GeoVista: Web-Augmented Agentic Visual Reasoning for Geolocalization

TL;DR

GeoVista addresses real-world geolocalization by uniting visual reasoning with external information retrieval through tool-augmented reasoning. It introduces GeoBench, a high-resolution, globally distributed benchmark with level-wise and nuanced evaluation, and a training pipeline combining cold-start supervised fine-tuning with reinforcement learning using a hierarchical reward. GeoVista achieves state-of-the-art performance among open models and approaches closed-source systems on key metrics, demonstrating the value of crop-zoom and web-search tools for agentic multimodal reasoning. This work lays a foundation for robust, scalable real-world geographic reasoning and provides a practical blueprint for future research in vision-language agents with external tooling.

Abstract

Current research on agentic visual reasoning enables deep multimodal understanding but primarily focuses on image manipulation tools, leaving a gap toward more general-purpose agentic models. In this work, we revisit the geolocalization task, which requires not only nuanced visual grounding but also web search to confirm or refine hypotheses during reasoning. Since existing geolocalization benchmarks fail to meet the need for high-resolution imagery and the localization challenge for deep agentic reasoning, we curate GeoBench, a benchmark that includes photos and panoramas from around the world, along with a subset of satellite images of different cities to rigorously evaluate the geolocalization ability of agentic models. We also propose GeoVista, an agentic model that seamlessly integrates tool invocation within the reasoning loop, including an image-zoom-in tool to magnify regions of interest and a web-search tool to retrieve related web information. We develop a complete training pipeline for it, including a cold-start supervised fine-tuning (SFT) stage to learn reasoning patterns and tool-use priors, followed by a reinforcement learning (RL) stage to further enhance reasoning ability. We adopt a hierarchical reward to leverage multi-level geographical information and improve overall geolocalization performance. Experimental results show that GeoVista surpasses other open-source agentic models on the geolocalization task greatly and achieves performance comparable to closed-source models such as Gemini-2.5-flash and GPT-5 on most metrics.

Figures (9)

• Figure 1: Agentic thinking of GeoVista for Real-world geolocalization. During the reasoning loop, our GeoVista seamlessly integrates the image–zoom-in tool to magnify regions of interest and the web-search tool to retrieve relevant information. This web-augmented visual reasoning process enables GeoVista validate or refine its geolocalization judgments.
• Figure 2: Image examples from GeoBench and the training data, and the agentic pipeline of GeoVista. Given a query and image, the policy model iteratively generates thoughts and actions; each action is parsed, executed, and yields a new observation, repeating this loop until it outputs a final geolocation prediction or reaches the maximum interaction turn limit.
• Figure 3: Localizable vs Non-Localizable. We remove the non-localizable (orange) and the landmarks (purple) from GeoBench, leaving only localizable images for rigorously evaluating models.
• Figure 4: Left: The evaluation pipeline of GeoBench dataset. The evaluation system consists of (1) Level-wise evaluation, which employs both rule-based and model-based verifiers to determine correctness at different administrative levels, and (2) nuanced evaluation, which extracts the predicted address, applies geocoding to obtain the predicted geolocalization point, and computes the haversine distance to the ground-truth location. Right: Geological distribution of GeoBench. GeoBench is a high-resolution, multi-source, globally annotated dataset to evaluate models’ general geolocalization ability.
• Figure 5: Illustration of GeoBench dataset, along with level-wise and nuanced evaluation.