Proceedings of the Annual Meeting of the Cognitive Science Society

Group-level cognitive diagnosis is fundamental in educational assessment, where examinees are organized into groups like schools or classes. However, most methods rely on mean-field assumptions that treat students independently, ignoring dependencies induced by shared resources. This oversight yields unstable estimates, particularly under sparse observations. We propose Group-aware Latent Ability Diagnosis (GLAD), a hierarchical variational framework that models within-group dependence via a latent group context while retaining scalable inference. We decompose student ability into a global mean, a group effect, and an individual residual, yielding three variants: a group-conditioned prior, a deterministic effect, and a stochastic effect. For efficient inference, we develop a permutation-invariant encoder to aggregate student representations. Experiments on real-world Eedi data demonstrate consistent improvements over baselines across multiple CDMs.