Controlling Parameter Drift in LoRA Continual Test-Time Adaptation via Reset Detection
When a model adapts to new data at test time, it can drift — over-fitting to a shifting domain until it forgets what it knew. I'm studying whether freezing the base weights and adapting only the LoRA layers keeps that in check, and whether watching the model's prediction entropy can detect the moment drift sets in.
When entropy spikes past a threshold, the system dynamically resets the adaptation. The question: does reset detection help more as the domain shifts get more severe? That's what the experiments are built to measure.