Domain Shift Stress Test for Continual Learning
Evaluating retention and plasticity tradeoffs under severe sequential domain shifts.
2025-12-15
Problem
Models fail to retain prior tasks after aggressive domain transitions.
Hypothesis
Regularized adapters provide better retention/plasticity balance than full-finetuning.
Method
Ran sequential fine-tuning and adapter baselines on three domain families.
Results
Adapters improved retention while maintaining competitive final-task accuracy.
Future Work
Explore hybrid replay + adapter strategies for long task horizons.