RubricArmor: Adversarial Evolution Improves LLM-Based Rubric Generation
Published in arXiv preprint, 2026
Recommended citation: Haocheng Yang, Yuchao Zhang, Licheng Pan, Jiajun Fan, Maolin Wang, Kangning Zhang, Shuai Shao, Shijian Wang, Yuan Lu, Chunyuan Zheng, Hao Wang. "RubricArmor: Adversarial Evolution Improves LLM-Based Rubric Generation." arXiv preprint arXiv:2610.05308, 2026. https://arxiv.org/abs/2610.05308
RubricArmor hardens LLM-generated rubrics for rubric-based RL against reward hacking before training starts. Attack and repair steps alternate over several rounds: the attack simulates how a policy would game the rubric with adversarial responses, and the repair closes the gaps it finds.