Tag
This paper introduces a diagnostic framework using Sparse Autoencoders to analyze concept-level forgetting in continual learning, finding that much forgetting is due to representational inaccessibility rather than erasure.