In active learning, whether to retrain a model from scratch each round or fine-tune it from the previous checkpoint is usually treated as a fixed implementation detail. A new paper accepted to the EMNLP 2026 Main Conference argues it should instead be a decision variable — and that the choice has exploitable structure.
The authors observe that retraining is most valuable in early rounds, when each newly labeled batch can substantially reshape the labeled distribution, while fine-tuning becomes safer once the model trajectory stabilizes. Based on this, they propose HybridAL, an adaptive training schedule that monitors an online stabilization signal and switches once — from retraining to fine-tuning — after sustained stabilization.
HybridAL is studied with two complementary signals that occupy different points on the time–calibration trade-off: spectral exponent change (Δα), a weight-based signal, and accuracy change (ΔAcc), a validation-based signal.
Across three encoder backbones and six text-classification tasks with five seeds each, the authors report that HybridAL keeps endpoint macro-F1 non-inferior to both retraining and fine-tuning at a 0.010 margin, saves up to 49% of retraining time, and recovers a substantial fraction of retraining's calibration advantage as measured by negative log-likelihood (NLL).
Compared with schedules that switch at a pre-committed round, HybridAL achieves lower NLL at moderate additional cost, suggesting that trajectory-dependent switching offers a stronger time–calibration trade-off than fixed early switching.