The authors replace a rubric-coverage reward with an F-beta score, a weighted harmonic mean of recall and precision.
Rewarding precision reduces unnecessary and incorrect content in rubric-trained models
Across medical and science benchmarks, the authors report that balancing rubric coverage with content precision improves response preferences while preserving similar coverage.
Top University
Hyun Ryu · Jong-Heon Kim · Yiqing Xie · Seungone Kim · Carolyn Rose · Eunho Yang
KAIST · Carnegie Mellon University
Research Digest··2 min read
Ryu and colleagues study a failure mode in reinforcement learning: rewarding models for covering a checklist can encourage longer answers with unnecessary or incorrect material.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.