PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning

A new method lets pretrained machine-learning models learn new tasks without retaining data from earlier training, a constraint that has long complicated the adaptation of large foundation models. The technique, called PLATE, exploits geometric redundancy inside pretrained networks to protect previously acquired knowledge while adding new capabilities. The approach was detailed in a paper posted to the arXiv preprint repository and last revised in June 2026 [1]. The authors observe that pretrained networks contain substantial geometric redundancy — neurons that encode overlapping or correlated information — and that this redundancy can serve two purposes [2]. First, redundant neurons act as a proxy for the dominant feature directions learned during pretraining, allowing the construction of update subspaces that are approximately protected from interference [2]. Second, redundancy suggests where to place new plasticity: by confining updates to a subset of redundant neurons and constraining the remaining degrees of freedom, the method yields update families with reduced functional drift on the old-data distribution and stronger worst-case retention guarantees [2].