This paper presents a novel solution to tackle the challenges that posed by the abundance of non-standard addresses, which input by users in modern applications such as navigation maps, ride-hailing apps, food delivery platforms, and logistics services. These manually entered addresses often contain irregularities, such as missing information, spelling errors, colloquial descriptions, and directional offsets, which hinder address-related tasks like address matching and linking. To tackle these challenges, we propose GeoAgent, a new framework comprising two main components: a large language model (LLM) and a suite of geographical tools. By harnessing the semantic understanding capabilities of the LLM and integrating specific geospatial tools, GeoAgent incorporates spatial knowledge into address texts and achieves efficient address standardization. Further, to verify the effectiveness and practicality of our approach, we construct a comprehensive dataset of complex non-standard addresses, which fills the gaps in existing datasets and proves invaluable for training and evaluating the performance of address standardization models in this community. Experimental results demonstrate the efficacy of GeoAgent, showcasing substantial improvements in the performance of address-related models across various downstream tasks.

Multimodal Large Language Models (MLLMs) have shown promising results in various tasks, but their ability to perceive the visual world with deep, hierarchical understanding similar to humans remains uncertain. To address this gap, we introduce CONSTRUCTURE, a novel concept-level benchmark to assess MLLMs’ hierarchical concept understanding and reasoning abilities. Our goal is to evaluate MLLMs across four key aspects: 1) Understanding atomic concepts at different levels of abstraction; 2) Performing upward abstraction reasoning across concepts; 3) Achieving downward concretization reasoning across concepts; and 4) Conducting multi-hop reasoning between sibling or common ancestor concepts. Our findings indicate that even state-of-the-art multimodal models struggle with concept structure reasoning (e.g., GPT-4o averages a score of 62.1%). We summarize key findings of MLLMs in concept structure reasoning evaluation. Morever, we provide key insights from experiments using CoT prompting and fine-tuning to enhance their abilities.