Reliability Assessment of BERT Models in Question Answering
August 12, 2026
This study evaluates the stability of BERT, RoBERTa, ALBERT, and DistilBERT under Monte Carlo Dropout and input paraphrasing. The findings highlight that while encoder-only models show high accuracy, their reliability remains inconsistent under perturbation.
HOW THIS AFFECTS YOU
●
researcherYou should incorporate Monte Carlo Dropout to better estimate uncertainty in transformer-based QA tasks.