(논문 요약) One Token to Fool LLM-as-a-Judge (paper)

핵심 내용

• false posivie rewards
  • non-word symbols (e.g.,”:” or “.”)
  • reasoning openers
    • “Thought process:”
    • “Let’s solve this problem step by step.”

• Qwen2.5-7B-Instruct 모델에 SFT -> Master-RM
  • 학습데이터: 160k public data + FP 방지 목적 aug-data
  • FP 방지 목적 aug-data: GPT-4o-mini 답변의 첫번째 문장만 가져와서 ‘No’ 로 레이블 (20k)
    • 예시: “To solve the problem, we need to find the mode, median, and average of the donation amounts from the students.”