EvoRubric: Self-Evolving Rubric-Driven RL for Open-Ended Generation
EvoRubric is a novel single-policy co-evolutionary Reinforcement Learning framework designed for open-ended generation tasks lacking static evaluation criteria. It unifies response generation and rubric creation under one policy, allowing the model to dynamically alternate between reasoning and generating evaluation rules. This self-evolving system creates better, more discriminative rubrics than static expert annotations alone.