This work presents a systematic evaluation of continual learning for medical visual question answering across diverse clinical objectives, including classification, detection, and report generation.

  • The study explores the ability of existing methods to mitigate catastrophic forgetting.
  • It analyzes sensitivity to task ordering and how sequences influence performance retention.
  • It examines the evolution of low-rank adaptation parameters to reveal weight drift patterns.

The findings suggest that current continual learning methods struggle to maintain a stability-plasticity balance when tasks with different objectives and supervision formats are interleaved.