Learning to Wander: Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning

Yushuo Zheng; Huiyu Duan; Zicheng Zhang; Xiaohong Liu; Xiongkuo Min

cs.CVMar 11, 2026

Learning to Wander: Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning

Authors: Yushuo Zheng, Huiyu Duan, Zicheng Zhang, Xiaohong Liu, Xiongkuo Min

Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory

Abstract

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.

Figures & tables

Dataset	Size	Geographic Coverage	Source	Open Access	Navigation Support	Bidirectional Evaluation
Google-WS [ 11 ]	15k	Global	Map Service	✗	✗	✗
GMCP [ 40 ]	105K	Local	Map Service	✗	✗	✗
StreetCLIP [ 17 ]	1M	Unknown	Map Service	✗	✗	✗
Im2GPS [ 19 ]	237	Local	Web-Scraped	✓	✗	✗
Im2GPS3K [ 36 ]	2997	Local	Web-Scraped	✗	✗	✗
YFCC4K citevo2017	4536	Local	Web-Scraped	✗	✗	✗

Table 1: Comparison of geolocation datasets. WanderBench is the only openly available, globally scoped dataset providing both navigation support and bidirectional geolocation evaluation.

Figure 2: Figure 2. (a) Global distribution of model-proposed locations across all continents. (b) Country-level distribution of the proposed locations. (c) Average navigation-graph structure of the WanderBench dataset.

Model	$O_{16}$	$H_{16}$	$A_{\text{hull}}$	$R_{\text{CE}}$	$\bar{r}_{\text{NN}}$
Model	( $\times 10^{-2}$ )	( $\times 10^{-1}$ )	( $\times 10^{-2}$ )	( $\times 10^{-1}$ )	( $\times 10^{-2}$ )
InternVL2-76B [ 9 ]	2.73	3.10	8.61	3.21	4.63
InternVL2.5-38B [ 8 ]	1.82	2.50	5.12	1.27	1.84
InternVL2.5-78B [ 8 ]	1.95	2.58	4.31	1.63	2.35
InternVL3-38B [ 32 ]	2.02	2.66	7.16	1.95	2.82
InternVL3-78B [ 32 ]	2.08	2.62	5.40	1.97	2.84

Table 2: Spatial diversity metrics of model-proposed locations. All coordinates normalized to

[0,1]^{2}

per continent.

O_{16}

: occupancy on

16{\times}16

grid;

H_{16}

: entropy;

A_{\text{hull}}

: convex-hull area;

R_{\text{CE}}

: Clark–Evans ratio;

\bar{r}_{\text{NN}}

: mean nearest-neighbor distance.

Figure 3: Visualization of the average navigation graph structure across all locations in the WanderBench dataset. Node sizes correspond to occurrence frequency within spatial bins, while edge thickness indicates transition frequency between bins.

Model Type	Model	Overall		Street			City			Country
		Dist. $\downarrow$	Score $\uparrow$	Acc. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$
Retrieval Based	GeoCLIP [ 35 ]	1925	65.81	0.158	0.158	0.158	0.302	0.302	0.302	0.611	0.611	0.611
	GeoCLIP-top5	1403	73.55	0.232	0.233	0.233	0.392	0.392	0.392	0.700	0.700	0.700
Open Source LMMs	InternVL2.5-78B [ 8 ]	1124	74.69	0.164	0.141	0.152	0.368	0.316	0.340	0.723	0.668	0.694
	InternVL3-78B [ 32 ]	1441	72.77	0.171	0.148	0.159	0.359	0.309	0.332	0.697	0.646	0.670
	Qwen2VL-72B [ 37 ]	988.7	83.46	0.270	0.233	0.250	0.525	0.451	0.485	0.837	0.813	0.825

Table 3: Baseline performance of models on the benchmark. Lower is better for Distance (km), higher is better for all other metrics. Blue bold values represent the best within the open-source model type. Red bold values represent the best performance across all model types (closed-source).

Model Type	Model Name	$F\,\uparrow$ Stat.	$p\,\downarrow$ $(10^{-2})$	$\rho\,\uparrow$ Spear	$p\,\downarrow$ Spear	Slope $\uparrow$ OLS	$p\,\downarrow$ OLS	Avg. Score Hard	Avg. Score Medium	Avg. Score Easy	Remark
Open Source	InternVL2-76B [ 9 ]	7.62	0.10	1.00	0.00	0.30	0.08	64.31	75.14	82.15	$\checkmark$
	InternVL2.5-38B [ 8 ]	2.47	0.92	1.00	0.00	0.15	0.28	75.41	76.29	84.64	$\checkmark$
	InternVL2.5-78B [ 8 ]	0.22	8.03	0.50	0.67	0.05	0.44	74.44	77.73	77.17	✗
	InternVL3-38B [ 32 ]	2.81	0.67	1.00	0.00	0.15	0.24	78.52	80.03	87.80	$\checkmark$
	InternVL3-78B [ 32 ]	0.55	5.82	1.00	0.00	0.07	0.27	75.64	79.60	80.07	$\checkmark$
	Phi-4 [ 1 ]	0.97	3.83	1.00	0.00	0.13	0.04	71.97	75.32	79.61	$\checkmark$

Table 4: ANOVA and trend analysis across coverage levels (20%, 50%, 80%). Group 1–3 correspond to means at 20%, 50%, 80%.

Figure 4: An overview of GeoAoT. Given an input image, GeoAoT first leverages a pre-trained LMMs to generate an initial geo-guess. It then iteratively refines this estimate through AoT based multi-turn interactions, where the model reasons about uncertainty and issues actions to gather more evidence, operating purely at inference time without any additional training.

Method	Overall		Street			City			Country
	Dist. $\downarrow$	Score $\uparrow$	Acc. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$
LLaVA OneVision 72B	4081	53.68	0.112	0.097	0.104	0.241	0.206	0.222	0.508	0.482	0.494
(w/ AoT)	3007	61.20	0.150	0.130	0.139	0.304	0.261	0.281	0.579	0.544	0.561
$\Delta$ (AoT $-$ base)	-1074	7.527	$0.038$	$0.033$	$0.035$	$0.063$	$0.055$	$0.058$	0.071	0.062	0.067
Gemma3 27B	1685	72.76	0.207	0.178	0.191	0.426	0.366	0.393	0.685	0.681	0.683
(w/ AoT)	1475	74.95	0.228	0.196	0.211	0.442	0.380	0.409	0.711	0.704	0.708

Table 5: Performance of GeoAoT on the benchmark. Blue bold values represent the best AoT result. Red bold values represent the strongest improvement per column.

Multimodal Large Language Models

Multimodal large language models combine language models with image, video, audio, and other inputs for grounded understanding and generation.

See all 1,069 papers

ERGeoBench:A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models
May 29, 2026Kaiwen Xue, Tao Wei, Guoxin Zhang +5Multimodal Large Language Models Visual Spatial Reasoning
Beijing University of Posts and Telecommunications, Beijing, China · State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China · School of Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China +2
GeoAgent: Evaluating VLM Geolocalization Through Embodied Navigation
Aug 30, 2026Arka Mukherjee, Soham Roy, Kartikeya Trivedi +1Geospatial Reasoning Benchmarks VLM Reasoning
KIIT Bhubaneswar · IIT Bhubaneswar
Locatability-Guided Adaptive Reasoning for Image Geo-Localization with Vision-Language Models
Mar 13, 2026Bo Yu, Fengze Yang, Yiming Liu +6VLM Reasoning Cross-View Geo-Localization
The University of Utah · The University of Oklahoma · Worcester Polytechnic Institute +2

May 29, 2026J/K move · Enter open · S save

ERGeoBench:A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models

Kaiwen Xue, Tao Wei, Guoxin Zhang, Zhonghong Ou+4

Beijing University of Posts and Telecommunications, Beijing, China · State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China · School of Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China+2

Multimodal Large Language Models Visual Spatial Reasoning Multimodal Model Evaluation Spatial Reasoning Benchmarks Geospatial Reasoning Visually Grounded Reasoning+1

Open paper AskPDF arXiv

Multimodal large language models (MLLMs) have shown strong potential as embodied agents, yet embodied geo-localization remains underexplored due to the lack of fine-grained evaluation. We introduce ERGeoBench, a diagnostic benchmark for vision-driven embodied geo-localization. ERGeoBench evaluates models under three progressive settings -- single-view, panorama-view, and embodied-view -- where agents may actively acquire observations through sequential changes in yaw, pitch, and zoom. The benchmark contains 2,207 globally distributed street-view panoramas and measures four complementary capabilities: foundational perception, spatial awareness, common sense reasoning, and geo-localization reasoning. Evaluations of leading proprietary and open-source MLLMs show that current models can infer high-level geographic semantics, but still struggle with fine-grained perceptual operations, metric localization, and spatial consistency across views. We further observe that geo-localization is strongly correlated with the other capability dimensions, suggesting that accurate localization depends on integrated perception, spatial reasoning, and commonsense inference rather than isolated visual recognition. Overall, ERGeoBench provides a unified framework for diagnosing and advancing human-like embodied geo-localization. Project Page: https://kaixuewen.github.io/ERGeoBench/

Abstract

Multimodal Large Language Models

Explore similar work

ERGeoBench:A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models

GeoAgent: Evaluating VLM Geolocalization Through Embodied Navigation

Locatability-Guided Adaptive Reasoning for Image Geo-Localization with Vision-Language Models