Effectiveness of Self-Proposed Rubric Rewards in Non-Verifiable Domains

Determine the effectiveness of RLCER—reinforcement learning with chain-of-thought supervision via self-evolving rubrics—for reasoning tasks in non-verifiable domains where automatic final-answer verifiers are unavailable, specifically assessing whether self-proposed rubric rewards provide useful learning signals and yield performance improvements.

Background

RLCER augments reinforcement learning with verifiable rewards (RLVR) by rewarding chain-of-thought using self-proposed, self-evolving rubrics, and the paper’s experiments focus on verifiable math reasoning tasks.

While rubrics provide meaningful supervision signals in RLVR settings, the generality of this approach to domains lacking automatic verifiers (non-verifiable tasks) is not assessed within the paper, leaving a concrete gap in understanding its broader applicability.