Large language models face catastrophic forgetting when continually fine-tuned on diverse tasks. Mode connectivity can balance plasticity and stability, leading to improved performance.
