sahitiy51/geochain

GeoChain: Multimodal Chain-of-Thought for Geographic Reasoning

[Paper on arXiv] [Code on GitHub]

GeoChain is a large-scale benchmark introduced for evaluating step-by-step geographic reasoning in multimodal large language models (MLLMs). Leveraging 1.46 million Mapillary street-level images, GeoChain pairs each image with a 21-step chain-of-thought (CoT) question sequence, resulting in over 30 million Q&A pairs. These sequences are designed to guide models from coarse attributes to fine-grained localization, covering four key reasoning categories: visual, spatial, cultural, and precise geolocation, with annotations for difficulty. Images within the dataset are also enriched with semantic segmentation (150 classes) and a visual locatability score. Our benchmarking of contemporary MLLMs reveals consistent challenges: models frequently exhibit weaknesses in visual grounding, display erratic reasoning, and struggle to achieve accurate localization, especially as reasoning complexity escalates. GeoChain offers a robust diagnostic methodology, critical for fostering significant advancements in complex geographic reasoning within MLLMs.

How to Use

The dataset can be loaded using the Hugging Face datasets library:

from datasets import load_dataset

# Load the mini_test split for quick experiments
mini_dataset = load_dataset("sahitiy51/geochain", split="mini_test")

# Load the full test split
full_dataset = load_dataset("sahitiy51/geochain", split="test")

print(mini_dataset[0])

Dataset Structure

This dataset provides two main splits for evaluation:

mini_test Split

A smaller subset for quick evaluation runs.

Features:

• image: A PIL Image object representing the street-level image.
• locatability_score: (float) The visual locatability score of the image.
• lat: (float) Latitude of the image.
• lon: (float) Longitude of the image.
• class_mapping: (string) Associated class mapping.
• sequence_key: (string) Unique sequence identifier.

test Split

The full-scale test set for comprehensive evaluation.

Features:

• key: (string) Unique identifier for the image.
• locatability_score: (float) The visual locatability score.
• lat: (float) Latitude of the image.
• lon: (float) Longitude of the image.
• city: (string) City where the image was taken.
• sub_folder: (string) Sub-folder information related to image storage/organization.
• class_mapping: (string) Associated class mapping.
• sequence_key: (string) Unique sequence identifier.
• image: This feature is None for the test split, as this split primarily provides metadata.

Citation

If you find our work useful, please cite the following paper:

@misc{yerramilli2025geochainmultimodalchainofthoughtgeographic,
      title={GeoChain: Multimodal Chain-of-Thought for Geographic Reasoning}, 
      author={Sahiti Yerramilli and Nilay Pande and Rynaa Grover and Jayant Sravan Tamarapalli},
      year={2025},
      eprint={2506.00785},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={[https://arxiv.org/abs/2506.00785](https://arxiv.org/abs/2506.00785)}, 
}