Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria

Researchers introduced Auto-Rubric as Reward (ARR), which extracts prompt-specific evaluation rubrics from a vision-language model.

They also propose Rubric Policy Optimization (RPO), turning rubric-based judgments into binary rewards for training.

The approach improves multimodal alignment on text-to-image and image editing benchmarks versus pairwise reward models and VLM judges.

It may make reward modeling more interpretable, data-efficient, and less prone to bias and reward hacking.