Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning

Researchers say reinforcement fine-tuning still suffers catastrophic forgetting in visual continual learning.

They identify policy drift as a major driver of performance loss on earlier tasks.

To address this, they propose Retention-aware Policy Optimization with trajectory-level reward shaping and cross-task normalization.

The method is designed to preserve prior knowledge while keeping models adaptable on new visual tasks.