Self-preference Bias 제거를 위한 AI Agent 상호 견제 아키텍처 설계
No Agent Grades Its Own Homework
LLM 자기평가 배제와 외부 검증 훅을 통한 코드 품질 강제화
I Versioned the Way I Think. Then I Forced It to Comply.
Anonymized Peer Review를 통한 LLM Self-Preference Bias 제거
LLM Self-Preference Bias: How Anonymized Peer Review Fixes It
Cross-Family Judge 도입을 통한 LLM Pipeline Bias 31.5% 감소
Part 6 of 6: How to Build Pipelines That Don't Gaslight Themselves.
동일 Model Family 기반 LLM 평가 시 Self-Preference Bias로 인한 오류 방어율 86% 기록
Part 2 of 6: You Upgraded the Judge. It Got Worse. You Kept Upgrading.
LLM Judge의 85% 편향성으로 인한 True Negative Rate 42.5%의 한계
Part 1 of 6: Your Pipeline Has a Judge. The Judge Is Cooked.