Abstract
Continual Generalized Category Discovery (C-GCD) faces a critical challenge:incrementally learning new classes from unlabeled data streams while preservingknowledge of old classes. Existing methods struggle with catastrophicforgetting, especially when unlabeled data mixes known and novel categories. Weaddress this by analyzing C-GCD's forgetting dynamics through a Bayesian lens,revealing that covariance misalignment between old and new classes drivesperformance degradation. Building on this insight, we propose Variational BayesC-GCD (VB-CGCD), a novel framework that integrates variational inference withcovariance-aware nearest-class-mean classification. VB-CGCD adaptively alignsclass distributions while suppressing pseudo-label noise via stochasticvariational updates. Experiments show VB-CGCD surpasses prior art by +15.21%with the overall accuracy in the final session on standard benchmarks. We alsointroduce a new challenging benchmark with only 10% labeled data and extendedonline phases, VB-CGCD achieves a 67.86% final accuracy, significantly higherthan state-of-the-art (38.55%), demonstrating its robust applicability acrossdiverse scenarios. Code is available at: https://github.com/daihao42/VB-CGCD