Thank you for the thoughtful and detailed feedback. We sincerely appreciate your recognition of the $**open-sourced code**$ and the $**comprehensiveness of our related work**$, as well as the time and effort you put into reviewing our paper. Below, we address your concerns point by point to further clarify and strengthen our contributions.

1. The explanation of potential correct-by-coincidence reasoning traces (W1)

A1: While some distilled reasoning traces may contain flawed logic that coincidentally yields correct predictions, we emphasize that $**these raw traces are not directly used in GRPO training**$. Instead, we extract key entities (e.g., architecture, plant) and cross-validate them against predictions from a semantic segmentation model. Only entities consistent with visual evidence are retained for supervision, significantly reducing the risk of propagating spurious reasoning. Since SFT directly learns from the full synthetic reasoning traces, it is more vulnerable to coincidental correctness, where flawed reasoning leads to correct predictions. In contrast, our GRPO-based approach relies on verified, structured signals (e.g., entity-level consistency), making it more robust to such artifacts. This distinction is reflected in Table 2 of the main manuscript (Row 2 vs. Rows 3–8), where GRPO variants consistently outperform standard SFT.

2. The explanation of validation with the semantic segmentation models (W2)

A2: We provide additional clarification on how semantic segmentation is used to validate reasoning traces. As described in Lines 149–152, we apply MaskFormer [1], a state-of-the-art panoptic segmentation model, to each image in our dataset to obtain pixel-wise semantic predictions across 150 common categories, including sky, building, tree, road, vegetation, car, and others. This provides a dense, localized understanding of the visual content. Crucially, this segmentation output is used to verify the factual consistency of entities mentioned in the distilled reasoning traces. For example, if the reasoning claims "the presence of snow-capped mountains suggests high altitude," we check whether mountain and snow regions are indeed present in the corresponding image regions. Similarly, if a trace references "urban infrastructure" or "dense road networks," we validate whether buildings and roads are detected with high confidence and spatial coverage. This cross-modal validation ensures that only reasoning paths grounded in actual visual content are preserved, reducing the risk of propagating hallucinated or coincidentally correct logic. We acknowledge that this part was briefly described in the original submission, and we will expand it in the revised manuscript with a clearer explanation and an example in the appendix.

3. The explanation of Visual Grounding Consistency Reward (W3)

A3: The reward design in Line 174 is based on the entity-level consistency between the reasoning trace $\hat{r_i}$ and the visual content of the image, as verified by semantic segmentation (see A2). Specifically, we extract the intersection of entities present in both the reasoning trace and the MaskFormer-predicted segmentation map. This does not require all detected visual entities to be mentioned (which would indeed be overly strict), nor does it consider a match valid if any single entity aligns (which would be too lenient). Instead, the reward reflects the proportion of mentioned entities that are visually grounded, normalized by the total number of relevant entities in the trace. As shown in Table 2 of the main manuscript (Row 7 vs. Row 2), this consistency-based reward leads to significant performance gains compared to standard SFT, demonstrating its effectiveness in guiding geolocation-capable reasoning. We will revise the text to include a formal definition of the reward and an illustrative example in the appendix.

4. The explanation of only considering the middle three levels (W4)

A4: We acknowledge that standard benchmarks are typically evaluated using five distance-based accuracy levels (1km, 25km, 200km, 750km, 2500km). In our work, however, the primary focus is on reasoning-based geolocation, where the model generates semantic reasoning traces rather than predicting precise coordinates. Given this design, the localization resolution is inherently limited to coarser granularity, making performance at the 1km level less reflective of the model’s reasoning capability. To maintain consistency with our method’s scope, we initially reported results on the middle three levels (25km, 200km, 750km) in the main text, which better align with city- and country-level predictions. As shown in Table 1 (updated), we now include results across all five levels.

Table 1. Performance of GLOBE on IMG2GPS3k

As expected, performance at the 1km level is lower, which is consistent with the fact that our approach does not aim for street-level precision.

5. Additional experiments for the proposed method (W5)

A5: We agree that deeper ablation and sensitivity analyses are important for understanding the robustness and generalizability of our method. In response, we have conducted three additional sets of experiments, which will be included in the revised manuscript:

$**a. Sensitivity to reward weights ($λ_1$,$λ_2$,$λ_3$)**$: The weights λ₁, λ₂, and λ₃ are fixed during training and were initially set to 0.2, 0.5, and 1, respectively. Since the primary objective is to accurately predict the city and country, we assigned the highest weight to λ₃, which directly supervises geolocation accuracy. Consistency is important for reducing hallucinations in the reasoning trajectory; thus, we assigned a relatively high weight to λ₂. Locatability, being a binary score that evaluates whether the image and its reasoning are geographically grounded, was assigned a smaller weight (λ₁) to provide auxiliary regularization. As shown in Table 2, we evaluate different weight combinations in the GRPO objective. The results confirm that our chosen setting (0.2,0.5,1) achieves the best performance.

Table 2. Performance comparison of weight selection ( λ₁, λ₂, and λ₃) for the GRPO framework on MP16-Reason-Test

$**b. Impact of backbone architecture**$: Table 3 compares performance across different vision-language model backbones (e.g., Qwen2.5-VL-7B vs. InternVL3-8B). The consistent gains of GRPO over SFT across backbones demonstrate the generalizability of our training framework.

Table 3. Performance comparison using different backbones on MP16-Reason-Test

$**c. Ablation on dataset curation**$: To evaluate the contribution of our validation steps in data curation, we are conducting experiments using ablated training sets (e.g., randomly sampled data, or data with only one LVLM validation). Table 4 shows clear performance drops when any of the validation steps are removed, confirming their importance in building a high-quality dataset.

Table 4. Performance comparison using different Training datasets on MP16-Reason-Test, with Qwen2.5-VL-7B

6. The writing suggestions for the experiment section. (W6)

A6: We thank the reviewer for the feedback. We agree that the current structure may be confusing and will revise the experimental section to clearly separate main results (Section 4.2.1, Table 1) from ablation studies (Section 4.2.2, Table 2). We will clarify the purpose of each table and improve the flow in the revised manuscript.

7. Some minors (W7)

A7: We thank the reviewer for the suggestions. We will clarify the model size used in GLOBE in Table 1 and move the case study in Figure 5 to a dedicated subsection (e.g., "Qualitative Analysis" or "Case Study") for better organization.

[1] Per-Pixel Classification is Not All You Need for Semantic Segmentation, NeurIPS 2021

Thank you for the authors’ rebuttal. Your response has addressed most of my concerns.

1. For Q1, should it correspond to a drop in street-level accuracy rather than continent-level accuracy?
2. For Q2, I have fully understood your explanation.
3. For Q3, I was expecting a comparison of GLOBE’s efficiency with other baseline methods. Even if GLOBE is less efficient than others (which is understandable), I still hope the authors can include this part in the final version for readers’ reference.

I will update my rating accordingly.

Thank you for your detailed review and critical feedback. We appreciate your positive assessment of our $**novel idea**$, $**curated benchmark**$, and $**extensive ablation study**$. We respect your concerns and provide clarifications below to address the key misunderstandings and better convey the contributions of our work.

1. The explanation of data curation and the biases of the LLMs (W1 & Q1)

A1: We clarify that the models used for data curation are $**distinct**$ from the training backbone, both in scale and capability. Specifically, our distillation sources include $**Qwen2.5-VL-72B**$ and InternVL3-78B, while the student model is $**Qwen2.5-VL-7B**$, a significantly smaller and less capable architecture, as documented in the Qwen2.5-VL technical report [1].

We emphasize that our data curation is a controlled distillation process aimed at extracting high-quality, reasoning-rich samples, not a source of bias. Rather than promoting memorization, it is designed to activate the model’s reasoning ability and leverage internal world knowledge, even with limited visual diversity. For a detailed explanation, please refer to our response to Reviewer wDfp in A2.

2. The discussion of GLOBE and GPT4.1 (W2)

A2: While our method performs strongly among open-source LVLMs, we acknowledge that it does not surpass closed-source models like GPT-4.1, an expected gap due to differences in scale and resources. Our aim is to advance open, reproducible, and data-efficient LVLM training. Notably, GLOBE (7B) outperforms the much larger Qwen2.5-VL-72B on both MP16-Reason-Test and IMG2GPS3K, demonstrating the effectiveness of our distillation framework. For details, please refer to our response to Reviewer wDfp in A3.

3. The explanation of the reward function $R_{loc}$ (W3 & Q3)

A3: We apologize if our original description of the $R_{loc}$ reward function was unclear and may have led to a misunderstanding. We appreciate the opportunity to clarify its design and importance. R_loc is not intended to trivially encourage all outputs to be "localizable." Although the training set is curated to contain generally localizable images, the key idea behind R_loc is to assess the locatability of an image–reasoning pair, rather than the image alone. During GRPO training, multiple reasoning traces are sampled for each image, and $R_{loc}$ is computed based on both the image and its generated explanation. As a result, different reasoning traces lead to different $R_{loc}$ scores. This mechanism implicitly encourages the model to generate higher-quality, more location-grounded reasoning.

4. The tracking performance on intermediate tasks (W4)

A4: As figures cannot be included in the rebuttal process, we present representative data in Table 1.

Table 1. Performance on intermediate tasks

The evaluation rewards exhibit steady improvement throughout training, with optimal performance observed at Step 1500. We observe that both rewards increase steadily during training, showing consistent improvement and stabilizing by Step 1000, with only marginal gains thereafter. By Step 1500, both reach their peak values, indicating convergence. We will also update charts showing the reward score trajectories throughout training to illustrate the optimization dynamics better.

5. The clarification of the proposed dataset (W5 & Q2)

A5: We agree that dataset statistics and geographical distribution are important for understanding the properties of MP16-Reason. As shown in Table 2, we have already reported key statistics of the dataset, including the number of images in the training and test sets, as well as the distribution across cities and countries. To further improve transparency, in the revised manuscript, we will include a geographical heatmap to illustrate the spatial coverage of both the training and test splits.

Table 2. Statistics of the proposed dataset

'#' indicates number of.

6. The clarification of the methodology section (W6)

A6: We thank the reviewer for the thoughtful feedback. Several methodological components, such as reasoning alignment, entity matching, and group definition, are central to our approach. To enhance clarity and accessibility for a broader audience, we will revise the Methodology section in the final version to include more precise definitions, improved descriptions, and illustrative examples.

7. The supplementary comparison experiments (W7)

A7: We thank the reviewer for pointing out the additional related work [2, 3, 4]. We will update both the Related Work section and the main results table accordingly (Since [4] does not have an open-source model, we only compared methods [2] and [3] on MP16-reasoning-Test and IMG2GPS3k in Table 3 and 4). The experimental results also show that the performance of the model trained only with street view data will be poorer in diverse scenarios, which is consistent with the viewpoint proposed in our work.

Table 3. Performance comparison on MP16-reasoning-Test

Table 4. Performance comparison on IMG2GPS3k

8. The reason for using GRPO-based RL ( W8 & Q4)

A8: We clarify that our choice of RL, specifically Group-based Reward PPO (GRPO), over standard supervised fine-tuning (SFT) is motivated by the need to guide multi-step, structured reasoning rather than simply mimic surface-level patterns in the training data.

In SFT, the model is trained to reproduce reference outputs (e.g., reasoning traces), which often leads to shallow imitation and limited generalization, especially when the test inputs differ in structure or context. In contrast, RL with carefully designed rewards allows us to shape intermediate behaviors: for example, encouraging the model to first identify geographic clues (e.g., vegetation, architecture), then narrow down the region. This stepwise reasoning is reflected in the generated traces and makes the decision process more interpretable. GRPO, in particular, improves upon standard PPO by optimizing based on relative preferences within a group of sampled responses, rather than absolute reward signals. As shown in prior work [5], this leads to more stable training and better alignment with human (or model-based) judgment, especially when rewards are sparse or noisy, as in reasoning tasks.

9. The distance-based metrics for evaluation (W9)

A9: We clarify that the benchmark Img2GPS3K is evaluated using distance-based metrics. Since GLOBE predicts semantic location labels (e.g., city and country), we report accuracy as the primary metric. However, we can also provide distance-based results (See Table 3 and 4) using an external geocoding tool (Microsoft Azure Maps tool).

10. Additional ablation experiments for CoT (W10)

A10: We conducted ablation studies on two backbones (Qwen2.5-VL-7B, InternVL3-8B) comparing with and without CoT. Table 5 shows consistent but modest gains, indicating the positive influence of CoT. We will include this analysis in the revision.

Table 5. Performance comparison on MP16-reasoning-Test

11. Some remarks (Remarks)

A15: We thank the reviewer for these helpful remarks. We will carefully address each point in the revision, including acronym refinement, citation corrections, terminology consistency, figure format improvements, and clarification of task definitions.

[1] Qwen2.5-VL Technical Report, arXiv 2025
[2] OpenStreetView-5M: The Many Roads to Global Visual Geolocation, CVPR2024
[3] Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation, CVPR2025
[4] Exploiting the Earth's spherical geometry to geolocate images, ECML2019
[5] Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning, arXiv 2025

Thank you for your positive feedback and insightful comments. We are greatly encouraged by your recognition of our $**curated dataset**$, the $**strong experimental performance**$, and $**the innovative ideas**$ presented in our work. Although no specific questions were raised, we would still like to address the weaknesses mentioned in your review to further clarify our motivation.

1. The explanation of bias in three VLMs (W1)

A1: We thank the reviewer for raising this important point regarding potential bias in our data distillation process. The use of three diverse, high-performing teacher models (Qwen2.5-VL-72B, InternVL3-78B, and GeoCLIP) is a deliberate design choice aimed at mitigating model-specific biases. Instead of relying on a single teacher, which may propagate its own systematic preferences or reasoning styles, our multi-teacher framework leverages the consensus and complementarity of multiple SOTA VLMs to enhance the robustness and diversity of the distilled signals [1]. As shown in Table 1, we compare performance using a single VLM versus our proposed multi-VLM strategy. Across both SFT and GRPO settings, distillation from three VLMs consistently yields better results, indicating improved training data quality and reduced bias compared to using only one VLM.

Table 1. Performance comparison on MP16-reasoning-Test

*Here, we report distance-based (Haversine) accuracy computed after converting predicted city names to GPS coordinates via the Microsoft Azure Maps tools.

2. The explanation of bias in MP16-Reason (W2)

A2: We emphasize that our data curation is not a source of bias, but a controlled knowledge distillation process designed to extract high-quality, informative samples for efficient LVLM training. This approach aligns with recent advances in the literature, where large models are used to construct compact, high-signal training sets for downstream tasks [2, 3, 4]. Crucially, the goal of distillation is not to replicate the full data distribution, but to identify a representative and semantically rich subset that maximizes learning efficiency. By prioritizing quality over quantity, our method enables state-of-the-art performance with significantly reduced training data and cost. $**Moreover, our objective is not to encourage memorization of specific samples**$, but to activate the model’s reasoning capabilities through well-structured, high-signal examples. These curated instances serve as cognitive anchors that guide the model to internalize spatial and contextual reasoning patterns, thereby promoting generalization, leveraging the model’s internal world knowledge even in the presence of limited visual diversity.

3. The results compared to the closed ones (W3)

A3: While our method achieves strong performance among open-source LVLMs, we acknowledge that it does not surpass closed-source models such as GPT-4.1. This gap is not unexpected, as industrial-scale systems benefit from vastly larger training datasets, proprietary infrastructure, and orders-of-magnitude more compute resources, typically beyond the scope of academic research. Our goal is not to compete directly with such closed systems, but to advance open, reproducible, and data-efficient LVLM training. Within this context, our approach demonstrates significant progress. Notably, GLOBE, built upon the Qwen2.5-VL-7B backbone, outperforms the much larger Qwen2.5-VL-72B (the very model used to generate the distilled data) on the MP16-Reason-Test and IMG2GPS3k benchmark. This "student surpassing teacher" result highlights the effectiveness of our distillation and training framework in extracting and refining high-quality knowledge, rather than merely replicating teacher behavior. It underscores the value of curated, reasoning-grounded supervision in boosting model capabilities beyond scale. To promote transparency and community advancement, we will release all $**code**$, $**models**$, and $**curated data**$.

[1] MTMS: Multi-teacher Multi-stage Knowledge Distillation for Reasoning-Based Machine Reading Comprehension, SIGIR 2024
[2] LESS: selecting influential data for targeted instruction tuning. ICML 2024
[3] LIMA: Less Is More for Alignment, NeurIPS 2023
[4] Jiuzhang3.0: Efficiently improving mathematical reasoning by training small data synthesis models, NeurIPS 2024

Thank you for the positive assessment and constructive feedback. We appreciate your support for our contributions ($**Novel Dataset**$, $**Innovative Methodology**$, $**Reasonable Motivation**$, $**Strong Results**$, and $**Data Efficiency**$) and address below the remaining concerns to further clarify and strengthen our work.

1. The explanation of fine-grained localization (W1)

A1: We thank the reviewer for this insightful comment. We fully agree that fine-grained localization is a critical capability, and we appreciate the opportunity to clarify and strengthen the evaluation of our work. Regarding localization precision, in this rebuttal process, all reported results are now computed using a high-precision evaluation pipeline. Specifically, we use Microsoft Azure Maps to convert predicted city names into precise GPS coordinates and compute the Haversine distance between the predicted and ground truth locations. Accuracy is then evaluated at five standard levels: Street (1km), City (25km), Region (200km), Country (750km), and Continent (2500km), following established protocols in prior work. Although performance at the 1km level is inherently limited (since our model operates at the city level or coarser), our approach achieves competitive results compared to other open-source methods at the city level and above. This demonstrates the effectiveness of our reasoning-driven framework within its intended scope (see our response to Reviewer T7Q7 and the updated results in Table 3 and Table 4). Second, we agree that retrieval-based methods can achieve significantly higher precision and are particularly effective for fine-grained localization. As suggested, we will revise the second paragraph of the Introduction to explicitly acknowledge this strength of retrieval approaches, and better position our method as complementary

2. The supplementary related work (W2)

A2: We thank the reviewer for pointing this out. We will revise the Related Work section to incorporate the suggested works and improve its overall completeness.

3. The clarification of computational costs (W3)

A3: For data curation, we deployed Qwen2.5-VL-72B and InternVL3-78B using 8×H20 GPUs under the VLLM framework, while GeoCLIP was run separately on a single H20 GPU. All models performed inference over the original MP16 dataset. For GRPO training, each experiment with our 7B-sized model was conducted using 8×H20 GPUs. With a batch size of 16, the training throughput was approximately 0.44 examples per second.

4. The explanation of bias in MP16-Reason (W4)

A4: We emphasize that our data curation is not a source of bias, but a controlled knowledge distillation process designed to extract high-quality, informative samples for efficient LVLM training. This approach aligns with recent advances in the literature, where large models are used to construct compact, high-signal training sets for downstream tasks [1, 2, 3]. Crucially, the goal of distillation is not to replicate the full data distribution, but to identify a representative and semantically rich subset that maximizes learning efficiency. By prioritizing quality over quantity, our method enables state-of-the-art performance with significantly reduced training data and cost. $**Moreover, our objective is not to encourage memorization of specific samples**$, but to activate the model’s reasoning capabilities through well-structured, high-signal examples. These curated instances serve as cognitive anchors that guide the model to internalize spatial and contextual reasoning patterns, thereby promoting generalization, leveraging the model’s internal world knowledge even in the presence of limited visual diversity.

5. The discussion of GLOBE and GPT-4.1 (W5)

A5: While our method achieves strong performance among open-source LVLMs, we acknowledge that it does not surpass closed-source models such as GPT-4.1. This gap is not unexpected, as industrial-scale systems benefit from vastly larger training datasets, proprietary infrastructure, and orders-of-magnitude more compute resources, typically beyond the scope of academic research. Our goal is not to compete directly with such closed systems, but to advance open, reproducible, and data-efficient LVLM training. Within this context, our approach demonstrates significant progress. Notably, GLOBE, built upon the Qwen2.5-VL-7B backbone, outperforms the much larger Qwen2.5-VL-72B (the very model used to generate the distilled data) on the MP16-Reason-Test and IMG2GPS3k benchmark. This "student surpassing teacher" result highlights the effectiveness of our distillation and training framework in extracting and refining high-quality knowledge, rather than merely replicating teacher behavior. It underscores the value of curated, reasoning-grounded supervision in boosting model capabilities beyond scale. To promote transparency and community advancement, we will release all $**code**$, $**models**$, and $**curated data**$.

[1] LESS: selecting influential data for targeted instruction tuning. ICML 2024
[2] LIMA: Less Is More for Alignment, NeurIPS 2023
[3] Jiuzhang3.0: Efficiently improving mathematical reasoning by training small data synthesis models, NeurIPS 2024

Thank you for the detailed responses to my concerns. I have no further questions. I strongly encourage that the authors update the final version based on the reviewers' comments and the author's own rebuttal, including the explanation of fine-grained localization, more discussion on related work, etc.

Thank you for the time and effort in reviewing our work. We sincerely appreciate your constructive feedback, which has helped us identify areas for improvement. We are also grateful for your recognition of our curated dataset, the novelty of our approach, and the strong performance demonstrated in the experiments. Below, we address the concerns raised and clarify key aspects of our contributions.

1. The explanation of bias in MP16-Reason (W1)

A1: We emphasize that our data curation is not a source of bias, but a controlled knowledge distillation process designed to extract high-quality, informative samples for efficient LVLM training. This approach aligns with recent advances in the literature, where large models are used to construct compact, high-signal training sets for downstream tasks [1, 2, 3]. The goal of distillation is not to replicate the full data distribution, but to extract a representative, high-signal subset that enhances learning efficiency. Rather than promoting memorization, our curated examples are designed to activate reasoning and guide the model to generalize via internal world knowledge, even with limited visual diversity. For a detailed explanation, please refer to our response to Reviewer wDfp in A2.

2. The clarification of performance across different conditions (W2)

A2: To better evaluate the generalization and external validity of our approach across diverse visual scenes, we have conducted a fine-grained breakdown of performance on the MP16-Reason-Test. Following the MP-16 dataset [4], we categorize images by scene condition (e.g., indoor, natural, urban) based on metadata that includes predicted scene probabilities. For each image, we assign the label with the highest predicted probability as its scene classification.

Table 1. Performance comparison across different scene conditions on MP16-Reason-Test. Each entry reports distance-based localization accuracy (km) at multiple scales: Street (1km) / City (25km) / Region (200km) / Country (750km) / Continent (2500km).

It can be found that GeoReasoner performs best in urban scenes, which is closely related to its training on the Street View dataset. Among them, our method consistently achieves the highest performance across all scene types, particularly excelling in nature scenes with substantial gains over both baselines. This detailed analysis confirms our claim of superior scene diversity and supports the broader applicability of our approach.

3. The clarification of key parameters (λ) in GRPO (W3 & Q3)

A3: The weights λ₁, λ₂, and λ₃ are fixed during training and were initially set to 0.2, 0.5, and 1, respectively. Since the primary objective is to accurately predict the city and country, we assigned the highest weight to λ₃, which directly supervises geolocation accuracy. Consistency is important for reducing hallucinations in the reasoning trajectory; thus, we assigned a relatively high weight to λ₂. Locatability, being a binary score that evaluates whether the image and its reasoning are geographically grounded, was assigned a smaller weight (λ₁) to provide auxiliary regularization. We have conducted additional experiments to validate the effect of different weight settings.

Table 2. Performance comparison of weight selection ( λ₁, λ₂, and λ₃) for the GRPO framework on MP16-Reason-Test

The results suggest that the setting λ₁=0.2, λ₂=0.5, λ₃=1, used in our main experiments, is relatively effective and well-balanced for the task.

4. The potential baselines (W4)

A4: We have additionally included Doubao-1.5-thinking-vision-pro [8] for comparison. While GPT-4.1 and Doubao-1.5-pro [8] still outperform GLOBE in absolute accuracy, they are large-scale proprietary systems with undisclosed architectures and training data. In contrast, GLOBE is fully open-source, data-efficient, and trained on curated supervision, offering a competitive and reproducible alternative for reasoning-based geo-localization. To promote transparency and community advancement, we will release all $**code**$, $**models**$, and $**curated data**$.

Table 3. Performance comparison on MP16-reasoning-Test

Regarding the model size, the parameter count of GLOBE (based on Qwen2.5-VL-7B) is $**stated in the caption of Table 2**$ in the paper. We acknowledge that this may not be sufficiently prominent, and in the revision, we will explicitly report the model size (7B) in the main text.

5. The discussion of privacy concerns (W5)

A5: We thank the reviewer for raising this important point. While our work does not aim to directly address privacy risks or adversarial robustness (such as [6] and [7]), we acknowledge the broader societal implications of geo-localization systems ([5]). The discussion in Section A.3 was intended as a responsible reflection on potential concerns, rather than a claim of technical contributions in this area.

6. The clarification of annotations among different datasets (Q1)

A6: Different annotations are provided in Table 4, showing that our multi-model distillation produces more diverse and independent reasoning paths compared to existing dataset baselines. We use multi-model voting to reduce hallucinations compared with a single model. During GRPO training, the reasoning paths themselves are not used directly; instead, entities extracted from the paths serve as the reward signal. Reviewer wDfp’s Table 1 also provides the performance impact of single- versus multi-model annotation.

Table 4. Comparison of different annotations

7. The discussion of failure cases (Q2)

A7: We categorize failure cases into two types: Error Reasoning and Right Reasoning, with representative examples shown in Table 5. Analysis reveals two common patterns: (1) visually similar features (e.g., domes, arches) leading to incorrect landmark attribution, and (2) correct reasoning biased toward locations more frequent in the training set.

Table 5. Comparison of failure cases

8. The discussion of robustness (Q4)

A8: To assess robustness, we conducted zero-shot testing on a 3K subset of the out-of-domain OSV-5M [10] dataset (a geolocalization dataset of streetview images). As shown in Table 6, GLOBE demonstrates strong generalization and outperforms all open-source baselines, including both LVLMs and non-LVLMs.

Table 6. Performance comparison on OSV-5M [10] (mini-3K)

[1] LESS: selecting influential data for targeted instruction tuning. ICML 2024
[2] LIMA: Less Is More for Alignment, NeurIPS 2023
[3] Jiuzhang3.0: Efficiently improving mathematical reasoning by training small data synthesis models, NeurIPS 2024
[4] The benchmarking initiative for multimedia evaluation: Mediaeval 2016. IEEE MultiMedia [5] Granular Privacy Control for Geolocation with Vision Language Models, arXiv 2024
[6] Images are Achilles’ Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models, ECCV 2024
[7] Agent Smith: A Single Image Can Jailbreak One Million Multimodal LLM Agents Exponentially Fast, ICML 2024
[8] Seed1.5-VL Technical Report, arXiv 2025
[9] G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models, NeurIPS 2024
[10] OpenStreetView-5M: The Many Roads to Global Visual Geolocation, CVPR2024

Thanks for the rebuttal with additional analysis. The results have addressed my concerns, and I have updated the score.