General continual learning (GCL) aims to learn from ever‑changing data streams without task identifiers, explicit boundaries, or repeated access to past data, making it a realistic yet challenging scenario for continual intelligence. Pre‑trained models (PTMs) provide rich prior knowledge that can alleviate the limited supervision and non‑stationarity of GCL, but existing PTM‑based approaches usually fine‑tune the pretrained representations directly, overlooking two critical gaps: the misalignment between upstream pretraining and downstream continual adaptation, and the unreliability of conventional output alignment under blurry streams. To address these issues we introduce MePo++, a unified post‑training framework that bridges pretrained knowledge to GCL through representation refinement and reconciliation. MePo++ consists of two complementary components:
- MetaPrep: performs unsupervised meta‑refinement over pseudo‑continual sequences to improve representation plasticity for downstream adaptation;
- StreamAlign: reconciles evolving online features with a stable pretrained geometry, reinforcing representation stability and preventing drift.
By enhancing learnability before adaptation and preserving alignment during continual learning, MePo++ enables PTMs to remain both plastic for new concepts and stable over evolving streams. Extensive experiments across diverse PTMs, datasets, and continual learning baselines demonstrate consistent effectiveness and generality of MePo++. The code is released at https://github.com/SunGL001/MePo_Plus.
Review