Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization

TL;DR

This work tackles image geolocalization by incorporating real-world maps into LVLM-based reasoning. It introduces Thinking with Map, a map-augmented agent that interacts with map tools in an agent-in-map loop, and pairs it with agentic reinforcement learning and parallel test-time scaling. Evaluations on MAPBench and other benchmarks show substantial improvements over open- and closed-source baselines, especially in fine-grained localization such as Acc@500m. The work highlights the importance of external, verifiable map knowledge for robust geolocalization and points to future directions in scalable, long-horizon map reasoning.

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.

Figures (6)

• Figure 1: (Up) The illustration of a complete Thinking with Map process. (Bottom) Comparison with up-to-date open- and closed-source models on three geolocalization benchmarks. Our method is built upon the model Qwen3-VL-30B-A3B. POI represents Point of Interest.
• Figure 2: The Thinking with Map trajectories from parallel sampling. The abundant map-API results make the trajectories easily verified based on their causal relationships.
• Figure 3: (a) The process of Thinking with Map, consists of an agent-in-the-map loop. During the loop, the agent implicitly maintains a candidate pool of hypotheses. (b) The agentic reinforcement learning for Thinking with Map. (c) The parallel test-time scaling with verifier pipeline for Thinking with Map.
• Figure 4: The comparison on parallel sampling.
• Figure 5: The evolution of pass@K accuracy across RL training steps on MAPBench.