Reproduce “RubricArmor: Adversarial Evolution Improves LLM-Based Rubric Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haocheng Yang, Yuchao Zhang, Licheng Pan, Jiajun Fan, Maolin Wang, Kangning Zhang, Shuai Shao, Shijian Wang and 3 more
RubricArmor tests draft rubrics by constructing defective answers that pass them, verifies the defects, and repairs the rubrics before reinforcement learning. It uses an attacker and review agents, with LLM judges confirming defects. Across nine evaluation sets, it reaches 77.09% macro preference accuracy, versus 73.97% for ChasingTail, but this measures benchmark response ranking, not universal prevention of reward hacking.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.