Enabling Intrinsic Reasoning over Dense Geospatial Embeddings with DFR-Gemma

2026-04-08

AI Summary

This work addresses the limitations of current large language models (LLMs) in handling dense geospatial embeddings, which typically rely on text-based conversion or retrieval indexing—approaches that introduce redundancy, computational inefficiency, and numerical distortion. To overcome these issues, the authors propose DFR-Gemma, a framework that employs a lightweight projector to directly align high-dimensional geospatial embeddings with the latent space of the Gemma model, enabling them to participate natively in reasoning as semantic tokens without textual mediation. This approach achieves the first end-to-end integration of dense geospatial embeddings into LLMs, significantly outperforming text-based baselines on a multi-task geospatial question-answering benchmark. Notably, DFR-Gemma accurately decodes spatial patterns even in zero-shot settings, demonstrating breakthroughs in both inference efficiency and accuracy.

Abstract

Representation learning for geospatial and spatio-temporal data plays a critical role in enabling general-purpose geospatial intelligence. Recent geospatial foundation models, such as the Population Dynamics Foundation Model (PDFM), encode complex population and mobility dynamics into compact embeddings. However, their integration with Large Language Models (LLMs) remains limited. Existing approaches to LLM integration treat these embeddings as retrieval indices or convert them into textual descriptions for reasoning, introducing redundancy, token inefficiency, and numerical inaccuracies. We propose Direct Feature Reasoning-Gemma (DFR-Gemma), a novel framework that enables LLMs to reason directly over dense geospatial embeddings. DFR aligns high-dimensional embeddings with the latent space of an LLM via a lightweight projector, allowing embeddings to be injected as semantic tokens alongside natural language instructions. This design eliminates the need for intermediate textual representations and enables intrinsic reasoning over spatial features. To evaluate this paradigm, we introduce a multi-task geospatial benchmark that pairs embeddings with diverse question-answer tasks, including feature querying, comparison, and semantic description. Experimental results show that DFR allows LLMs to decode latent spatial patterns and perform accurate zero-shot reasoning across tasks, while significantly improving efficiency compared to text-based baselines. Our results demonstrate that treating embeddings as primary data inputs, provides a more direct, efficient, and scalable approach to multimodal geospatial intelligence.

Problem

• geospatial embeddings
• Large Language Models
• intrinsic reasoning
• representation learning
• multimodal integration

Innovation

• geospatial embeddings
• direct feature reasoning
• large language models
• multimodal reasoning
• zero-shot inference