Organizations: Southwest Jiaotong University Chengdu, China · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Zhejiang University Hangzhou, China
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 Existing methods predominantly rely on implicit parametric memory, which often exploits outdated knowledge and generates hallucinated reasoning.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.
Figures & tables
Figure 1. The evolution of global visual geo-localization paradigms. Unlike traditional feature-based matching and static LLM reasoning, our proposed GeoSkill utilizes an evolving Skill-Graph to iteratively refine geographic knowledge through autonomous feedback loops. A double-column diagram illustrating the evolution of visual geo-localization paradigms, from feature-based matching and static LLM reasoning to the proposed GeoSkill framework based on an evolving Skill-Graph.
Figure 2. The overall framework of GeoSkill . The system follows a four-step workflow: (1) Offline Skill Library Initialization , which distills human heuristics into a reusable skill library; (2) Online Inference and Rollout , where the LVLM agent retrieves relevant skills and dynamically constructs a task-specific Skill-Graph; (3) Offline Logging and Validation , which records reasoning trajectories and verifies them with external tools; and (4) Offline Skill & Graph Updating , which refines and expands the skill library based on outcome feedback.
Methods
Im2GPS 3k
EarthWhere
GeoRC
10km
25km
200km
750km
2000km
10km
25km
200km
750km
2000km
10km
25km
200km
750km
2000km
GeoCLIP
0.277
0.322
0.486
0.665
0.807
0.085
0.110
0.200
0.333
0.610
0.082
0.096
0.188
0.412
0.605
GAEA
0.283
0.335
0.489
0.665
0.824
0.133
0.163
0.270
0.438
0.650
0.115
0.142
0.245
0.510
0.680
GeoReasoner
0.218
0.242
0.347
0.506
0.661
0.075
0.088
0.200
0.338
0.535
0.085
0.102
0.215
0.480
0.650
GLOBE
0.166
0.194
0.283
0.423
0.581
0.018
0.018
0.050
0.123
0.358
0.025
0.035
0.085
0.210
0.420
GeoVista
0.252
0.281
0.359
0.446
0.517
0.088
0.108
0.165
0.225
0.333
0.109
0.136
0.203
0.452
0.661
Table 1. Comparison of geolocation accuracy under different error thresholds on Im2GPS 3k, EarthWhere, and GeoRC. Best results are in bold.
Methods
Precision
Recall
F1
GeoCLIP
51.84
47.26
48.93
GAEA
53.12
46.88
49.27
GeoReasoner
47.85
47.68
46.47
GLOBE
43.27
40.94
41.08
GeoVista
57.28
50.73
53.67
Ours
58.64
63.18
60.31
Table 2. Comparison of precision, recall, and F1 score.
Iteration
#Skills
10km
25km
200km
Precision
Recall
F1
T=0
1080
0.108
0.125
0.280
52.10
56.05
54.00
T=1
1350
0.122
0.142
0.315
54.88
60.42
57.51
T=2
1510
0.135
0.158
0.334
57.24
61.30
59.20
T=3
1425
0.141
0.165
0.342
58.64
63.18
60.31
Table 3. Evolution of the Skills across autonomous optimization iterations.
Ablation
10km
25km
200km
Precision
Recall
F1
w.o. skill
0.091
0.109
0.244
55.70
58.50
56.88
random_skill
0.096
0.118
0.271
56.40
59.00
57.56
shuffled_order
0.129
0.154
0.321
58.20
61.50
59.71
atomic_only
0.114
0.146
0.307
57.40
60.10
58.42
Ours
0.141
0.165
0.342
58.60
63.20
60.31
Table 4. Ablation study of different skill settings.
Figure 3. Case study of GeoSkill.
Figure 4. Sensitivity analysis of multi-step rollouts. (a) Distance-threshold accuracy across varying rollout steps. (b) The reasoning faithfulness (Precision, Recall, F1) and inference runtime per sample.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Field
Description
skill_text
Reusable natural-language geographic rule
region_hint
Coarse region applicability prior
confidence
Historical reliability score
visual_cues
Key visual triggers associated with the skill
source_game_id
Source game identifier for provenance
source_round
Source round identifier for provenance
Appendix
Table 6. Structured representation of an atomic skill in GeoSkill.
Iteration
#Skills
Main effect
T=0
1080
Expert-initialized cold-start skill library
T=1
1350
Skill synthesis expands coverage
T=2
1510
Continued accumulation of corrective heuristics
T=3
1425
Merge and prune improve compactness and quality
Appendix
Table 7. Illustrative evolution dynamics of the skill library.
Figure 5. Failure analysis of a representative Kazakhstan–Kyrgyzstan confusion case.
Figure 6. Failure analysis of a representative Oman–UAE confusion case.
Component
Setting
Online inference model
Qwen3.5-122B-A10B
Offline evolution model
GPT-5.2
Initial skill count
1080
Retriever
BM25 + dense semantic retrieval
Top- k retrieval
7
Main reasoning temperature
0.2
Appendix
Table 8. Core implementation settings used in GeoSkill.
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.
Bo Yu, Fengze Yang, Yiming Liu +6
The University of Utah · The University of Oklahoma · Worcester Polytechnic Institute +2
Geolocation, the task of identifying the geographic location of an image, requires abundant world knowledge and complex reasoning abilities. Though advanced large multimodal models (LMMs) have shown superior aforementioned capabilities, their performance on the geolocation task remains unexplored. To this end, we introduce WanderBench, the first open access global geolocation benchmark designed for actionable geolocation reasoning in embodied scenarios. WanderBench contains over 32K panoramas across six continents, organized as navigable graphs that enable physical actions such as rotation and movement, transforming geolocation from static recognition into interactive exploration. Building on this foundation, we propose GeoAoT (Action of Thought), a Geolocation framework with Action of Though, which couples reasoning with embodied actions. Instead of generating textual reasoning chains, GeoAoT produces actionable plans such as, approaching landmarks or adjusting viewpoints, to actively reduce uncertainty. We further establish an evaluation protocol that jointly measures geolocation accuracy and difficulty-aware geolocation questioning ability. Experiments on 19 large multimodal models show that GeoAoT achieves superior fine-grained localization and stronger generalization in dynamic environments. WanderBench and GeoAoT define a new paradigm for actionable, reasoning driven geolocation in embodied visual understanding.
Yushuo Zheng, Huiyu Duan, Zicheng Zhang +2
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory
Geospatial reasoning requires solving image-grounded problems over the complex spatial structure of a scene. However, developing this capability is hindered by the cost of annotating a vast and combinatorial question space. We propose GeoX, a self-play framework that acquires spatial logic through executable programs that yield verifiable rewards, without relying on large-scale human-curated data Given a satellite or aerial image, our framework employs a single multimodal policy that proposes spatial problems as executable programs and solves them under three reasoning modes-abduction, deduction, and induction-over spatial primitives and an image understanding tool. A verifier executes each program to covert a reward signal that jointly optimizes the two roles via reinforcement learning. GeoX consistently improves its base VLMs by up to 5.5 points on average, matching or exceeding conventional baselines trained on millions of curated data. Along-side the proposed method, we release a benchmark for geospatial understanding accumulated through self-play.
Kyeongjin Ahn, Seungeon Lee, Krishna P. Gummadi +1