Train Smarter, Not Harder: Switching Signal-Guided Training in Active Learning
HybridAL switches active-learning runs from retraining to fine-tuning once the model shows sustained stabilization.
The paper treats the retrain-vs.-fine-tune choice as a decision to make during active learning, not a fixed setup. It says retraining helps most in early rounds, while fine-tuning becomes safer after the trajectory settles. HybridAL uses either a weight-based spectral signal or a validation-accuracy signal to decide when to switch. Across the reported text-classification tests, it kept macro-F1 within a 0.010 non-inferiority margin while saving up to 49% of retraining time and improving the time-calibration trade-off versus fixed switch schedules. HF Daily Papers' note
The paper treats the retrain-vs.-fine-tune choice as a decision to make during active learning, not a fixed setup. It says retraining helps most in early rounds, while fine-tuning becomes safer after the trajectory settles. HybridAL uses either a weight-based spectral signal or a validation-accuracy signal to decide when to switch. Across the reported text-classification tests, it kept macro-F1 within a 0.010 non-inferiority margin while saving up to 49% of retraining time and improving the time-calibration trade-off versus fixed switch schedules. HF Daily Papers' note
score 4