Rubric Response Theory for Scalar Reward Aggregation
September 27, 2026
Rubric Response Theory (RRT) provides a method for converting multi-criterion rubrics into scalar rewards for reinforcement learning. It uses monotone indicators to select and weight criteria based on how effectively they distinguish between different model rollouts.
HOW THIS AFFECTS YOU
●
builderYou can more efficiently use complex rubrics to train models on non-binary tasks.
●
researcherThis improves the statistical validity of reward signals in complex RLHF tasks.