arXiv: 'Can Scale Save Us From Plasticity Loss in LLMs?' probes continual-learning limits

arXiv·medium signal

This paper investigates plasticity loss — a network's declining ability to learn new information after prior training — and asks whether sheer model scale mitigates it in large language models. The answer matters for any agent that must adapt over time (memory updates, fine-tuning on fresh data, run-over-run self-improvement) without catastrophically forgetting. It's a direct empirical look at whether bigger models are inherently better at lifelong learning or whether plasticity loss persists regardless of scale.