CriPO
Notes on Criterion-Distilled Policy Optimization - a rubric-based RL method that uses on-policy self-distillation to explore unsatisfied criteria and preserve useful criterion-level signals otherwise suppressed by scalar reward aggregation.
Notes on Criterion-Distilled Policy Optimization - a rubric-based RL method that uses on-policy self-distillation to explore unsatisfied criteria and preserve useful criterion-level signals otherwise suppressed by scalar reward aggregation.