Preference-Aware Rubric Learning for Personalized Evaluation

Researchers introduced PARL, a framework for personalized evaluation of large language models.

PARL learns evaluation rubrics directly from raw user histories, rather than relying on generic benchmarks.

It adds a self-validation step and a discriminative reinforcement learning objective to assess personalized text generation.

The authors say the learned rubrics better capture user-specific preferences and generalize across users and tasks.