Large language models for environmental modeling: Framework, capabilities, constraints

Environmental modeling faces growing complexity, yet practical pathways for integrating Large Language Models (LLMs) remained limited. We introduce two practical, reproducible frameworks that embed LLMs into core modeling workflows: a human-AI collaborative "Copilot" and an LLM-driven automation "Autopilot". Using the Rainfall-Runoff-Inundation (RRI) model in Japan's Kuzuryu River basin, we evaluate both frameworks on parameter calibration and real-time correction. Copilot framework demonstrated robust performance, leveraging prompt engineering to guide LLMs in algorithm comprehension and code generation. It implemented advanced optimization and correction algorithms, achieving strong performance in parameter calibration (Nash-Sutcliffe Efficiency of 0.91 for calibration and 0.81 for validation) and delivering stable real-time correction. Autopilot framework showed competence in physics-constrained parameter calibration but failed in the long-sequence, data-intensive task of real-time correction, revealing instability caused by "attention decay". This stark contrast underscores the importance of strategic task design and human oversight. Our findings highlight that LLMs currently excel as powerful knowledge engines and coding assistants within human-supervised workflows (Copilot), whereas full automation (Autopilot) remains constrained by context window limitations and inherent weaknesses in processing long numerical sequences. This study provides a methodological framework, practical implementation guidance, and design principles (task decomposition, physics constraints, oversight checkpoints) to support generalizable deployment of LLMs in environmental modeling, and we outline testable directions for improving robustness, scalability, and reproducibility.