Runsheng Chen

LLM judges aren't created equal: lenient judges yield optimized prompts that generalize better than those from strict judges, revealing a critical dependency on judge disposition for effective prompt optimization.