Number of Players, Zero: Using Multi-Agent Collaboration to Evaluate Unsolvable Collaborative Items

While the importance of collaborative problem-solving (CPS) is widely acknowledged, developing and evaluating CPS items remains challenging.This study extends to multi-agent systems (MASs) prior work on CPS item templates and the use of large language models for evaluation of collaborative interdependence in math problems designed for two solvers working together.Specifically, we test the hypothesis that MASs more accurately handle evaluation of flawed items that are unsolvable even if both students collaborate with respect to the information and/or answer affordances provided to them.