Explicit geometric grounding enables data-efficient and interpretable geospatial reasoning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Explicit geometric grounding enables data-efficient and interpretable geospatial reasoningThis repository contains the official implementation, datasets, and demo examples for the paper "Explicit geometric grounding enables data-efficient and interpretable geospatial reasoning".LandAI is a family of multimodal large language models designed to bridge the gap between predictive performance and interpretability in Earth Observation.🌟 Key FeaturesExplicit Geometric Grounding: Enforces a "look-then-reason" mechanism where semantic conclusions must be supported by bounding-box evidence.GRPO-L1 Algorithm: A geometry-aware reinforcement learning approach that rewards valid coordinate generation and penalizes hallucinations.Data Efficiency: Achieves State-of-the-Art (SOTA) accuracy using only 25% of the training data required by comparable models (e.g., LandGPT).Robustness: Demonstrated significant resilience to visual-textual conflicts (e.g., misleading POI data).🛠️ PrerequisitesOur implementation relies on the ms-swift framework for efficient LLM training.You only need to install the ms-swift framework to easily run our code.We have listed the versions of some key libraries. ms_swift 3.6.0.dev0 vllm 0.8.5.post1 transformers 4.51.3 torch 2.6.0 torchaudio 2.6.0 torchvision 0.21.0 flash_attn 2.7.4.post1 📊 Data Preparation1.Training Data (landai_base_sft_dataset.jsonl)This file contains the composite corpus used to train LandAI-Base (Stage 1). It aggregates:Geo-Base-Thinking-14K: A distilled dataset of ~14.7k geospatial reasoning chains derived from geography examinations and textbooks.General Reasoning Corpus: Subsets from OpenR1-Math, Chinese-Data-Distill-From-R1 and OpenThoughts to activate logical reasoning capabilities.2.Demo Data (landai-test-image-demo/)We provide 5,000 representative samples from the MSLU test set to facilitate rapid evaluation. Each sample allows you to reproduce the multimodal analysis:images: High-resolution remote sensing imagery.model_output.jsonl: LandAI-L1's generated grounded reasoning (JSON bounding boxes + CoT).3.MSLU-100K Data The MSLU-100K dataset used in this article is now open source; you can find the raw data files used in our experiments without any processing here. 🤗 Hugging Face.🚀 Training & InferenceWe employ a two-stage training strategy using the scripts provided in landai-msswift/.Stage 1: Supervised Fine-Tuning (SFT) Activate domain-specific reasoning capabilities using the landai_base_sft_dataset.jsonl cd landai-msswift bash sft.sh Stage 2: Reinforcement Learning (GRPO) This is the core contribution of the paper. We provide scripts for both the proposed method and the baseline.Option A: Proposed Method (LandAI-L1) bash GRPO_L1.sh Option B: Ungrounded Baseline (LandAI-L1-Zero) bash GRPO.sh Please note that the parameters in the script file may need to be adjusted according to your local machine's characteristics. 🤖 Model ZooWe have synchronously open-sourced three variants of the LandAI model family on Hugging Face. This allows researchers to reproduce our results or initiate their work from any stage of the training pipeline .ModelStageDescriptionLinkLandAI-L1Stage 2 (Final)The proposed model featuring explicit geometric grounding. It achieves SOTA accuracy (86.41%) and superior hallucination resistance.🤗 Hugging FaceLandAI-L1-ZeroStage 2 (Baseline)The ungrounded experimental baseline trained via standard GRPO. Useful for comparative analysis of multimodal hallucinations and reasoning stability.🤗 Hugging FaceLandAI-BaseStage 1 (SFT)The foundation model fine-tuned on the landai_base_sft_dataset.jsonl dataset.🤗 Hugging Face