Title: Treating Continual Learning as a Multiphase Moving-Boundary Problem

Abstract

Stefan-CL offers a sophisticated solution to the stability-plasticity dilemma inherent in continual learning—the ongoing challenge of preserving established knowledge while integrating new information. Drawing inspiration from the physics of melting, this approach conceptualizes consolidated knowledge as a rigid "solid" state and unused neural capacity as a fluid, adaptable "liquid." As the model undergoes training, the interface between these two phases shifts, regulated by a mechanism analogous to "latent heat." By mathematically stabilizing the learned core, Stefan-CL effectively eliminates catastrophic forgetting, achieving performance comparable to data-intensive baseline methods without the need to retain raw datasets. This physics-based framework establishes an elegant new trajectory for artificial intelligence development.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC