#1002: Exploring Cognitive, Social, and Metacognitive Dimensions of Student-Student-LLM Collaboration in Problem Solving

If large language models (LLMs) are to be framed as collaborative tools, it is important to understand the nature of LLM collaboration and how it may differ from human collaboration. To this end, we studied problem solving episodes between undergraduate dyads and ChatGPT, observing the extent to which students’ interactions with ChatGPT mediated their cognition, social interactions, and/or metacognition. We found mixed results among dyads (including one that avoided ChatGPT altogether) and elaborate on four cases in-depth, highlighting interactions between students and juxtaposing them with ChatGPT collaboration. Dyads that collaborated with ChatGPT spent more talk turns monitoring its output whereas dyads who collaborated with each other developed more cognitive and metacognitive awareness of the task. Our findings raise implications for tradeoffs that educators should consider when deciding if and when to introduce LLMs into problem solving.