Towards Understanding the Robustness of LLM-based Evaluations under Perturbations