Paper 2601.05432
Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 8
- GitHub
- 177 stars
01 In brief
Summary
The paper introduces Thinking with Map, a map-augmented agent for image geolocalization that equips a large vision-language model (LVLM) with map tools (POI search, static/satellite map queries, image zoom) to iteratively propose and verify location hypotheses.
The method uses a two-stage optimization: agentic reinforcement learning (GRPO) to improve sampling efficiency, followed by parallel test-time scaling (TTS) with a verifier to aggregate multiple trajectories.
The authors also present MAPBench, a new benchmark of 5,000 up-to-date Chinese street-view images, and evaluate on IMAGEO-Bench and GeoBench.
Results show the method outperforms open- and closed-source models, improving Acc@500m from 8.0% to 22.1% compared to Gemini-3-Pro with Google Search/Map grounded mode.
Ablations show map tools significantly boost fine-grained localization, RL reduces variance and improves pass@K, and larger verifiers help with more parallel samples.
02 From the paper
Abstract
The image geolocalization task aims to predict the location where an image was taken anywhere on Earth using visual clues. Existing large vision-language model (LVLM) approaches leverage world knowledge, chain-of-thought reasoning, and agentic capabilities, but overlook a common strategy used by humans -- using maps. In this work, we first equip the model \textit{Thinking with Map} ability and formulate it as an agent-in-the-map loop. We develop a two-stage optimization scheme for it, including agentic reinforcement learning (RL) followed by parallel test-time scaling (TTS). The RL strengthens the agentic capability of model to improve sampling efficiency, and the parallel TTS enables the model to explore multiple candidate paths before making the final prediction, which is crucial for geolocalization. To evaluate our method on up-to-date and in-the-wild images, we further present MAPBench, a comprehensive geolocalization training and evaluation benchmark composed entirely of real-world images. Experimental results show that our method outperforms existing open- and closed-source models on most metrics, specifically improving Acc@500m from 8.0\% to 22.1\% compared to \textit{Gemini-3-Pro} with Google Search/Map grounded mode.