Explain the limitations of using LLM-as-a-judge for evaluating instruction-following capabilities, and how would you mitigate bias in this setup?
