Leveraging Large Language Models to Promote AI-Infused STEM Problem-Solving for Middle School Students
Conference
Rao, A, Piryani, K, Jiang, S et al. (2025). Leveraging Large Language Models to Promote AI-Infused STEM Problem-Solving for Middle School Students
. CEUR Workshop Proceedings, 4019
Rao, A, Piryani, K, Jiang, S et al. (2025). Leveraging Large Language Models to Promote AI-Infused STEM Problem-Solving for Middle School Students
. CEUR Workshop Proceedings, 4019
AI-infused STEM problem-solving is becoming an increasingly important skill for the future STEM workforce, requiring innovative systems to support student learning. Integrating AI into STEM problem-solving requires building a strong foundation in students' computational thinking skills, which can be supported through carefully designed, technically advanced systems. In this paper, we propose augmenting a block-based programming environment specialized for AI-infused STEM problem-solving with large language model (LLM) capabilities to reinforce key computational thinking skills. Through our prior experimentation with students, we have identified three major computational thinking skills essential for mastering learning and transferring knowledge between contexts: abstraction, algorithmic thinking, and generalization. To this end, we enhance an LLM with knowledge about the high-level concept of breadth-first search (abstraction) and the ability to situate students' steps within the required steps of the BFS algorithm (algorithmic thinking). We then evaluate the LLM's ability to provide adaptive feedback as students implement BFS in various STEM contexts (generalization). We present a proof-of-concept evaluation demonstrating how an LLM trained with general BFS knowledge can provide adaptive, contextualized feedback in three different scientific scenarios: pathfinding (mathematics), contact-tracing (biology), and the time-to-live (TTL) package algorithm (networks). This functionality allows our environment to support students in developing abstraction, algorithmic thinking, and generalization skills when applying BFS to scientific problem-solving.
