LocationAgent: A Hierarchical Agent for Image Geolocation via Decoupling Strategy and Evidence from Parametric Knowledge
Qiujun Li, Zijin Xiao, Xulin Wang, Zhidan Ma, Cheng Yang, Haifeng Li
TL;DR
This work reframes image geolocation as an abductive reasoning problem involving hypothesis verification, addressing hallucinations from static memory by decoupling reasoning from evidence verification. It introduces LocationAgent with a Reasoner-Executor-Recorder (RER) architecture and a suite of external evidence tools to guide multi-step localization, formalized through updates such as $a_{t+1}= \\mathcal{R}(I, \\mathcal{A}_{tool}, S_{1:t})$, $e_{t+1}= \\mathcal{V}(a_{t+1}, \\mathcal{W})$, and $L_{t+1}= \\{ l \\in L_t \\mid \\text{Consistent}(l, e_{t+1}) \\}$. A new China-focused benchmark, CCL-Bench, mitigates data leakage and China-region data scarcity, enabling rigorous evaluation of reasoning under realistic conditions. Empirical results show LocationAgent outperforms state-of-the-art baselines by large margins, achieving over 50% accuracy at 1 km and 100% beyond 200 km, with high city-level reliability and robust reasoning trajectories. Overall, the approach offers improved generalization for open-world image geolocation and practical applicability to region-specific localization tasks, with future work expanding tool design and agent scalability.
Abstract
Image geolocation aims to infer capture locations based on visual content. Fundamentally, this constitutes a reasoning process composed of \textit{hypothesis-verification cycles}, requiring models to possess both geospatial reasoning capabilities and the ability to verify evidence against geographic facts. Existing methods typically internalize location knowledge and reasoning patterns into static memory via supervised training or trajectory-based reinforcement fine-tuning. Consequently, these methods are prone to factual hallucinations and generalization bottlenecks in open-world settings or scenarios requiring dynamic knowledge. To address these challenges, we propose a Hierarchical Localization Agent, called LocationAgent. Our core philosophy is to retain hierarchical reasoning logic within the model while offloading the verification of geographic evidence to external tools. To implement hierarchical reasoning, we design the RER architecture (Reasoner-Executor-Recorder), which employs role separation and context compression to prevent the drifting problem in multi-step reasoning. For evidence verification, we construct a suite of clue exploration tools that provide diverse evidence to support location reasoning. Furthermore, to address data leakage and the scarcity of Chinese data in existing datasets, we introduce CCL-Bench (China City Location Bench), an image geolocation benchmark encompassing various scene granularities and difficulty levels. Extensive experiments demonstrate that LocationAgent significantly outperforms existing methods by at least 30\% in zero-shot settings.