Reproduce “CRISP: Cultural Reward Modeling for Implicit Situated Propriety” 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
Zekun Yuan, Yangfan Ye, Baohang Li, Shuaibo Zhao, Zekun Zhou, Ziming Li, Qichen Hong, Kun Chen and 1 more
The authors introduce CRISP, a reward model that evaluates culturally situated actions and guides language-model optimization. It learns preferences between responses to scenarios, while optional Norm Grounding Supervision rewards explanations tied to a target norm. On Qwen3-4B, NormCompass scores rise from 3.27 to 3.69 with CRISP-RM and 3.86 with NGS; these setup-specific results do not establish universal cultural rules.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.