Deep research agents integrate fragmented evidence through multi-step tool use. BrowseComp offers a text-only testbed for such agents, but existing multimodal benchmarks rarely require both weak visual cues composition and BrowseComp-style multi-hop verification. Geolocation is a natural testbed because answers depend on combining multiple ambiguous visual cues and validating them with open-web evidence. Thus, we introduce GeoBrowse, a geolocation benchmark that combines visual reasoning with knowledge-intensive multi-hop queries. Level 1 tests extracting and composing fragmented visual cues, and Level 2 increases query difficulty by injecting long-tail knowledge and obfuscating key entities. To support evaluation, we provide an agentic workflow GATE with five think-with-image tools and four knowledge-intensive tools, and release expert-annotated stepwise traces grounded in verifiable evidence for trajectory-level analysis. Experiments show that GATE outperforms direct inference and open-source agents, indicating that no-tool, search-only or image-only setups are insufficient. Gains come from coherent, level-specific tool-use plans rather than more tool calls, as they more reliably reach annotated key evidence steps and make fewer errors when integrating into the final decision. The GeoBrowse bernchmark and codes are provided in https://github.com/ornamentt/GeoBrowse
Figures & tables
Figure 1: GeoBrowse couples a tool-use framework with a geolocation benchmark: Level 1 emphasizes visual cue composition, while Level 2 contains BrowseComp-style queries, all paired with expert-annotated stepwise traces.
Dataset
Geographic Coverage
Human Annotation
Reasonable Localizability
CoT
Agentic Tool-use
MP-16 [ 22 ]
Local
✗
✗
✗
✗
GLDv2 [ 50 ]
Local
✗
✗
✗
✗
VIGOR [ 59 ]
Local
✗
✗
✗
✗
Google-WS [ 8 ]
Global
✗
✗
✗
✗
OSV [ 2 ]
Global
✗
✗
✗
✗
GeoComp [ 43 ]
Global
✓
❍
✗
✗
Table 1: Comparison of existing geolocation datasets with GeoBrowse. “Local” denotes region-level coverage, whereas “Global” spans multiple continents. ✓ and ✗ indicate the presence and absence of a feature, respectively; ❍ indicates partial availability.
Figure 3
Figure 5: The pipeline of GATE, our proposed Geolocation Agentic-workflow with Tool Enhancement approach. The input image is first registered into stable img_id references. GATE then follows a ReAct-style loop: <Think> summarizes the latest evidence and plans the next step, <Action> invokes an image or knowledge tool, and the tool response is returned as <Obs> to update the agent state. Any new images in <Obs> are re-registered, and the loop repeats until the final answer.
Backbone
Level 1: Visual Cues
Level 2: BrowseComp-style
City
State
Country
Avg.
City
State
Country
Avg.
Direct Inference
GPT-4o
5.8
16.4
35.7
23.1
10.5
15.8
11.1
11.9
GPT-4.1
5.8
14.5
30.4
19.6
5.3
5.3
15.9
11.9
GPT-5
11.5
21.8
38.0
27.6
15.8
10.8
20.6
17.8
GPT-5.2
7.7
9.1
30.4
18.6
21.1
21.1
20.6
20.8
Table 3: Main results on GeoBrowse. All accuracy scores are reported as percentages. Each entry is averaged over 3 runs. Avg. is the macro-average over city/state/country.
Figure 6: Tool-use distribution on GeoBrowse. Statistics are aggregated over all tool calls produced by GATE with the Gemini-3-Pro backbone, across Level 1 (geolocation tasks) and Level 2 (multi-step reasoning tasks requiring external knowledge).
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Subset
N
Agree(%)
κcountry
κstate
κcity
κtracew
Level 1
199
82.0
0.93
0.86
0.72
0.62
Level 2
101
76.0
0.91
0.82
0.66
0.55
Macro avg. (L1/L2)
300
79.0
0.92
0.84
0.69
0.59
Appendix
Table 8: Inter-annotator agreement (IAA) on a blinded double-annotated subset. Agree(%) is strict city-level exact match. κw denotes weighted Cohen’s kappa (quadratic weights) for ordinal 1–4 trace-quality ratings (verifiability/completeness/coherence averaged).
Judge
Agreement (%)
95% CI
Cohen’s κ
GPT-4o
95.0
[91.9, 97.9]
0.90
Claude-3.5-Sonnet
91.1
[83.9, 95.2]
0.82
Gemini-2.5-Pro
94.8
[90.9, 96.5]
0.89
Qwen2.5-72B-Instruct
93.1
[86.4, 96.6]
0.86
Appendix
Table 9: Inter-rater agreement between LLM judges and human experts on GeoBrowse (binary 0–1 labels, N=300 ). 95% Wilson confidence intervals are reported for raw agreement.
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
Open-world geo-localization requires models to reason over ambiguous visual cues through multi-step reasoning and external knowledge grounding. While recent large vision-language models exhibit strong multimodal reasoning capabilities, existing approaches still suffer from perceptual hallucination and context drift due to the lack of explicit evidence-grounded verification. In this work, we reformulate geo-localization as a human-like perceive-then-verify reasoning problem and propose GeoPAVE (Geo-localization Perception-and-Verification-Engine), a bi-level agentic framework that contains perception-based hypothesis generation via single-pass rollouts and verification-based evidence grounding for decision actions: support, refute, and refine. To support rigorous evaluation, we further introduce PAVED, a novel dataset derived from real-world user check-in data, equipped with comprehensive reasoning trajectories featuring multi-hop queries, multi-round tool invocations, and structured perception-verification traces. The dataset and code are available at https://github.com/Arandinglv/GeoPAVE.
Yutian Jiang, Ruijie Li, Sisuo Lyu +4
The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology
Google Earth AI is a family of geospatial AI models, together with an agent that orchestrates them. We present this agent, Google Earth AI's Geospatial Reasoning agent (GRA), a Gemini-powered research system that combines geospatial data sources, foundation models, and tools to answer questions about our planet. GRA organizes these into modular domain skills and pairs them with upfront geospatial planning to decompose complex queries into multi-stage workflows. Evaluating such systems is hard: real-world problems admit multiple valid workflows, require data discovery across large catalogs, and rely on evolving ground truth. We introduce two benchmarks and release their development splits. GeoExpertQueries poses complex analytical problems grounded in practitioner use cases and refined by geospatial experts; each is paired with a consensus golden workflow from three independent experts and scored by a human-validated, tool-agnostic autorater. GeoKTools consists of templated queries with deterministic reference solvers that systematically vary the number of required tool types. We also use two external benchmarks, CBGB and UnivEARTH; the latter we find to be saturated. To isolate the contribution of GRA's tools and its agentic orchestration, we compare four configurations, each with Gemini 3.1 Pro and 3.8 Flash: a standalone LLM, the LLM with basic tools, the LLM with GRA's skills and tools but without its upfront planning and execution guidance, and GRA itself. With at most basic tools, an advanced model recovers much of the answer on templated queries but satisfies only a small fraction of the expert workflows on open-ended problems; GRA performs well on both. The largest step comes from GRA's skills and tools; on open-ended problems, its planning and execution guidance add a further gain with Gemini 3.8 Flash, letting Flash match or exceed Pro at lower cost.