Geo-R1: Unlocking VLM Geospatial Reasoning with Cross-View Reinforcement Learning

This repository contains the Geo-R1 model, a reasoning-centric post-training framework that unlocks geospatial reasoning in vision-language models, as introduced in the paper:

Geo-R1: Unlocking VLM Geospatial Reasoning with Cross-View Reinforcement Learning

Geo-R1 combines "thinking scaffolding" (supervised fine-tuning on synthetic chain-of-thought exemplars) and an "elevating" stage using GRPO-based reinforcement learning on a weakly-supervised cross-view pairing proxy. This approach enables models to connect visual cues with geographic priors and harness reasoning for accurate prediction, achieving state-of-the-art performance across various geospatial reasoning benchmarks.