Locatability-Guided Adaptive Reasoning for Image Geo-Localization with Vision-Language Models
Authors: Bo Yu, Fengze Yang, Yiming Liu, Chao Wang, Xuewen Luo, Taozhe Li, Ruimin Ke, Xiaofan Zhou, Chenxi Liu
Organizations: The University of Utah · The University of Oklahoma · Worcester Polytechnic Institute · Rensselaer Polytechnic Institute · University of Illinois Chicago
The emergence of Vision-Language Models (VLMs) has introduced new paradigms for global image geo-localization through retrieval-augmented generation (RAG) and reasoning-driven inference. However, RAG methods are constrained by retrieval database quality, while reasoning-driven approaches fail to internalize image locatability, relying on inefficient, fixed-depth reasoning paths that increase hallucinations and degrade accuracy. To overcome these limitations, we introduce an Optimized Locatability Score that quantifies an image's suitability for deep reasoning in geo-localization. Using this metric, we curate Geo-ADAPT-51K, a locatability-stratified reasoning dataset enriched with augmented reasoning trajectories for complex visual scenes. Building on this foundation, we propose a two-stage Group Relative Policy Optimization (GRPO) curriculum with customized reward functions that regulate adaptive reasoning depth, visual grounding, and hierarchical geographical accuracy. Our framework, Geo-ADAPT, learns an adaptive reasoning policy, achieves state-of-the-art performance across multiple geo-localization benchmarks, and substantially reduces hallucinations by reasoning both adaptively and efficiently.
Figures & tables
Figure 1: Illustration of locatability-guided adaptive reasoning. Current reasoning VLMs for image geo-localization are trained on fixed chain of thoughts, generating predetermined reasoning trajectories step by step without considering the actual image locatability during inference. In our work, we propose Geo-ADAPT, an adaptive reasoning framework that dynamically adjusts reasoning depth based on the image’s Optimized Locatability Score Lopt .
Figure 2: Overview of Geo-ADAPT framework. Our approach comprises three components: (1) an Optimized Locatability Score Lopt quantifying reasoning feasibility, (2) locatability-stratified dataset curation Geo-ADAPT-51K with enriched implicit reasoning trajectories, and (3) two-stage GRPO training with adaptive depth, grounding, and hierarchical accuracy rewards for dynamic reasoning allocation.
Figure 3: Motivation for Optimized Locatability Score Lopt and locatability-guided adaptive reasoning strategies. (a) RAG retrieves visually similar candidates but fails due to being unable to convert implicit visual cues to rich semantic priors. (b) Standard reasoning initially fails similarly, but integrating implicit cues from retrieved image candidates into deep reasoning enables correct localization. Images where dReason>dRAG+τmargin receive lower Lopt , triggering adaptive deep reasoning.
IM2GPS3K
YFCC4K
Methods
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
[L]kNN, sigma=4 [ 52 ]
7.5
18.3
27.7
38.7
57.2
2.4
6.2
9.8
23.9
43.1
PlaNet [ 58 ]
7.8
25.8
33.7
48.7
63.2
6.5
14.2
22.4
36.0
56.4
CPlaNet [ 45 ]
11.3
25.7
34.7
47.7
65.3
6.6
15.4
21.6
36.9
54.5
ISNs [ 36 ]
10.1
29.2
35.5
50.5
64.5
6.7
15.7
25.1
36.7
55.1
Translocator [ 39 ]
12.1
30.5
47.6
57.7
80.5
7.6
19.7
25.6
41.8
60.2
Table 1: Performance comparison for Geo-ADAPT-8B on IM2GPS3K and YFCC4K with baseline methods. For all metrics, higher is better. The best performances are highlighted in bold , while the best performances among baselines are underlined . Change represents the relative improvement ↑ or decline ↓ of our method over the best baseline.
Geo-ADAPT-51K Test
Methods
City Name Acc.
Country Name Acc.
Gemma3-27B [ 49 ]
37.6
71.9
InternVL3.5-38B [ 56 ]
36.1
79.4
Gemini 2.5 Flash † [ 9 ]
54.1
87.2
Qwen3-VL-30B [ 3 ]
43.9
83.5
GeoReasoner [ 25 ]
31.8
67.9
Table 2: Localization accuracy for city and country name prediction on the Geo-ADAPT-51K test split compared with baseline reasoning VLMs. † denotes models that are not publicly available.
IM2GPS3K
Methods
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
w/o Daug
17.7
43.2
60.1
75.4
89.2
w/o Rdepth
17.2
44.3
60.5
76.7
88.9
w/o Rvis
17.4
43.7
60.8
76.8
89.0
w/o Rgeo
15.3
40.4
58.9
75.7
87.9
w/o TSFT
17.8
44.9
61.3
76.5
89.1
Table 3: Ablation study on IM2GPS3K. The worst performance results in ablation study are underlined .
Geo-ADAPT-51K Test
Methods
City Name Acc.
Country Name Acc.
w/o Daug
53.1
87.5
w/o Rdepth
54.7
88.6
w/o Rvis
53.9
87.8
w/o Rgeo
51.4
86.3
w/o TSFT
54.9
88.4
Table 4: Ablation study on Geo-ADAPT-51K test. The worst performance results in ablation study are underlined .
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Setting
GPU
NVIDIA H200 × 4
Dataset Samples
51K
Batch Size
64
Batch Size per Device
16
Training GPU Memory
90 GB / GPU
Base Model
Qwen3-VL-8B-Instruct
Appendix
Table 5: More Details on Training and Inference.
Hyper-parameter
Value
Optimizer Settings
Optimizer
AdamW
Adam β1
0.9
Adam β2
0.95
Weight Decay
0.1
Learning Rate Schedule
Appendix
Table 6: Hyper-parameter Settings of Proposed Geo-ADAPT.
Figure 4: Qualitative comparison on a weakly localizable sample ( Lopt=0.42 ). Facing visual ambiguity, Geo-ADAPT adaptively expands its reasoning depth (202 tokens), conducting a detailed forensic scan to extract the address “17510” and localize the exact shopping center. In contrast, the baseline generates a longer but less grounded response (318 tokens) and fails to identify the city.
Figure 5: Qualitative comparison on a highly localizable sample ( Lopt=0.67 ). Recognizing the strong explicit cues (Olympic Rings), Geo-ADAPT adaptively shortens its reasoning path for efficiency (156 tokens), rapidly converging on the correct location. This contrasts with the baseline, which engages in verbose description (329 tokens) without improving precision.
Vision-language models (VLMs) have shown a promising ability in image geolocation, but they still lack structured geographic reasoning and the capacity for autonomous self-evolution. Existing methods predominantly rely on implicit parametric memory, which often exploits outdated knowledge and generates hallucinated reasoning. Furthermore, current inference is a "one-off" process, lacking the feedback loops necessary for self-evolution based on reasoning outcomes. To address these issues, we propose GeoSkill, a training-free framework based on an evolving Skill-Graph. We first initialize the graph by refining human expert trajectories into atomic, natural-language skills. For execution, GeoSkill employs an inference model to perform direct reasoning guided by the current Skill-Graph. For continuous growth, an Autonomous Evolution mechanism leverages a larger model to conduct multiple reasoning rollouts on image-coordinate pairs sourced from web-scale data and verified real-world reasoning. By analyzing both successful and failed trajectories from these rollouts, the mechanism iteratively synthesizes and prunes skills, effectively expanding the Skill-Graph and correcting geographic biases without any parameter updates. Experiments demonstrate that GeoSkill achieves promising performance in both geolocation accuracy and reasoning faithfulness on GeoRC, while maintaining superior generalization across diverse external datasets. Furthermore, our autonomous evolution fosters the emergence of novel, verifiable skills, significantly enhancing the system's cognition of real-world geographic knowledge beyond isolated case studies.
Chenjie Yang, Yutian Jiang, Yutong Deng +1
Southwest Jiaotong University Chengdu, China · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Zhejiang University Hangzhou, China
Image geolocalization has traditionally been addressed through retrieval-based place recognition or geometry-based visual localization pipelines. Recent advances in Vision-Language Models (VLMs) have demonstrated strong zero-shot reasoning capabilities across multimodal tasks, yet their performance in geographic inference remains underexplored. In this work, we present a systematic evaluation of multiple state-of-the-art VLMs for country-level image geolocalization using ground-view imagery only. Instead of relying on image matching, GPS metadata, or task-specific training, we evaluate prompt-based country prediction in a zero-shot setting. The selected models are tested on three geographically diverse datasets to assess their robustness and generalization ability. Our results reveal substantial variation across models, highlighting the potential of semantic reasoning for coarse geolocalization and the limitations of current VLMs in capturing fine-grained geographic cues. This study provides the first focused comparison of modern VLMs for country-level geolocalization and establishes a foundation for future research at the intersection of multimodal reasoning and geographic understanding.
Geo-localization from a single image at planet scale (essentially an advanced or extreme version of the kidnapped robot problem) is a fundamental and challenging task in applications such as navigation, autonomous driving and disaster response due to the vast diversity of locations, environmental conditions, and scene variations. Traditional retrieval-based methods for geo-localization struggle with scalability and perceptual aliasing, while classification-based approaches lack generalization and require extensive training data. Recent advances in vision-language models (VLMs) offer a promising alternative by leveraging contextual understanding and reasoning. However, while VLMs achieve high accuracy, they are often prone to hallucinations and lack interpretability, making them unreliable as standalone solutions. In this work, we propose a novel hybrid geo-localization framework that combines the strengths of VLMs with retrieval-based visual place recognition (VPR) methods. Our approach first leverages a VLM to generate a prior, effectively guiding and constraining the retrieval search space. We then employ a retrieval step, followed by a re-ranking mechanism that selects the most geographically plausible matches based on feature similarity and proximity to the initially estimated coordinates. We evaluate our approach on multiple geo-localization benchmarks and show that it consistently outperforms prior state-of-the-art methods, particularly at street (up to 4.51%) and city level (up to 13.52%). Our results demonstrate that VLM-generated geographic priors in combination with VPR lead to scalable, robust, and accurate geo-localization systems.
Sania Waheed, Na Min An, Michael Milford +2
Univ. of Southampton, UK · KAIST, South Korea · QUT, Australia +1