A computationally efficient PID control-based self-healing framework can improve the robustness of pre-trained large language models against a wide range of perturbations.