Modeling is a key scientific practice emphasized in the Next Generation Science Standards (NGSS), but students often need significant support to engage in the practice meaningfully. This project aims to develop an augmented reality (AR) learning environment integrated with a large language model (LLM)-powered pedagogical agent to guide middle school students' modeling practice. It will bring together computer scientists, human-computer interaction researchers, science education researchers, and K-12 teachers to co-design the learning environment. The AR environment will allow students to interact with simulated objects and phenomena as they develop their models. The LLM-powered agent will provide timely assessments of students' diagrammatic models, offer personalized feedback, and share insights with teachers to inform instruction. This project has the potential to transform the way students engage in modeling practice in K-12 science classrooms. Over three years, it will directly impact approximately 10 middle school teachers and 400 middle school students. The project is guided by three core objectives: 1) to build AR simulations for modeling tasks, 2) to develop an on-device multitask LLM to assess diagrammatic models, and 3) to investigate student modeling practice within the AR-LLM learning environment. A key innovation of the project is the closed-loop feedback system: students first receive suggestions for model revisions from the LLM and revisit AR components, a