Introduction


  • Uncertainty estimation을 활용한 selective abstention
  • Safety 향상을 위한 “know what you don’t know” 접근

Related Papers


  • Uncertainty quantification
  • Selective prediction

Methods


  • 다양한 uncertainty estimation 방법 비교
  • Uncertainty threshold 기반 abstention 메커니즘

Results


  • Uncertainty-based abstention으로 safety metrics 향상
  • Accuracy-coverage trade-off 분석

Discussion


  • Self-knowledge를 safety mechanism으로 활용
  • Uncertainty estimation의 metacognitive 해석