Scholarlink

The integration of Large Language Models (LLMs) into geospatial analysis presents significant opportunities for advancing spatial intelligence. However, questions remain regarding the optimal balance between predictive performance and computational efficiency. This study provides a comprehensive investigation into modern LLMs for geospatial reasoning tasks. We compare the state-of-the-art proprietary model GPT-4o against two fine-tuned lightweight open-weight models: Gemma-2B and DeepSeek-R1. To ensure robust evaluation, we conduct experiments on the complete 100,000-prompt GeoLLM benchmark, using the original fine-tuned LLaMA-7B as our primary baseline. Our experimental results demonstrate that GPT-4o achieves a new performance record with an average Pearson's r2 of 0.862. More significantly, the fine-tuned Gemma-2B model surpasses the larger LLaMA-7B baseline while maintaining exceptional computational efficiency with inference latencies under 0.6 s. These findings validate the effectiveness of parameter-efficient fine-tuning techniques, offering a practical path for deploying advanced geospatial AI in resource-constrained environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fine-Tuning Generative LLMs for Geographic Analysis: Comparing GPT-4o, Gemma-2B, and DeepSeek-R1

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kim-Son Nguyen,
  • The-Vinh Nguyen,
  • Huu-Huy Ngo,
  • Van-Viet Nguyen,
  • Thi Minh-Hue Luong,
  • Huu-Khanh Nguyen,
  • Duc-Binh Nguyen

摘要

The integration of Large Language Models (LLMs) into geospatial analysis presents significant opportunities for advancing spatial intelligence. However, questions remain regarding the optimal balance between predictive performance and computational efficiency. This study provides a comprehensive investigation into modern LLMs for geospatial reasoning tasks. We compare the state-of-the-art proprietary model GPT-4o against two fine-tuned lightweight open-weight models: Gemma-2B and DeepSeek-R1. To ensure robust evaluation, we conduct experiments on the complete 100,000-prompt GeoLLM benchmark, using the original fine-tuned LLaMA-7B as our primary baseline. Our experimental results demonstrate that GPT-4o achieves a new performance record with an average Pearson's r2 of 0.862. More significantly, the fine-tuned Gemma-2B model surpasses the larger LLaMA-7B baseline while maintaining exceptional computational efficiency with inference latencies under 0.6 s. These findings validate the effectiveness of parameter-efficient fine-tuning techniques, offering a practical path for deploying advanced geospatial AI in resource-constrained environments.