Distort, Distract, Decode: Instruction-Tuned Model Can Refine its Response from Noisy Instructions