The Stable Recovery Manifold: geometric principles of recoverability in continual learning

arXiv·low signal

This work reframes catastrophic forgetting in continual learning not as outright destruction of prior knowledge but as a question of recoverability, identifying a 'stable recovery manifold' that governs which forgotten capabilities can be cheaply restored. Theoretical, but relevant for anyone running sequential fine-tuning who wants to reason about what's truly lost versus dormant.