Reproduce “Intra-Rubric Reward Divergence: Mitigating Rubric Hacking through Latent Dimension Recovery and Balancing” 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
Jianqiang Xia, Han Xu, Dianxi Shi
The authors introduce FRDA and RSDC to address rubric-based reinforcement learning, where scores can rise while response quality dimensions decline. FRDA normalizes scores across rubric dimensions covered by a prompt, while RSDC adds a preference-model residual signal. On Qwen3-4B, FRDA reached 72.46 versus 67.41 on Arena-Hard v2.0; RSDC reached 79.58, but was tested only on chat.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.