The paper studies when to retrain an active-learning model from scratch versus fine-tune it, finding retraining wins early and fine-tuning becomes preferable once the model stabilizes. HybridAL monitors an online stabilization signal — a weight-based spectral-exponent change plus a validation-based accuracy change — and automatically switches from retraining to fine-tuning once sustained stabilization is detected. Across multiple encoder backbones and text-classification tasks it matches the best of either fixed strategy while cutting retraining time by up to 49% and improving calibration.
