Adaptation with Self-Evaluation to Improve Selective Prediction in LLMs