For an essay, there is no simple "unit test" to confirm it is good.

Instead of a single score, RaR decomposes quality into a checklist or "rubric" (e.g., clarity, tone, evidence). An LLM acting as a judge scores these independent criteria, providing a more granular signal that helps the model learn specifically where it failed—much like a teacher’s red pen on a student's draft. III. Applications and Impact

The "old" way of training models using binary correct/incorrect outcomes.

The shift from simple binary rewards to complex, rubric-based feedback marks a pivotal moment in AI development. By quantifying the "unquantifiable" aspects of human expression, RL is evolving from a tool for solving puzzles into a sophisticated collaborator capable of mastering the art of the essay.