Putting Large Language Models into Environmental Context

Abstract The emergence of large language models (LLMs) has fundamentally transformed information interaction, yet their application in environmental science and engineering (ESE) is hindered by multimodal data formats, dynamic context requirements, and hallucination risks. To transition from general-purpose models to domain expert agentic systems, this Perspective introduces a three-pillar framework for ESE-LLMs: domain perception, knowledge grounding, and action capability. Implementing these capabilities introduces potential pitfalls such as geographic representation bias, catastrophic forgetting, and source authenticity vulnerabilities, which require strategic model blending and careful data curation to overcome. Furthermore, rigorous evaluation must extend beyond standard knowledge benchmarks to systematically assess factuality, fairness, consistency, and safety using both human and automated metrics. Finally, we present a decision-guided deployment workflow that helps researchers in ESE fields integrate LLMs into their workflow. Ultimately, achieving trustworthy environmental intelligence relies on continuous evaluation, robust governance, and sustained interdisciplinary collaboration to safely deploy agentic ESE-LLMs for high-stakes environmental challenges.